Precision viticulture service at TerraNIS aims at optimizing the management of the vine, facilitating
maturity control, optimizing harvesting plans and increasing in plot profitability through different bio-
physical and vegetation parameters produced using Oenoview, which was developed by Airbus Defence and Space’s Geo-intelligence line program. One of the first steps in order to provide these maps is to define vineyards’ boundaries. Nonetheless, generating these boundaries manually is a non-trivial task especially when the number of parcels is large. Recent advances in machine learning and more specifically
deep learning have fostered the computer vision field and its application. Deep learning algorithms allow machines to understand images, extract their features and hence performing various computer vision tasks such as semantic and instance segmentation as well as object detection. The Mask R-CNN [1] is a deep learning model that extends the well-known Faster R-CNN model [2] in order to perform instance detection and segmentation. The Mask R-CNN has proven to be successful in various computer vision
tasks, e.g., nuclei segmentation in microscopy images [3] and label-free cell tracking in phase contrast microscopy [4]. The Mask R-CNN approach consists of four main steps. The first one is the backbone model, which is a convolutional network that extracts features from input data. Common choices of the backbone model are ResNet 50 and ResNet 101. The second step aims at detecting objects in the input images using the Region Proposal Network (RPN). The third step produces an output class and
a bounding box for the regions of interest (ROI) proposed by the RPN. Finally, a segmentation step is performed on the ROIs in order to generate segmentation masks. It is worth noting that transfer learning can also be applied when using the Mask R-CNN model through pretrained weights of very large datasets such as COCO and ImageNet.
The zone of interest (ZOI) used for parcel delineation is located in Bordeaux, France (-0◦359179’W 44◦ 519578’N: -0◦081316’W 44◦ 797136’N), which is characterized by its richness of vineyards. This ZOI covers an area of 481.5 km2. A SPOT 6 image acquired during August 2017 of size 24080 × 19612 in the RGB color space with a spatial resolution of 1.5 × 1.5 m2 was used (see Fig. 1[a]).
To prepare the training dataset, 4.3 km2 of the same ZOI were labeled manually in a first step. This surface was divided into 5 training patches in order to capture the variety of vineyards in the area. In a second step, random crops of size 384 × 384 were applied on these patches yielding a dataset that contains 1767 images. The Mask R-CNN was trained using the COCO dataset training strategy [5] and the prepared dataset. Transfer learning and image augmentation techniques were also applied which allow training the model with such a small number of annotated images. The detected vineyard parcels in this
ZOI are shown in Fig. 1[b].
To further investigate the obtained results, Fig. 2 depicts the ground truth vineyard parcel boundaries of a small zone in the ZOI (Fig. 2[a]) (0.33 km2 ) and the corresponding estimated ones (Fig. 2[b]) . From a visual perspective, these results confirm the ability of such a model in detecting vineyard parcel boundaries. Dice and Intersection over Union (IOU) were also estimated. On the one hand, the Dice coefficient is a statistic to compare the similarity between two samples, which allows comparing the segmentation results. On the other hand, the IoU coefficient is a metric to evaluate object detection using objects’ bounding boxes. The estimated
Dice and IOU coefficients of 0.93 and 0.86 between the ground truth and the estimated results are also coherent with the visual comparison. However, the availability of training data captured from different sites is essential for creating a generalized model that is able to detect vineyard parcels with different spatial and spectral structures. In this context, parcel registration databases such as the Geographic Parcel Registration (RPG) in France [6] have a great potential in training such a model for large scale parcel detection. Unfortunately, the French RPG has a very small number of registered vineyard parcels, which limits its usefulness for this specific task. Nonetheless, Terranis has been developing a Mask R-CNN-based model to detect large crop parcels such as wheat, maize and sunflower using the RPG and Sentinel-2 images with promising first results.
Future work will focus on expanding the training dataset through an iterative strategy, i.e., by refining the model predictions and feed them to the training dataset. A further performance analysis will then be conducted by evaluating the algorithm trained using the extended dataset on other geographical areas.

REFERENCES
[1] K. He, G. Gkioxari, P. Dollar, and R. Girshick, “Mask R-CNN,” in Proc. ICCV, 2017, pp. 2961–2969.
[2] S. Ren, K. He, R. Girshick, and J. Sun, “Faster R-CNN: Towards real-time object detection with region proposal networks,” in Advances in neural information processing systems, 2015, pp. 91–99.
[3] J.W. Johnson, “Adapting Mask-RCNN for automatic nucleus segmentation,” arXiv preprint arXiv:1805.00500, 2018.
[4] H.-F. Tsai, J. Gajda, T. Sloan, A. Rares, and A. Shen, “USIIGACI: Instance-aware cell tracking in stain-free phase contrast microscopy enabled by machine learning,” bioRxiv, 2019.
[5] Abdulla W., “Mask R-CNN for object detection and instance segmentation on Keras and Tensorflow,” https://github.com/matterport/Mask RCNN, 2017.
[6] Institut Géographique National, “Le Registre Parcellaire Graphique,” http://professionnels.ign.fr/rpg.

Geodesign describes a collaborative design activity by different people: scientists, architects, engineers, planners, and people of the place where computers and technology provide rapid responses to changes to the designs in real-time. Geodesignhub provides a systematic way to create a design and enables the different stakeholders to work together on a design project. There are three key aspects to the Geodesign process: understand the past, build an assessment of the current state and then engage stakeholders to manage the future. To build a site assessment, one needs to understand how a site functions, what are the major human, physical and ecological processes at play and how are they linked. Once the processes are understood an assessment of whether they are working well must be made. For this site assessment sometimes using geo-spatial data is desirable, generally these are data such as roads, rivers, land use etc. In addition, it could be social and economic data. This is where space related data comes in.
Advances in satellite technology and the open availability of high-resolution earth observation data (e.g. Sentinel 2) have important implications for geodesign. This when combined with the advances in cloud computing infrastructure mean that a new class of geo-applications can be built that leverage supercomputers and huge data processing and analysis capability.
Starting with Sentinel we have a huge repository of continuous earth observation imagery and more recently drones are also playing a part. This data is used to understand the processes at work in the study area and are used to make site assessments in the form of maps that communicate to the different stakeholders the state of play. For making “site evaluation maps” we leverage Earth Observation (EO) technology. Evaluation maps are red/ yellow/ green maps that the stakeholder what is happening on the site. For example, below is an Agricultural assessment map made using Sentinel data. The Green are areas that can support agricultural interventions and changes and yellow are areas that are not suitable for agriculture (e.g. there may be housing already existing). The map on the right hand shows a Green Infrastructure assessment for a site, the areas in red mean there is already Green Infrastructure in place (no management needed) areas in green mean that new green infrastructure can be added to these locations.
Geodesignhub has built a sophisticated and intuitive user interface and a process to engage stakeholder with design and help them negotiate about the future using the data. In this talk Dr. Ballal will focus on an open-source toolset built by Geodesignhub (www.geodesignhub.com) that uses advanced geo-computing environment built by Mundialis GmbH (www.mundialis.de). In addition, Dr. Ballal will also talk about other open source tools built on other ESA products (e.g. ESA Climate Change Initiative data) that help with the geodesign workflow. Dr. Ballal will demonstrate how this toolset enables high performance collaborative design by providing citizens and professionals site assessment and impacts of design interventions in near real time.

In the words of Etzioni, “thought is not a vector and artificial intelligence (AI) is not a problem in statistics” [1], where statistics is related to the quantitative/unequivocal notion of information-as-thing, typical of data transmission. Rather, as part-of the multi-disciplinary domain of cognitive science, AI is related to the qualitative/equivocal notion of information-as-data-interpretation. The race to commercialize AI has put enormous pressure on researchers, whose accepted AI paradigm is to produce massively overparameterized deep learning (DL) network which take hundreds of thousands of examples to work “well enough” (with low bias, i.e., low error rate) on narrow tasks, despite featuring large variance (dependence on input data), synonym for low transferability/scalability, which is monotonically increasing with model complexity (equal to the number of free-parameters, to be learned from data, plus hyperparameters, to be user-defined based on heuristics) [2], [3]. Indeed, the fundamental machine learning (ML) principle well known as “bias variance” dilemma seems to be ignored by the accepted paradigm in DL and AI [3]. Actually, a low-bias high-variance ML trade-off was dismissed as out of hand by Geman et al. more than 25 years ago [4]. What can we assume is built into the brain, and what must be learned from examples? What about the visual cortex? The current deep learning paradigm assumes the visual cortex is largely plastic, where all weights are randomized and must be learned from data. Is this a plausible hypothesis? Geman et al. argued from the theory of nonparametric statistics that this cannot be the case, unless AI researchers are willing to use up exponentially many samples to learn classes of objects from data. Hence, the goal of developing trustworthy (low-bias low-variance) AI will require a deeper investigation into humans’ remarkable abilities to learn from small datasets and new insights into the cognitive mechanisms humans employ to reliably (at low bias) and robustly (at low variance) understand the world. In the words of Mitchell, “unlocking AI’s barrier of meaning (semantics) is likely to require a step backward for the field, away from ever bigger networks and data collections, and back to the field’s roots as an interdisciplinary cognitive science studying the most challenging of scientific problems: the nature of intelligence” [2]. To comply with an ongoing critical appraisal of DL by AI pioneers, encompassing Hinton [5], grandfather of supervised data learning (inductive) Deep Convolutional Neural Networks (DCNNs), the present work presents and discusses a computational model of low-level vision in mammals, synonym for physical model-based (deductive) parameter-free automatic image-independent DCNN for image-contour detection, image segmentation and keypoint detection as raw primal sketch. Our working hypothesis is ‘Human vision -> Computer vision (CV)’, where symbol ‘->’ identifies relationship part-of, pointing from the supplier to the client. Encompassing both biological vision and CV, vision is synonym for scene-from-image reconstruction and understanding, which is an inherently ill-posed cognitive (information-as-data-interpretation) problem. Vision is inherently ill-posed because affected by: (i) a data dimensionality reduction, from the 4D space-time scene-domain to the (2D) image-domain, and (ii) the semantic information gap, from ever-varying sensations (sensory data) in the image-domain to stable percepts (concepts) in the mental model (ontology) of the physical 4D world. Since it is ill-posed, vision requires a priori knowledge in addition to data to become better conditioned for numerical solutions. One of Marr’s legacies is the notion of computational constraints, equivalent to genotype, reflecting properties of the world and embodied through evolution into the human visual complex system [6]. Marr’s computational constraints are Bayesian priors in a Bayesian approach to CV. Based on these observations, to enforce a ‘Human vision -> CV’ paradigm, capable of overcoming the black box problem affecting ML solutions in general and DL solutions in particular, whose traceability/interpretability typically scores “low”, the following original constraints can be inferred from human vision to make an inherently ill-posed CV system better conditioned for numerical solution. (1) In biological vision, spatial information typically dominates color information. For example, humans typically wear sunglasses because their panchromatic and chromatic visual systems are nearly as effective in scene-from-image reconstruction and understanding. If a chromatic CV system does not down-scale seamlessly to achromatic image analysis, then it tends to ignore the paramount spatial information in favor of subordinate (secondary) spatial context-insensitive color information, such as multi-spectral (MS) signatures typically investigated in traditional pixel-based single-date or multi-temporal Earth observation (EO) image understanding (EO-IU) algorithms, where relationship ‘CV -) EO-IU" holds, with symbol ‘-)’ identifying relationship superset-of (with inheritance), pointing from the parent to the child. In other words, a necessary and sufficient condition for a CV -) EO-IU system to fully exploit primary spatial topological information (e.g., adjacency, inclusion, etc.) and spatial non-topological information (e.g., spatial distance, angle distance) components, in addition to secondary colorimetric information, is to perform nearly as well when input with either panchromatic or color imagery. (2) The CV system is required to comply with visual illusions, such as the Mach bands illusion: where a luminance (radiance, intensity) ramp meets a plateau, there are spikes of brightness (perceived luminance), although there is no discontinuity in the luminance profile, independent of the ramp’s slope. It means that local variance, contrast and first-order derivative (gradient) are popular local statistics not suitable to detect image-segments/image-contours required to be consistent with human visual perception, including ramp-edge detection. This straightforward (obvious), but not trivial observation is at odd with a large portion of the existing CV and remote sensing (RS) literature, where many semi-automatic image segmentation/image-contour detection algorithms, based on thresholding the local variance, contrast or first-order gradient, require a system’s free-parameter for thresholding image-objects or image-contours to be user-defined in range (0, +∞) based on heuristics. The proposed CV system for pre-attentional (low-level) vision is a physical model-based (deductive) parameter-free DCNN consisting of: (i) multi-scale multi-orientation banks of 2D Gaussian (low-pass) and even-symmetric (band-pass) spatial filters, i.e., odd-symmetric filters are omitted, whose size is 3×3, like in the popular inductive Visual Geometry Group (VGG) DCNN architecture [7], (ii) 2D spatial filters are separable into products of 1D spatial filters for computational efficiency, and (iii) oriented 1D even-symmetric spatial filters, computed as a second-order derivative of an oriented 1D Gaussian function, are local concavity estimators, eligible for zero-crossing local concavity detection as candidate image-contour pixels where local concavity changes in sign. The proposed deductive DCNN is fully interpretable and capable of parameter-free (automatic, no human-machine interaction and no training data set are required): (i) panchromatic and color image decomposition (analysis/encoding) and lossless (exact) image reconstruction (synthesis/decoding), (ii) brightness (perceived luminance) estimation, in compliance with the Mach bands illusion, (iii) image-contour detection, where image-contours are localized by an original 2D local indicator of spatial autocorrelation (LISA) scoring “low”, (iv) well-posed (deterministic) two-pass image segmentation into connected (closed) image-objects (segments), estimated from (not necessarily connected) binary image-contours, (v) keypoint (salient point, fixation point) detection, (vi) eccentricity-dependent vision (foveated imaging system simulations). Further developments of the proposed physical model-based DCNN for low-level vision first stage as raw primal sketch [6] will focus on texture segmentation (perceptual spatial grouping of texture elements, texels) as full primal sketch at the low-level vision second stage. No physical model-based automatic parameter-free image texture segmentation algorithm exists to date in the CV or RS literature.
References
[1] Oren Etzioni, What shortcomings do you see with deep learning?, Quora, 2017. Available online: https://www.quora.com/What-shortcomings-do-you-see-with-deep-learning
[2] M. Mitchell, Artificial Intelligence hits the barrier of meaning, The New York Times, 2018.
[3] S. Mahadevan,”What problems can occur when a neural network has too many neurons?, Quora, 2019.
[4] S. Geman, E. Bienenstock and R. Doursat, Neural Networks and the Bias Variance Dilemma, Neural Computation, vol. 4, pp. 1-58, 1992.
[5] Axios (2017), Artificial intelligence pioneer, Geoffrey Hinton, says we need to start over, Sept. 15, 2017.
[6] Marr, D. Vision; Freeman and C: New York, NY, USA, 1982.
[7] K. Simonyan and A. Zisserman, Very deep convolutional networks for large-scale image recognition, ICLR 2015. Available online: arXiv:1409.1556v6 [cs.CV] 10 Apr 2015.

Territorial changes, natural or man-made, can have a strong negative impact on land, buildings and infrastructures both in terms of market value and utility. The owners or, more importantly, those who manage real estate assets, need to continuously monitor what happens in the properties (cadastral parcels) under their control and possibly anticipate negative repercussions deriving from unexpected or undesirable changes. Furthermore, land assets often appear as small patches of land scattered across a larger territory, which makes the satellite monitoring operation uneconomic for the large amount of EO resources needed.
The potential of EO solutions in the field of real estate management and related industries is in an early stage of development, and promising business opportunity is expected in the next few years. There are a wide range of potential users in the field of satellite monitoring of real estate assets, such as real estate and land assets management companies, investment trusts, insurance, appraisals and property valuation sector.
This proposal aims to capture the induced opportunity of these industries to open up to new technologies, using earth observation tools and precision satellite navigation systems to offer data-driven solutions to support decision-making.
We propose an innovative EO monitoring procedure based on pay-per-use business model and resources optimization strategy. It’s a cloud-based platform, accessible from a multi-device web dashboard for real-time automatic service of satellite monitoring and geoanalysis of land parcels and buildings. The service speeds up and improves accuracy in geo-spatial data-based decision-making. Employing cognitive extraction technologies and self-developed artificial intelligence algorithms applied to Copernicus satellite imagery it allows users to mine relevant information from various unstructured sources enabling the following outputs:
· Automatic change detection and ML classification of changes (with AI);
· Monitoring of abuses and remote alerting system (for single cadastral parcel);
· Calculation of solar potential (irradiation and slopes) and precision NDVI;
Also, added-values services from non-EO data sources can be easily integrated such as:
· Identification of territorial dynamics and regulatory framework monitoring (territorial area);
· GIS mapping and digitization of the Land Registry information (public and private);
· Recognition of RE market trend (OMI - online market);
Moreover, high precision field services based on GNSS positioning technology can be natively integrated to the platform as web applications:
· Routing and route optimization service;
· Real-time field reporting service;
· Geofencing or virtual fence service;
· Field cadastral interrogation (reverse cadastral consultation).
A distinctive element of the proposal is the use of a data connection infrastructure to facilitate the processes of satellite data acquisition and geo-analysis ensuring high quality standards. In this regard, the ONDA service (Serco DIAS) integrated with ESRI Image Server is used as the main access infrastructure to satellite data. In relation to other data sources, both open and proprietary, the acquisition takes place through automatic procedures of web-scrapping and data mining.
The data organization model is designed to minimize the need for redundant information storage space by favouring on-the-fly processing in most analyses. The innovative key of this model is the satellite images acquisition method. Copernicus satellite images are used to identify the areas of greatest interest (where the automatic change detection procedure identified significant changes) while higher resolution images of commercial constellations are used to classify the changes recognized in the first analysis.
The monitoring process is divided into three phases:
· Automatic identification and assessment of changes,
· Classification of findings through artificial intelligence;
· Reporting of anomalies and alerting to responsible bodies/operators.
The first level analysis – the “tipping”
Once all the cadastral parcels to be monitored have been identified and mapped on GIS, the first automatic change detection procedures can be launched. The change detection geoprocessing requires the use of Copernicus Sentinel2 multispectral images with a resolution of 10m. Starting from the comparison between two satellite images of the same area at different times an algorithm recognizes the differences between the images, pre-classifies the results in macro categories, cluster them within the cadastral geometry, and represents all the results on a web map. The results of the first level change detection procedure are the starting point for a second level of detailed geoanalysis on the specific areas where suspicious changes have previously been detected.
The second level (in-depth) analysis – the “cueing”
Territorial changes may occur due to normal seasonal phenomena, planned changes or as a result of unwanted or unpredictable actions. The latter represent a risk for land assets as they can create damage to third parties and have a negative impact on their market value, in both cases it is preferable to anticipate or minimize damage.
The second level analysis (cueing) deepens the results of the first level analysis (tipping) in which the cadastral parcels have been identified and classified according to the type of change found. The cueing phase gives added value to the process as it allows to automate spatial analysis using very high-resolution commercial satellite images only in very precise geographic areas within the chosen cadastral parcel geometries. For the second level geoanalysis it is used multispectral satellite very high-resolution images from commercial constellations with resolutions of 0.70/0.50cm.
Supported by ad hoc machine learning procedures, the changes are further classified according to their size, type and level of hazard and communicated through a distributed alert system to the various operators and responsible bodies.

Mapmakers, such as the contractors of the European Environmental Agency who deliver the CORINE Land cover, commonly perform statistical classification of pixels or of surface shapefiles. From photograph or from satellite EO images, they are using supervised machine learning generated decision trees to support in understanding the semantics of the image, and also to speed up the design of the thematic layer. Yet, from one image to another, the classification map, based on the same classes, may vary although the objects have not changed. This is due to objects’ inherent or apparent optical properties which change alongside the inter-and-intra class probability distributions.
Thematic chart accuracy relies on operators experience who select the training data sets, perform the quality controls and amend the charts. All the more than the raw information content of the image might not be sufficient to perform the classification according to the classes’ definition, compelling the operator to also rely on contextual analysis while automatic classifiers are merely pixel-based.
Thanks to satellite missions such as MetOp, ENVISAT, Landsat and Sentinel, Land Cover has been thoroughly mapped from year to year by remote sensing labs and geographic department of universities. This has been performed by time-series analysis on evolving characteristics instead of classifying on individual images and then stacking the classifications for data fusion (e.g. work of Jordi Inglada and coworkers at CESBIO and Centre d'Expertise Scientifique "Occupation des sols" /CES-OSO). This new approach using variabilities should improve automatic classification although the need of classifier training is still an issue. It has reached accuracies ranging from 60% to 90%, which although impressive is not always sufficient. Some possible explanations for these shortcomings are inconsistencies between the semantics of images and classification, non-optimal choices of variables and insufficient quality of the training set.
In this context, our work focuses on the issue of the choice of input variables, specifically in the case of random forest based classification. The performance of such methods depends not only on the number of nodes of each tree and on the number of trees, but also on the choice of the optimal differentiators. In fact, in most cases raw remote sensing variables are used (plain radiances or reflectance at sensors’ wavelength for instance) rather than derivatives of those which could be more telling. Notably, enriching the input variables with scientific knowledge could impact performance and improve interpretability of the results. The key is thus to use both geographic and statistical methods to select the relevant features to be used as input variables instead of directly using physical quantities. Several selection procedures have been proposed, in particular two straightforward methods, the mean decrease impurity and the so-called Perm method, as well as another popular one, the recursive feature elimination. Alternatively, some feature selection methods are not used to remove redundant variable, but instead to classify parameters based on their statistical impact and potentially eliminate the least relevant ones. For instance, we may think of the previously quoted Perm method and its parametric variant Altmann as well as the Boruta algorithm, but also recently published approaches like recurrent relative variable importance and the Vita method. In any case, working with experts is key in order to interpret the results of the statistical analysis, confirm its relevance with regards to the classes and eventually propose new variables derived from scientific expertise.
In order to illustrate the relevance of such feature selection methods, some LULC examples on selected regions in Quebec along the Saint Laurent will shown. The work performed on this example could support efforts in sketching an automatic change detection scheme to monitor the impact of climate change on the coastal morphology, as well as the related changes in terms of biota and land use.

Nowadays, the use of Deep Neural Networks for the detection and segmentation of specific objects from satellite images and the creation of fictitious VHR/HR datasets is a key point in the AI4EO community.
The purpose of this work is to support continuous and near-real-time Maritime Domain Awareness services through the classification, detection and segmentation of maritime vessels and the generation of a digital report containing information such as position, length, width and heading of each detected vessel.
Moreover, these results are an ideal starting point to generate Synthetic Datasets by means of Generative Adversarial Networks beyond the state-of-the-art.
Several types of AI architectures have been investigated in the development phase and at, eventually, the final tool is based on a state-of-the-art Deep Learning algorithm, a Region based Convolutional Neural Network, well readjusted on Maritime Case.
In this work, the results of 3 CNN models (Segnet, U-Net, and Mask R-CNN) are presented for the detection of ships from VHR and HR optical images starting from annotated VHR datasets.
SegNet and U-Net have a very similar basic structure VGG16-based, the first one more flexible on several fields and the second coming from the biomedical field, and show good results in the detection of ships of any size in the open sea, even with unfavorable environmental conditions, with slight differences in accuracy in favor of the second. The chosen model is a customized version of Mask-R CNN and has a precious boost for ship detection close to sea coasts, harbor, other vessels and etc. despite classical Fully Convolutional Neural Networks. The training dataset is a sample of thousands flagged VHR maritime images split into 80% for training and 20% for validating reaching very good performances on both of them. The model contemplates Transfer Learning for initial weights for more precise and faster feature extraction and as Overfitting avoidance, Dropout and Early Stopping monitoring validation loss are used, finally as Optimizer a custom Stochastic Gradient Descent is used. The online prediction workflow expects as input any kind of VHR image and this is divided into subtitles, each of them containing a fixed area in kilometers, that is forecasted according to the trained weights. The output is represented by the real image, the image processed by the model weights and a digital report containing vessels information. Any forecasted ship is fitted into an ellipse of which any measure is known as length and width, the area in m² along with longitude, and latitude of the ship. The model reaches outstanding performances even on different VHR datasets like DOTA and Google Earth as well as, obviously, the starting dataset. The ultimate goal is to create a continuous flow of generation and discrimination of satellite image regardless of the source. We are introducing the latest Generative Adversarial Networks techniques to face the fact that Flagged VHR and HR satellite images dataset are few and limited and the accuracy on VHR test datasets clearly exceeds that on HR.
GANs are used in order to create new VHR synthetic samples starting from the open datasets of VHR images and to create synthetic HR flagged datasets by means of Cycle-GAN.
Cycle-GAN are capable to translate the style from a dataset to another one, in this case (from VHR to HR) in order to improve the accuracy of the models on HR images.

The use of Deep Learning techniques for detection and segmentation of objects in satellite imagery is nowadays consolidated and the present work is focused on demonstrating how these techniques accelerate the feature extraction process in the context of traditional mapping services, while ensuring the full respect of cartographic quality standards and providing a real added value, especially in time critical operations.
Translating images into maps requires an intensive effort, in particular in a short timeframe and on a large-scale. Traditionally, cartographic feature extraction process is a time and effort consuming activity, mainly relying on the so-called Computer Assisted Photo-interpretation (CAPI) which has been evolving along with most used GIS suites (e.g. Quantum GIS, ArcGIS) to improve productivity (e.g. editing tools, feature management tools, etc.). The uptake of automated production of cartography has been generally limited by the following factors: a) non satisfactory geometric and thematic quality level achievable, which almost always lead to feature manual editing and reshaping; b) the high amount of time necessary to training feature detection algorithms; c) the relatively high availability of manpower compared to the required mapping tasks; d) the scarce availability of sufficiently large training datasets.
Today, all the above mentioned limitations can be overcome through the exploitation of Deep Learning techniques and the availability of large reference databases (cartographic layers and satellite imagery), coupled with novel ICT resources.
A set of tools has been developed to detect and extract roads, buildings (building footprints or build-up area depending on the resolution class) and hydrographic network in satellite Very High Resolution (e.g. WorldView-2, WorldView-3, Pleiades) and High Resolution (Sentinel-2A/B) images.
Fully Convolutional Networks (FCN), U-Net like, were trained and some post-processing refinements were added in order to regularize the results. Both the multispectral and the RGB approaches were explored with performing results depending on the target.
The data set is composed of Spacenet competition data (images and annotations) and labels scraped from OpenStreetMap. Internal cartographic products for roads, hydrography and buildings are also available and they are manually annotated by expert photointerpreters with high accuracy.
As an example, these models are extremely useful for emergency mapping services: a mapping of transportations and buildings is necessary when a natural disaster happens in order to identify damaged areas and their accessibility. Some parts of the world have not been mapped, especially those most vulnerable and the availability of a tool like this one allows to accelerate the production of reference maps. Post-disaster information and damage assessment are a critical point for speeding up decision making.
A change detection method has been also implemented to detect damaged buildings. In case of emergency damage assessment, the distortion due to very high off-nadir angles needs to be faced. The ability to use this kind of images allows to save time that is precious during emergency management. Therefore, a new deep neural network was trained with the same architecture as the previous one by using a fine tuning approach with a data set of images characterized by very high off-nadir angles. The goal is to have a reliable model that works fine with images having any off-nadir angle and to create a pipeline that requires only an easy validation at the end of the entire process.
Based on the encouraging results achieved, a second problem addressed was how to put emergency mapping operators in the best conditions to handle and exploit the developed models in different geographic areas, without having specific competences on Deep Learning and Artificial Intelligence. To this extent, different production environments have been tested in order to identify the most suitable one to be used by the emergency mapping operators.

Could in principle a machine be as creative as humans are? Is creativity the key ingredient that distinguishes us from computers? These questions are brain teasers for philosophers.

The usual argument against computers being creative goes that they can only do what they are told to do. Humans are the ones programming them, they are just slaves to the instructions given and carry out tasks accordingly. Humans, on the other hand, are not bound by rules. They can innovate, think outside the box. In other words, they can be creative and can create things out of nothing. However, it is egotistical to think that humans are free from being bound, because in reality, just as computers, we inherently run on a set of basic rules, except instead of in zeros and ones, our bits are A, C, T, and Gs.

Still, even though we are somehow programmed, we are able to beat or even lick computers in a creativity match. The computer scientist that coded us is not a beginner. Our brain is an incredible machine, forged by nature for thousands or even millions of years. Contrarily to some AI experts' claims, creativity is not just the clever move taken by AlphaGo. Creativity is changing the rules, is connecting different fields. Human creativity is Einstein’s Theory of Relativity (Special & General), Feynman’s Path Integral, Dante's poems or Kandinsky’s paintings.

There are at least three kinds of creative works. There is exploratory creativity, which is taking the rules of the game and pushing them to the extreme, like in Bach music. There is combinatorial creativity, where you are taking two ideas that have nothing to do with each other to see how associations in one can help stimulate new ideas in the other. The third creativity, which is somehow the most mysterious, are those moments that somehow seem to come out of nowhere.

Exploratory creativity seems suitable for a computer since it can do many more calculations than a human brain could ever do. Combinatorial creativity might be feasible in the future, but the third one seems to be really challenging for a machine.

In our presentation, we are going to show some examples of creative AI such as Google QuickDraw, Cris Valenzuela's AttnGAN, some Neural Style Transfer and the Google Magenta Project. On the other hand, we are going to emphasize some examples, from photography to literature, that in our opinion AI will not be able to generate in the near future.

If AI will not beat human creativity, it might be the paintbrush for the future of art. A paintbrush needs a wielder, so we expect that creativity itself will enhance thanks to the cooperation between humans and machines.

Remote sensing has recently profited from appearance of multiple sensors, like Sentinel-1A/B and Sentinel-2A within the European Copernicus program. Datasets collected by these sensors are publicly available, which presents opportunities for researchers to use them in various domains. Agriculture is one of the domains that benefits the most from this kind of temporal and spatial high resolution data. Among the most essential tasks is crop mapping, that can further allow to estimate crop areas, analyze crop rotation and be used in precision farming. In this work we show how we approach obtaining high-quality field-level crop maps on a global scale.

We choose to operate on an object-based level. Main reasons for that are relative simplicity of data gathering and filtering (compared to pixel-based approach) and natural denoising effect due to averaging within a field. This choice results in the necessity to have field-detection tool. We at OneSoil developed our own instruments for this task, but there exist other tools, like eCognition software. Another essential tool for working with satellite imagery is cloud detector. We also use our own tool, but there exist many tools with decent quality, from F-mask to Sen2Cor. These steps are out of the scope of current work, so we will not discuss them in detail.

There exist many open datasets that contain field boundaries and corresponding crops (France, Belgium, Netherlands). There also are high-quality datasets for California and Washington states, that allow to compare with USDA Cropscape which can be considered as a benchmark in crop classification.

We propose simple robust algorithm for accurate object-based crop classification that uses all available Sentinel-2 images for an agricultural season that is automatically detected for each field. Our approach to feature extraction allows to deal with sparse nonaligned multisource multiresolution data, and our automated season detection based on smoothed NDVI time series greatly enhances transfer of models between various regions.

Data preprocessing and feature extraction process for each field consists of the following steps:
From all Sentinel-2 images (with appropriate validity masks) for a calendar year we crop data using the defined field boundary. We use all 13 bands except for SWIR band 10, which is used for cirrus cloud detection and does not provide any useful information for crops.
Prepared NDVI time series for whole year is smoothed using Savitzky–Golay algorithm, after which the plant processing season is estimated using a combination of threshold methods.
For each band and index, we extract spatial spectral features. Due to object-based approach we can easily use bands with different resolution.
We select all clean dates available for a field within previously selected season. From spectral time-series data, we compute temporal features for all spatial features computed at previous steps.
We also add timestamp descriptors that include several statistics about time of observations

As a result, for any field, we obtain feature vector of size 794 that reflects spectral dynamics over the season. We train the classifier as a fully-connected deep neural network. To train this model we used a large dataset of roughly 8 million fields in European region and US.

To see the effect of automated season detection, we use California data to compare our model with the one trained on fixed season typical to central European region - from March to October (figure 1).



Using ground truth data for the us (California and Washington), we compared our model with the USDA results of predictions (Figure 2). It is seen that our approach outperforms USDA in almost all crops.

To further validate our results, we made predictions for regions where we didn’t have field-level ground truth data, but had region-level aggregated statistics on hectares planted published by the governments (Mexico 2017, figure 3, 4).



Results show that Sentinel-2 data can be successfully used for high-resolution scalable end-of-season crop classification. Optical data sparsity can be efficiently dealt with using our proposed approach for feature extraction. Integration of automatic season detection allows successful transfer to new geographies. Training on large data is achievable due to batch gradient descent learning within deep learning framework.

Given the existence of cloud filtering and fields segmentation tools, this approach results in fully-automated pipeline that can create crop maps for any region. The simplicity of solution ensures high performance even for large scale inference.

We see that results can be further improved by fusing optical and synthetic aperture radar (SAR) data, that is also available through ESA services.

Earth observation from remote sensing satellites allows us to monitor the processes occurring on the land cover, water bodies and the atmosphere, as well as their interactions. In the last decade machine learning has impacted the field enormously due to the unprecedented data deluge and emergence of complex problems that need to be tackled (semi)automatically. One of the main problems is to perform estimation of bio-geo-physical parameters from remote sensing observations. In this model inversion setting, Gaussian processes (GPs) are one of the preferred choices for model inversion, emulation, gap filling and data assimilation. GPs do not only provide accurate predictions but also allow for feature ranking, deriving confidence intervals, and error propagation and uncertainty quantification in a principled Bayesian inference framework.

Here we introduce GPs for data analysis in general and to address the forward-inverse problem posed in remote sensing in particular. GPs are typically used for inverse modelling based on concurrent observations and in situ measurements only, or to invert model simulations. We often rely on forward radiative transfer model (RTM) encoding the well-understood physical relations to either perform model inversion with machine learning, or to replace the RTM model with machine learning models, a process known as emulation. We review four novel GP models that respect and learn the physics, and deploy useful machine learning models for remote sensing parameter retrieval and model emulation tasks. First, we will introduce a Joint GP (JGP) model that combines in situ measurements and simulated data in a single GP model for inversion. Second, we present a latent force model (LFM) for GP modelling that encodes ordinary differential equations to blend data and physical models of the system. The LFM performs multi-output regression, can cope with missing data in the time series, and provides explicit latent functions that allow system analysis and evaluation. Third, we present an Automatic Gaussian Process Emulator (AGAPE) that approximates the forward physical model via interpolation, reducing the number of necessary nodes. Finally, we introduce a new GP model for data-driven regression that respects fundamental laws of physics via dependence-regularization. All models attain data-driven physics-aware modeling.

Empirical evidence of performance of these models will be presented through illustrative examples of vegetation/land monitoring involving multispectral (Landsat, MODIS) and passive microwave (SMOS, SMAP) observations, as well as blending data with radiative transfer models, such as PROSAIL, SCOPE and MODTRAN.

Statistical inference is at the heart of learning from data. Inference is not only about predictive capabilities, but about process understanding through observations. Compared to other data-driven empirical learning methods, such as probabilistic modeling, kernel machines, or deep learning, which mainly focus on prediction and classification, causal inference methods aim at discovering and explaining the causal structure of the underlying system. Essentially, causal discovery aims at extracting potential causal relationships from multivariate datasets, which goes beyond the commonly adopted correlation approach.

In recent years, new developments in statistics, machine learning, and philosophy, along with a rapidly increasing amount of observational and simulated data have paved the way to the use and adoption of novel data-driven causal inference methods. However, new causal inference methods appear very often, along with particular assumptions and hyperparameters, but are only compared on small sets of examples, and the practitioner does not understand subtle differences, good practices, and their limits. In addition, Earth system data has very particular properties featuring major challenges for such methods, so often methods are not well suited for particular Earth system causal problems.

We here review the field of causal inference for remote sensing and Earth observation data analysis and understanding. We will first discuss the main families of methods encountered in the literature, and characterize their advantages and shortcomings. Second, we will introduce four new causal methods in each one of the families: kernelized Granger causality, sensitivity measures in additive noise models, multivariate kernel conditional deviance measures, and unbiased CCM. We will illustrate performance through examples of applications in remote sensing that allows us to discover cause-effect relations from observational data (e.g. temperature, gross primary productivity, soil moisture, vegetation parameters), as well as to assess consistency and plausibility of radiative transfer model simulations -and emulations- of vegetation at leaf and canopy levels (PROSAIL, SCOPE). Finally, we will present a web platform called CAUSEME for benchmarking causal inference methods and solutions automatically.

Nowadays, thanks to the modern satellite constellations delivering high resolution optical data, we are able to study in deeper details many aspects of the Earth status. Also, due to the large amount of acquired data coming from several missions we can monitor constantly the Earth environmental condition.

Unfortunately, due to the physical nature of optical sensors, the acquired information suffers from several atmospheric natural events such as rain, clouds, fog, fire smoke and volcanic gases. All these phenomena, in particular the clouds, affect the availability of useful data for scientific purpose. Indeed, on average at any moment in the day, about 52% of Earth is usually covered by clouds. Some regions are almost always cloudy such as the Amazon Rainforest whilst some others are almost always clear such as the Sahara Desert. In addition to natural phenomena, we have to consider the revisit cycle of satellite too (for instance each of the Sentinel-2 satellite has an individual revisit cycle of 10 days at the equator with one satellite, and 5 days with 2 satellites under cloud-free conditions which results in 2-3 days at mid-latitudes) and that optical acquisitions are limited to daytime operations. These factors together reduce the amount of data available for any kind of application especially time series analysis over the same area. For example, in Northern Europe, the cloud free optical imagery decreases dramatically with respect to the total data acquired, affecting analysis based on optical time series.

Differently from optical data, that covers the range of 390 / 700 nanometres of the visible spectrum, SAR (Synthetic Aperture Radar) sensors exploits microwave wavelengths that are able to penetrate the clouds and so provide cloud-free products. These products, being the result of a reflected signal travelling not in the visible spectrum, cannot be rendered as natural colours images and usually are rendered as grey scale images or processed in order to produce an RGB image with fake colours highlighting specific features. This kind of colorization do not help visual interpretation in the same easily way as per optical counterpart.

Nowadays, in the Artificial Intelligence field, several studies, papers and applications are oriented in image and video restoration process (e.g. aged ancient photos or movies coming from black and white era) exploiting the possibility to inject missing colour inside monochrome images. Starting from this concept and considering a less variance in terms of possible different situation w.r.t to real-life scenes, the aim of the project is to provide a system, based on a Deep Neural Network, able to infer colour information inside SAR images to obtain synthetically reproduced optical data in natural colours totally cloud-free. The model is designed on a Convolutional Neural Network (CNN) trained with ad-hoc selected Sentinel-1 and Sentinel-2 (cloud free) products constituting our training and test data sets. Each single training example is based on a precise pair of Sentinel-1 (our input feature) and Sentinel-2 (our ground truth and expected output) products covering the same area and having their acquisition time as closest as possible to reduce feature differences. A pre-processing of the training is needed, such as co-registration, to achieve a perfect match between radar and optical data to train in the best way the Neural Network. As a preliminary case study, the area of Nile delta has been used considering the morphology of the landscape that includes deserts, urban and vegetated area, different water resources such as river and sea.

Thanks to this trained Neural Network, it is possible to achieve images in natural colours starting from almost any kind of radar images overcoming the problem of the lack cloud free optical images. Considering also to the huge amount of historical data currently preserved (e.g. ERS 1-2, ENVISAT ASAR and MERIS, Landsat just to name a few) it will be possible to apply our model to them to generate cloud free historical data sets. Future applications of the model developed, could be employed to recreate synthetically the missing information directly on optical images (due to cloud presence or others hiding phenomena) starting from radar information working on every different optical bands & wavelength.

The complexity of archaeological sites and religious monuments combined with the requirements for high accuracies, the need for 3D visualization of geospatial information in high resolution, the change detection and monitoring through time, require the development and implementation of contemporary techniques and procedures for multiple (spatial and qualitative) documentation. However, the available techniques and methods are either inadequate (e.g., the lack of 3D-GIS systems with 3D topology) or are time-consuming and require of high computing power (e.g., for the compilation of georeferenced photorealistic 3D modeling).
The Meteora is a rock formation in central Greece hosting one of the largest and most precipitously built complexes of Eastern Orthodox monasteries, second in importance only to Mount Athos. The six (of an original twenty four) monasteries are built on immense natural pillars and hill-like rounded boulders that dominate the local area. Meteora is included on the UNESCO World Heritage List under criteria I, II, IV, V and VII. The broader area is inhabited by an active monastic community of multi-century-long history, which is recognized as a Holy Area and an important archaeological site.
The METEORA research project aims to the integrated documentation of Meteora archaeological sites and religious monuments as an international research issue in Geoinformatics. In general, the METEORA project will provide an easy-to-use digital platform for multiple applications and potential users, supporting an interactive information system for a multi-dimensional documentation of Cultural Heritage sites under a 3D/4D modeling and augmented reality setting.
In this study, the data collection from various sources and processing is carried out. Simultaneous high resolution RGBNIR imagery of 8 cm resolution and LIDAR point clouds (Near Infrared & Green bands) of high density (40 points/m2) were acquired through an aircraft equipped also with an accurate GNNS/INS system. Terrestrial measurements were also, performed. The fusion of the aerial imagery with the LIDAR data contributed to the extraction of accurate 2D/3D photogrammetric products.

ο Building an ecosystem where every geospatial project can take shape, scale-up and succeed

sobloo is a cloud based solution hosted on a powerful & highly-reliable infrastructure designed to be a one-stop-shop for data users, developers, services providers & insights end-users. Through a secure cloud computing architecture, the sobloo platform provides a broad access to Copernicus Sentinel data, as well an extensive library of complementary data collections. Far beyond data, sobloo offers a wealth of value-added services to turn data into analytics, a wealth of tools & APIs and a full set of powerful GitLab features, along with a convenient desktop.

All users, public or private, are encouraged to take advantage of sobloo’s universal docking station to plug their own data, service, applications. Building the best ecosystem possible to help developers turning data into invaluable insights is our core priority.




ο Bringing together “Medium Resolution” and “Very-High Resolution” to unleash the potential of Earth Observation data

The main rationale for Copernicus is to ensure European Institutions, industry and citizens to get an easy access to useful and relevant medium and high-resolution data, in a free and open manner. Given the volume and variety of data it provides every day, the Copernicus programme is one of the key originators of the Big Data Analytics market within Earth observation. Integration & fusion of large amounts of multi-source information are nowadays considered to be the main challenges in remote sensing data exploitation.

Paving the way to a new paradigm in data fusion, sobloo signed a very strategic partnership with Airbus to provide its users with a privileged access to Airbus optical satellite constellation imagery by making SPOT 6, SPOT 7 and Pléiades catalogue fully integrated within sobloo data library. Therefore, sobloo becomes the one-stop-shop solution offering a living library of imagery with the largest combination of resolution and coverage. Beyond the comprehensive collection of Copernicus data, very-high resolution (VHR) brings sobloo to a new strategic step: every day, every point on the globe can be viewed both in high-resolution by a SPOT satellite and in very high-resolution by a Pléiades satellite. While SPOT 6 and SPOT 7 cover wide areas with a resolution of 1.5m, Pléiades 1A and 1B focus on more targeted zones with a greater level of detail by providing 50cm products.

By merging these huge collections in a unique ecosystem, sobloo tackles the main challenge of geospatial digital industry: building a new paradigm for geospatial multimodal data fusion. Observations derived by different acquisition sources can be coupled and jointly analyzed by data fusion practices to achieve a richer description of any scene. This becomes a key asset to improve insights in a wide range of application areas, among which characterization of land use, damage detection (e.g. disasters, oil-spills), monitoring of natural resources, or climate change.

While removing the need to download bulky files from several access points and process them locally, sobloo lowers the complexity of data integration from heterogeneous sources, which was one of the main barriers for value-added service providers. sobloo enables them to build their own applications faster, in a secured and universal ecosystem. On sobloo, service providers can access tailored EO products with real added value to their business: they can obtain the right imagery in the resolution needed to meet their specific requirement at any time through to a trustful highly-secured gate.

Thanks to multi-mission capabilities, large-scale image processing capacity & and a performant computing environment, it becomes almost effortless to develop new applications & services.

Thus, citizens, policy makers, researchers, commercial users, private and public users, scientific actors can benefit from smarter insights resulting from an improved geospatial data inclusion.

Water resources managers, responsible for hydropower reservoirs consult periodically collected information from different sources, such as: detailed data about the condition of the distinct hydropower reservoirs, water economic data, meteorological data and forecasts, geographical and geospatial information about the water reservoirs and their contingent environment. Satellite data about liquid and solid precipitations, sunshine intensity, soil moisture, air temperature, vegetation index, to name a few, are instrumental for building reliable water inflow forecasting models when mixing them with in-situ measurements. They are all presented as numeric data and measurement values. However, when analyzing a particular situation, for instance preparing a monthly report about the status of the managed dams, it becomes necessary to combine these numeric values with locations, geographic concepts, elements of hydropower reservoirs, temporal information, and compose queries that retrieve data based on these heterogeneous facets. Questions of interest in this context can be “What is the highest water level of Cascade Arda for April for the period 2013-2017”, “What is the expected inflow for Cascade Arda in June 2019”, “What is the biggest snow stock for Cascade Arda in the last 10 years”, “What was the precipitations volume and the sunshine intensity at the time of the biggest snow stock for Cascade Arda within the last 10 years”, “Water balance of Cascade Arda for 2016”, ““Gate opening for dam Ivaylovgrad when inflow is 222 m3/s”, etc. To answer these questions an infrastructure that integrates and interprets all types of data mentioned above is necessary. Linked data technologies offer an optimal standardized framework to deal with this issue, as they allow easy data integration, resource economy and semantic search capability, that allows the composition of queries with conceptual components and the retrieval of results from heterogeneous data. We present semantic search over a linked data infrastructure that semantically integrates data by means of specially crafted water management ontology based on INSPIRE Data Specification on Hydrography , that includes semantic elements from GIS, satellite data, proprietary measurement data and domain knowledge. It allows interconnection of and logical reasoning over the introduced semantic elements and generation of additional relevant related facts that become available for querying. We demonstrate the potential of the linked data infrastructure with a series of semantic queries that relate to and retrieve explicit and implicit facts, use GeoSPARQL , and interoperate with a GIS component (ESRI ArcGis Server ) to display relevant for the query results geospatial information. We compare the proposed semantic search with non-semantic search capabilities, and discuss the interest of employing it and linked data for hydropower reservoirs exploitation and for water management in general.

Similarity matrix shows the similarity degree between each data pairs, it actually plays a core role in a number of dimensionality reduction methods, since the objective function builds upon this matrix. While compression-based similarity measures are effectively employed in applications on diverse data types as basically parameter free approach, a fast compression distance (FCD) metric has been proved to be able to achieve similar classification performance comparing with the Normalized Compression Distance (NCD) method, regarding small- to medium- size datasets [1].
The FCD is claimed as combining a fast speed without skipping the joint compression step which obtains better performance compared with NCD [2]. The idea behind is: the LZW algorithm extracts a dictionary D(x) from each image patch, and encode into a string x, in ascending order. The definition of FCD is defined as an operation which mainly takes account of the joint number of patterns within two dictionaries D(x) and D(y).
In this research, we use FCD together with t-SNE to visualize a large semantic annotated TerraSAR-X dataset as a study case. The dataset contain image patches from 288 TerraSAR-X images with a total number of over 60,000 individual image patches.
The visualizations represent the annotated semantic labels in such an intuitive way which helps us to better understand the relationships between their annotated semantics and how their actual similarities are in manifold space. Our obtained results show that the FCD based similarity matrix effectively provides us a fast yet performance preserved insights in high-dimensional datasets with a non-parametric distance metric. Via the visualization on TerraSAR-X dataset, we have gained quick intuition and better understanding of the connections between the annotated semantics and the relationships within the data which is revealed as similarities in manifold space.
Change Detection methods are dependent on the extracted image features and measures of similarity used for the comparison of the observed scene at different time moment. The NCD has an important advantage; it does not use features and compares the intrinsic data information. The change detection is thus an un-polarized estimator for temporal changes. The method is validated on two areas with visible changes such as flooding and tsunami effects. The results are compared with the ground truth data.
The influence of natural disasters, as well as climate warming of the global environment increased in the past decades. Therefore, the detection of changes in a satellite image time series is a trivial task [3]. In order to do this, the calculation processes can be divided into several phases. The first represents the preprocessing, which includes an alignment of all images as well as the creation of patches in a region of interest. We propose here to use the Sentinel-1 SAR data. The second phase encloses the generation of a distance matrix of the patches from two images. In the last phase a threshold is applied to the created matrix, in order to show the changes in a binary way as binary change map (BCM). The results show that a compression-based approach is working with very good results on SAR data. Moreover, a visual evaluation of the resulting images shows that the compression-based approach detects the flooded areas within a region. Furthermore, if the input is reordered with the Burrows and Wheeler transformation [4], the resulting image is better in some areas that the base method. This shows the robustness of NCD.


[1] C. Daniele and M. Datcu, “A fast compression-based similarity measure with applications to content-based image retrieval,” Journal of Visual Communication and ImageRepresentation, vol. 23, pp. 293–302, February 2012.
[2] M. Li, X. Chen, X. Li, Ma. B., and P.M.B. Vitanyi, “The similarity metric,” IEEE Transaction of Information Theory, vol. 50, pp. 3250–3264, 2004.
[3] M. Coca, A. Anghel, M. Datcu, “Normalized compression distance for SAR image change Detection”, pp. 1-3, 2018.
[4] R. Giancarlo, A. Restivo, M. Sciortino, “From first principles to the Burrows and Wheeler transform and beyond, via combinatorial optimization”, Elsevier Theoretical Computer Science, vol. 387, pp. 236-248, 2007.

Mangroves ecosystems are distributed across the tropical and subtropical regions of the world and provide important ecosystem goods and services to nature and society. Mangrove forests are under threat from both natural and anthropogenic forces. Companies and consumers increasingly demand that source industries develop commodities without negative impacts to the environment and particularly to prevent deforestation. Sustainable supply chain initiatives include fisheries and aquaculture, for example the “GLOBALG.A.P” aquaculture standard and Global Sustainable Seafood Initiative, which require that aquaculture is not established within mangrove ecosystems. Worldwide, aquaculture sustainability is itself impacted by climate change, competition for space and natural resources, pollution and environmental degradation.
Effective management and sustainability of the aquaculture sector requires appropriate and accessible data to inform management decisions. Satellite earth observation technology has developed rapidly in the past five years, with numerous advances in spatial resolution, spectral information, frequency of observation, and increased access to open-data. While the growing volume of satellite data represents a unique opportunity for sustainable aquaculture development, the acquisition, processing, classification and interpretation of remotely sensed data is time consuming, computationally demanding and involves methods which require expertise. As a result, achieving the full potential of these data to extract valuable management information poses a major challenge. A variety of big-data platforms currently seek to address these issues by bringing together direct access to satellite data, and analytics processing systems that can be scaled horizontally on-demand. These platforms include the European Space Agency funded Coastal and Food Security Thematic Exploitation Platforms and Google’s Earth Engine.
In response to the aquaculture sustainable supply chain requirements and data management challenges, we developed a mangrove mapping and change detection algorithm written in python and running within a Docker container environment. The Docker container was deployed on the ESA-supported Food Security, hosted within the Creodias platform that provides access to Copernicus data and processing resources. The portable, containerized application collects the imagery over the user-defined study area and creates seamless, cloud-free, best available pixel Sentinel-2 composites and calibrated Sentinel-1 cross polarization images for a given season. Multiple Sentinel-1/2 derived metrics that are sensitive to discriminate mangroves are derived from the image composites and used as predictor variables for a Random Forest (RF) machine learning model. To train the RF model we used thousands of reference data points in two study areas – Tanzania and Indonesia. The exercise covered the full Tanzania’s coastline, or 13 Sentinel-2 tiles, and the analysis of three consecutive years for more than 600 Sentinel-2 images. Out of bag error estimates for Tanzania indicate mangrove prediction accuracy in excess of 90% and both Sentinel-2 and Sentinel-1 metrics were important predictors for the model
The development and deployment of a mangrove mapping application demonstrates how the new paradigm of open data, portable containerized algorithms, and scalable processing platforms can enable earth observations services to address emerging industry requirements. World aquaculture production continues to grow and now provides half of all fish for human consumption according to the United Nations Food and Agriculture Organization (FAO) – this growth demands innovative new tools and methods for more efficient monitoring and improved environmental stewardship.

In 2018 ESA published a tender for developing EO services to support the resilience and sustainability of critical infrastructures and to make such infrastructures “greener”.
Planetek Italia, as prime contractor, together with ENEA and INGV as sub-contractors selected the “integrated water cycle management” as object for their feasibility study. The project, whose name is SCIRES - Supporting Critical Infrastructures Resilience from Space, is currently on going.
The integrated water cycle management relies on different infrastructures: water basin and impoundment, transport and distribution pipelines, wastewater collection, and as interconnected infrastructures, power distribution networks on which the infrastructures rely on for energy supply, and power plants that exploits the path of the water to produce energy. Two innovative EO services, eventually integrated into the CIPCast system (an existing Decision Support System developed by ENEA) are object of the study: (1) impoundment stability service and (2) pipeline monitoring service
1. The Impoundment Stability service
The service is focused on the monitoring by means of mean ground velocity and time series of deformation (Persistent Scatterers) computed by Rheticus®’ services, integrated with in situ geological and geomorphological data. The objective of the service is to provide a comprehensive analysis of horizontal and vertical surface and near sub-surface deformations and dislocations that affect hydraulic works, such as water impoundments and distribution networks that could be damaged by hazardous events. In particular, landslides (LS), deep-seated gravitational slope deformation (DGSD), and differential ground subsidence (GS) phenomena are taken into account. Both LS, DGSD and GS can damage or threaten the functionalities of the infrastructures itself (e.g. excessive pipeline deformations, differential subsidence close or below hydraulic works), especially if triggered by heavy rains or earthquakes.
2. Pipeline monitoring by using EO data, IoT (in-situ sensors) and AI technique
The service aims at identifying precursor phenomena related to the breaking of pipeline, responsible for the losses of water, and subsequent water infiltration in the ground and potentially causing damages to the surrounding infrastructures. The goal is to prioritise the maintenance of the pipelines, basing the priority on the computed vulnerability and on the subsequent risk computed in the CIPCast DSS.
Damages to pipelines may be caused by several factors, coming from the environment (weather, subsidence-causing stress on the ground…) or from the working condition of the network (pressure and velocity of the fluids and relevant gradients…).
Many statistic models have been implemented to correlate the probability of breaks, causing losses, to environmental and endogenous parameters. In this scenario, it is possible to insert an element of innovation based on the concept of Artificial Intelligence.
Artificial intelligence (intelligence demonstrated by machine) contains several methodologies, algorithms and strategies that can “instruct” the machine to execute some operations autonomously. One of the most important characteristic of AI methodologies is the ability to extract useful information from data that means, highlighting only the relevant data. Therefore, in the case of PSI, this kind of analysis is particularly important due to the enormous amount of available measured points. These methodologies can be found in fields like dimensionality reduction, clustering and anomaly detection.
To reach the best possible accuracy, when you want to monitor the ground movement of a certain area, is necessary to improve the displacements analysis by observing the dynamical behaviour, not only of a single point but also, of a set of neighbour points that can describe the condition of a certain area on the network. Indeed, the analysis of a single PSI done independently from the others, it could be not enough to describe completely the mechanism of ground movement reaching the proper accuracy over a selected area. For this reason, the understanding of how combining a group (cluster) of PSI, in terms of spatio-temporal analysis, is an important challenge in this field. In this context, artificial intelligence can help to achieve this improvement.
Artificial Intelligence can provide methods and algorithms that are able to augment the knowledge of PS movements. This is a fundamental aspect when we want to characterise the instability of a certain area using information coming from PS. The service will be integrate AI methodologies in order to highlight only the points that have “real” information in fact, one of the most important feature of some AI algorithm, is the ability to reduce data dimension without reduce the information, the core concepts of dimensionality reduction strategies.
AI algorithms for PSI analysis can therefore be adopted for extract points, whose behaviour are “similar” in terms of some selected properties. In fact, there are points that present similar movements where the term “similar” is used to indicate different point features (like displacement time series, average velocity, acceleration and amplitude of the signal or combination of them). Therefore, as first analysis, is possible use this kind of methods to assemble PSI (PSI cluster) improving the ground movement physical interpretation in order to better assess the effect on the pipelines infrastructures stability. In fact, the movement of one measured point in a stationary state area, could have a different impact on the pipelines infrastructure stability with respect to the same point that is in moving area. Combining dimensionality reduction techniques and cluster approaches, we can “draw” the different zones and inspect these zones separately.
Once clusters have been created, more accurate analysis can be performed in the selected area. In fact, the movement of one PSI, can be compared with other PSI that have the same velocity and then, the displacement interpretation and its correlation with the pipeline stability can be improved. It is clear that, as mentioned before, there are different artificial intelligence methodologies that can be used for this kind of analysis but, methods like PCA or SVD for dimensionality reduction and K-means or SVM for clustering, are the most relevant to reinforce the basis of displacement analysis.
Using clustering methodologies is possible divide big zone, with a huge number of PS (greater than 10 millions) and study these “different” zones separately. The aim is to characterize the pipeline using all available information on PSI. The final step is to see each PSI like a time series and then, is possible to study anomalies in the time. To conclude, a combination of clustering, dimensionality reduction and anomaly detection algorithms can improve the statistical methods used in the services.

The aim:
Funded under the Innonetwork Initiative of the Apulia Region, DECiSION (Data-drivEn Customer Service InnovatiON) is an industrial research project aimed at the development of a knowledge and customer service system, based on machine learning and deep learning methodologies.
The project proposes an innovative solution in the field of Information Seeking Support Systems (ISSS). The aim is to recover all the data related to a decision-making process, managing and making them useful to address the user requests. DECiSION will implement an awareness ability to understand and to interpret the user’s request made in natural language. DECiSION will be able to identify the information sources from which it recovers, in an automatic way, the information needs for the sense-making task. For instance, the System will be able to manage several data sources in order to provide, to the technical users, a response combining natural language with infographics and geographical maps, to address a user’s question as: "Which are the pipelines that need maintenance based on - the weather conditions and the stability of the territory? Which of these have a greater impact on the citizens in the event of an accident? ”. The project implements a chat-bot, which will play the role of virtual assistant, and conversational recommender systems, able to dialogue with the user to discover their needs and direct answers in a personalized way. Therefore, the main goal is to create an intelligent system capable of responding, autonomously and comprehensively, questions asked by the user in natural language, regarding a specific reference domain in order to support the decision-making process.
The idea:
The project intends to develop knowledge, technologies, algorithms, interfaces, all organised in a software platform that will allow to interrogate / interact / consult heterogeneous data sources, to get information to satisfy the user’s needs in line with the profile of the user / citizen / organisation. The question of the users, made in natural language, will start the interpretation process of the expressed need and activate the generation of multi-format responses that can satisfy it. The system is composed of various modules that allow the request to be analysed semantically. Through the analysis of heterogeneous data sources, the system elaborates the answers in dynamic, customized and contextualised formats. These answers contain information "enriched" by new knowledge, generated through processes of artificial intelligence and automatic reasoning.

The DECiSION project idea is to fulfil the lack of tools to access to structured and unstructured service information to provide a support for decision-making processes. Decision-making is a complex process, which requires the integration of heterogeneous information generated by different sources, which must be interpreted to define its knowledge value. In complex decision situations the problems are not well structured or clearly defined, these conditions can produce uncertainty. For this is necessary to consider that this system can understand and process a large amount of information, that helps a single user to address this type of problem. The project aims to achieve a decision support system that can be used through an artificial dialogue system able to query heterogeneous data sources and provide relevant and complete answers to the user.
The technologies:
The main element of DECiSION is the adoption of strategies, in a decision-making process, that takes care of time and space constraints collected from the context of the request, to generate the right answers for the user. Fig. 1 shows the diagram of the logical-functional model, developed and detailed in the various design activities planned in the project. The capabilities of the System to interpret heterogeneous data is managed in block "a". It provides the processing of heterogeneous data sources in order to extract the content necessary for the semantic interpretation of the information. The enriched indexing component (block "b"), represents the layer for "normalization" and "enrichment" of the annotations, to allow semantic interpretation.
(Figure 1 - Architectural model of the DECiSION platform)
This process uses ontologies to activate processes of reasoning and creating new knowledge. The components in the blocks "c", "d" and "e" have the responsibility to manage all the processes related to the interpretation of the request and the execution of the appropriate queries on the data services and other available data sources.

Once the information needs are understood, the query management module (labelled with "c") will automatically build and execute the queries to obtain the information useful for the composition of the answers. The "f" block has the task for combining the results of the queries to obtain a personalized response, in terms of content and layout. The recommendation component (block “g”) supports the decision-making process of the user from which the request is derived, integrated with the response composition process (block “f”). It will produce possible relevant suggestions in respect to the needs extracted from the request (for example, available services that can answer to an expressed user need). The component will use algorithms that will lead the suggestions to generate socially responsible behaviour. The front-end component (block "h") is designed to communicate to the user the answer in a simple and effective way, and will allow the use of the system through mobile applications, chat-bots or more generally user-friendly interfaces on different devices.
The innovation:
The question interpretation module represents one of the most innovative components of the project. The module is able to process the natural language extracting all its syntactic and semantic information. Then, using this annotation level, this module is able to identify the entities and relationships involved. In this second level of analysis the application identifies, the action that the user wants to perform, and any temporal, geographical or other constraints concerning the subject of the user request.
The language analysis and disambiguation process consists of a series of tools organised within a pipeline for natural language text analysis.
In particular, it is able to:
• Carry out a syntactic analysis of the text where each node is a word and the connection between the words indicate the dependence that interchange between two terms, for instance: subject, object and predicate.
• Carry out the disambiguation of the text assigning the correct meaning to each word.
• Perform entity linking between the words (people, places, organisations) and the concept of the base of knowledge.

The Land Cover is one of the major reference data-sets necessary for almost any kind of Earth Science applications. A land-cover product for the Bulgarian Black Sea Coastal Zone, as part of the information layers of the SatWebMare system developed under an ESA-PECS project, is considered to be generated. It will be used as a reference data-set in the analysis, modeling and visualization of the affected territories by various natural and man-induced hazards along Bulgarian Black Sea coast. The main goal of this poster is to present the preliminary results of Land Cover layer over Bulgarian Black Sea coast generated from multi-temporal Copernicus satellite imagery using the complementary Cloud Infrastructure of the Bulgarian Academy of Sciences, i.e. Avitohol Supercomputer and Google Earth Engine.
The used data-sets for the case study areas of Bulgarian Black Sea coastline and North Bulgarian Black Sea coastline and adjacent mainland are Copernicus optical and radar data from Sentinel-1 and Sentinel-2. Different image classification algorithms, such as Maximum Likelihood (ML), Random Forest (RF), Neural Networks (NN), and Support Vector Machines (SVM), are tested over multi-temporal imagery consisting of Intensity image from S1 and reflectance bands of S2 with 10 m and 20 m spatial resolution. The achieved results are tested against accuracy of the high-resolution ESA CCI Land Cover and CORINE Land Cover (CLC2018) databases extracted over Bulgaria which performs similar to the results achieved with an expense of the lower spatial resolution of the latter. The target mapping accuracies of above 85 % is achieved by all image classification algorithms with slightly better results of above 94 % for RF. In conclusion of the validation study performed and presented in this poster is provided a discussion with suggestions and recommendations for improvement of the land cover products over the coastal territory of Bulgaria in view with the coastal management and monitoring.

Deep learning is revolutionizing the way we solve problems whose solutions were previously within the grasp of only experts and seasoned professionals. Nonetheless, its use still requires specific knowledge and tools in order to be effectively implemented in real applications. Notably, certain large scale activity areas which could stand to greatly benefit from this technology, such as agriculture, still don’t use this new tool. We propose in this study to develop agriculture related deep learning algorithms that could help companies and farmers by bringing them new useful information. In fact, predicting the yield of crops can be particularly difficult as information about surrounding crops may not be readily accessible. But, remote sensing and deep learning could bring new data to farmers, data that could be essential in their decision-making process. Deep learning algorithms now bring new ways of detecting complex settings by using only remote sensing data; however, they are, as of yet, too complex to use for most users. That is where the Sentinel platform allows for the construction of earth observation processing chains that are easy to use while also allowing experts to add any processing. Using the tools available, we want to show that deep learning methods can be used to produce reliable results in crop detection problems. We firsts create processing chains that download, cut, filter and save deep learning datasets. Visualizations and statistics are then added to control the quality of said data. Using the Keras and Tensorflow frameworks, a RefineNet and U-net neural network is trained to produce crop-type maps using only Sentinel-2 data. The resulting neural network is included in another processing chain. This new chain allows the user to quickly predict crop type maps at any given coordinates by simply giving a polygon as input of the resulting processing chain. This processing chain can now be used inside a web-service to be made available to farmers or by scientist to study the state and evolution of crops in a certain area.

Cloud detection and segmentation in remote sensing represents valuable information for assessing the image quality, for image analysis tasks and also for localizing the regions of interest. Various applications in meteorology, agriculture, maritime surveillance can be based on this meta-information to improve analysis results. This information is also essential for weather prediction or evaluating climate change.

On the other hand, deep convolutional neural networks have led to a series of breakthroughs for image processing task as classification, detection, segmentation and image generation. Recently promising results on multi-level cloud detection have been achieved with convolutional neural networks (CNN). It was shown that segmentation CNN architectures are well-suited to process multi-spectral remote sensing data and to classify clouds.

This poster presents our approach to tackle cloud segmentation task using Convolutional Neural Networks. Our goal is twofold : apply these methods on SENTINEL-2's multi-spectral data (Copernicus program) and extend the learning on other sensors (SPOT-6/Pléiades).

In this work, at first we successfully trained and validated state-of-the-art neural networks : U-Net like (backbone ResNet, ResNeXt, DPN), DeepLabV3+, ENet, RefineNet for cloud segmentation task on SENTINEL-2 data. Our training and validation datasets are based on the data built up by Hollstein et al.

In the second part of the work, we evaluated fine-tuning and semi-supervised domain adaptation method "self-training" aimed to generalize on SPOT-6/Pléiades sensors.

Our implementation is based on open-source deep learning framework Pytorch. It also benefits from the latest implementation of the loss functions (e.g. Lovasz-softmax loss), data augmentation and training libraries. We also adapted for our task an open-source self-training method for domain adaptation.

Pollution of sea and ocean surfaces due to oil has become a serious worldwide environmental problem over the past few decades. Detection through SAR image analysis is now widely used for maritime surveillance and monitoring.

Satellite SAR imaging is very effective for sea monitoring and oil spill detection due to its day-night acquisition capabilities and its insensitivity to atmosphere conditions. Under favorable wind conditions, oil surface roughness is much smaller than that of water, preventing the Bragg scattering mechanism. Therefore, the SAR electromagnetic waves are not strongly backscattered by oil spill surfaces, which therefore appear much darker than the surrounding water surface. In general, VV polarizations are preferred because they allow a better contrast between water and oil spills than HH ones.
However, there are some challenging aspects that cannot be underestimated:
• In areas with low-wind conditions (typically for wind velocity less than 2 m/s), the Bragg mechanism is not active and the low backscattering of the water surface does not allow an easy separation from oil spills;
• Under very high wind conditions, oil surfaces become rough enough to be indistinguishable from water;
• Dark areas on sea surfaces can be originated by natural sources as upwelling and seepage of oil from sea bottom.
The task of detecting oil spills, which is well accomplished by photointerpretation, can actually be a challenging task for fully automatic algorithms based on classical computer vision techniques such as filtering and thresholding of images, since many factors must be considered: contrast between oil and water , shape, roughness border. In addition to the difficulties of extracting oil spills from a complex marine scene, the different types and aspects of oil spills in different weather and marine conditions and in different SAR modes and resolutions make the task of automatically detecting oil spills even more difficult.
Many studies exist for the automatic detection of hydrocarbon spills based on machine learning approaches that require preliminary segmentation and feature extraction. Recently, several convolutional neural networks for semantic segmentation purposes, like SegNet and Unet, have been developed.
In this work, we propose a new CNN architecture that implements semantic segmentation. This development is based on a very large set of training data. In fact, thanks to our long experiences as a service provider for EMSA Clean Sea Net (CSN) and for similar services for other customers (AMSA, oil and gas companies), we have collected a very large dataset of SAR images (COSMO SkyMed, Radarsat , Sentinel-1) labeled with classification and confidence indices. Moreover, the portions of sea with oil spills in each image are exactly delimited, which allows effective pixel-wise training of our semantic segmentation.
The preliminary results we have obtained are very promising, and confirm the validity of the proposed approach.

Soil underpins 90% of human resources of all food, fiber and fuel production and it is essential for water and ecosystems conservation. It is a global carbon sink, playing an important role in the potential slowing of climate change (EEA, 2016). These characteristics have led to high requirements in terms of detailed soil information, which is mainly provided by agrologic surveys.

Initiatives such as the Global Soil Partnership, conducted by the United Nations Food and Agriculture Organization (FAO) highlight the importance of accessible information. Thus, the availability of documented, standardized and interoperable agrological information is an imperative condition to plan and carry out any research/ productive agriculture project in Colombia.

The soil information system for the Cundiboyacense highlands, IRAKA (indigenous word that refers to the sacred city of the Muisca culture), is an service that aims to provide soil information to improve knowledge of the soil resource, guide management practices, promote the development of agriculture in a sustainable manner and implement actions aimed at guaranteeing the food security of territories. IRAKA systematizes secondary information, from the general and semi-detailed soil surveys from the Instituto Geográfico Agustín Codazzi (IGAC) and in-situ samples from the soil laboratory of The Colombian Corporation for Agricultural Research – AGROSAVIA. These data are then used to develop models of the spatial distribution of physical and chemical properties of soil (pH, Organic matter NOM, Cation-exchange capacity CEC, Bulk density, among others), using machine learning techniques. The outputs of these models are used in the construction of a soil quality index for the cultivation of welsh onion (Allium fistulosum L.) and bulb onion (Allium cepa), which is validated through field samplings and experimental tests in a greenhouse.

The IRAKA system interface is developed under a Visual Studio version 2017 environment, through the programming language .NET C # with MVC (Model View Controller). Likewise, the geospatial information visualization service is delivered using the open source web server GeoServer version 2.15 and the OpenLayers 5.3 library. The data is stored in a database made up of 34 objects, of which 6 have a spatial attribute and 1,716 records. The database design stage began with the definition of the data models (logical, conceptual and physical). Then, came the identification of the domains and the provision of data to the final users, which was performed using a PostgreSQL 10.6 object database management system (ODBMS), with a PostGIS 2.4.3 module that supports geographic objects.

Currently, the system is available at the following link: http://paginad.corpoica.org.co:8086/NetCorpoicaMVC/Iraka#

Floods and in particular coastal floods are natural disasters having a great impact worldwide, in terms of both frequency and damage costs. There are currently operational services based on remote sensed data aiming to estimate the flood extent and the early damage assessment. For instance, the early flood detection and characterization based on satellite data is one of the core elements of the Copernicus Emergency Management Service (CEMS) to provide geospatial information within hours or days from the activation, in support of emergency management activities immediately following a disaster.

The main approach to detect flooded areas based on remote sensed image is to exploit satellite Synthetic Aperture Radar (SAR) systems, which provide global and continuous monitoring capability in almost all-weather conditions and day-and-night acquisitions. SAR data offer a strong capability to highlight flooded areas thanks to the high contrast of backscattering between flooded areas and water-free surfaces.
Multispectral optical data can be used to support the analysis and to complete the flood extent definition if weather conditions are favourable, but generally SAR data are preferred for early monitoring purposes.

The main idea underlying the technique proposed within this work is the exploitation of a priori knowledge of permanent water bodies in the acquired scene to generate reliable training and test data sets in a completely automatic way. The AI model training and the reliability calculation (test) are performed using satellite sensor data available at the processing time. For this purpose, each one of the following cases can be considered:
- A single SAR or multispectral optical acquisition post-event;
- A couple SAR and multispectral optical acquisitions post-event;
- A couple of SAR data with pre and post event acquisitions.
The above cases have advantages and disadvantages which have been analyzed in this study. For instance, the use of a couple of pre and post event SAR acquisitions allows the removal of false alarms due to dark areas related to pre-existing water bodies and to not flooded areas as aerodrome runway.

The developed technique is based on the application of a support vector machine classification (OneClassSVM, based on “Support Vector Method for Novelty Detection” by Bernhard Scholkopf) trained with samples of surface water surface, extracted from the input images using open source layers as OpenStreetMap and Global Surface Water data sets. Following this strategy, the training and test dataset are generated automatically by labelling samples extracted from input data with the information of the co-located targets contained in the open source data sets. This approach is optimized to work in coastal areas on which it is expected to find many pixels of water surface to be used for training and for accuracy assessment (test).
Both sea and internal waters are used to generate the training data set. Since open seawater surface is more stressed by winds than coastal and internal waters, the water samples are extracted mostly from internal waters and sea areas close to the coast in order to be representative of the inland flooded areas.

Data visually interpreted by expert operators within the CEMS Rapid Mapping service are being used to validate the performance of this approach, and the first results are promising and confirm the validity of the proposed approach. The challenges are in the selection of the most suitable samples to train the classification algorithm and finding an effective way to choose the most effective hyper-parameters for the SVM to operate proficiently.

Agriculture is one of the least digitised and thus most obvious industries with the potential to benefit from EO data. Factors such as climate change, population growth and increasing food insecurity call for innovative approaches to protect and improve crop yield. Recent adaptation of data analytics into the EO domain of using AI, deep-learning and machine learning models help to better map, monitor and predict agricultural assets.
The transformation of the EO data into derived information of value to various players requires the development and advancement of numerous “component” services, such as refined crop type determination, field parcel delineation, crop growth assessment and soil-moisture determination. Furthermore, natural hazard mapping, crop yield forecasts and agricultural statistics are key issues that can be addressed with such services.
Several business-to-business opportunities arise with the advent of the unprecedented observations collected by the Sentinel satellite family such as operational risk and impact assessments for the insurance sector or the support of global agricultural and rural development activities by Multilateral Development Banks and their national stakeholders. Further key customer groups are public institutions, paying agencies and farmers (e.g. CAP monitoring, precision farming), the financial sector (e.g. agro-insurances and commodity traders), as well as the AgroFood industry. The latter is a key sector to benefit from EO data, however, it is as yet little used to support other elements of the AgroFood industry beyond the farmer, which contribute a significant proportion of the added value in the entire “farm-to-fork” chain and its support web. For example, considering the potato industry in particular, there is potential for EO data to improve disease control, logistics, storage, trade, and insurance amongst other services.
A newly developed, scalable system will provide the means to integrate satellite data and other information streams directly in existing end-user workflows in a business-friendly format, starting with the AgroFood sector. EO PLUG-IN is a new project supported by ESA’s “Investing in Industrial Innovation in Earth Observation” (InCubed) initiative, the first of its kind for the newly founded ESA Phi-lab. The project will provide a system of user enabled onboarding to ready-to-use satellite data-based information streams for the AgroFood industry to support, amongst others: business operations, production management, and operational efficiency.
A key component is to build a scalable processing infrastructure and an API, which turns satellite and ancillary data streams into valuable derived information solutions, delivered directly to the commercial end users who evoke a specific data stream. A generalised and flexible infrastructure is being developed to provide a number of specific services initially requested by and designed in collaboration with AgroFood industry partners in the Netherlands and to be rolled-out globally. The setup of this service is however in a way that it can be also expanded to further customers and topics, thereby growing the opportunities of the Earth observation industry in the agricultural sector.

The aim of the Q-COLT project is to develop a Quality Control (QC) software tool to use with optical Earth Observation data that will employ machine learning (ML) and deep learning (DL) algorithms and techniques to support the detection of anomalies within the data products.
Q-COLT extracts and stores a database of relevant information from the satellite data under investigation that are used by the QC Operator as part of their inspections, supporting their analysis of the products. The first release of the tool will focus on Landsat data products, and will contain the relevant Landsat metadata files, quicklook (QL) images for inspection and AMALFI quality reports (a pre-existing QC tool).
Initially, the QC operator will use the Q-COLT tool to inspect and tag/log anomalies within the data products. This activity will help to generate a training data set that can be deployed to develop and teach the machine learning algorithms. Once at a suitable state of maturity, these algorithms will enable the Q-COLT tool to then automatically flag particular products if it detects an anomaly, triggering a deeper inspection by the QC Operator to determine if the anomaly has been correctly assigned. Ultimately, the objective will be to create a semi-autonomous EO data QC system, able to analyze large datasets for anomalies in a bulk manner without the need for a manual inspection of each product.
Q-COLT has a user interface that allows the QC Operator to search all the products available in the database and list all their specific attributes. For each of the products, a QC interface is provided to analyze the data and provide different options to detect and log possible anomalies. Each of the products has metadata associated with different information to allow the QC Operator to have all the relevant information available for better understanding of the anomaly.
Q-COLT enables the QC Operator to create specific anomalies classes, assign these anomalies to a product, inspect and overlay browse and quality quicklook (QL) images of the data, zoom/pan around linked QLs, add comments, display selected metadata information alongside the QLs, and load pages containing metadata & AMALFI quality control report for each product.
In literature, anomaly detection is usually called novelty detection, which refers to the identification of novel or abnormal patterns embedded in a large amount of data. These types of problems may be seen as "one-class classification", in which a model is built to describe "normal" training data. Novelty detection approach is used when the quantity of available "abnormal" data is insufficient to construct explicit models for non-normal classes. Traditionally anomaly detection has been done using statistical methods, either via one-dimensional normal distribution, or multidimensional case. Recently new techniques to detect anomalies have come up based on ML/DL in particular Neural Networks. Despite these techniques have been around for decades, two factors have made them to be reconsidered: (1) the amount of data available; (2) the reducing cost on computational resources based on GPU allowing fast linear algebra operations.
Recently, more sophisticated ML/DL approaches are promising to tackle multidimensional data. In particular Q-COLT has chosen Convolutional Neural Networks (CNN) as a candidate to detect anomaly features on images in order to detect strange features easily identifiable by a human from a visual inspection. It is known that CNN does not work well in case of unbalanced classes, however enrichment data methods can be used to cope with this problem. Computer vision methods such as Pyramiding or Blurring techniques will allow to produce square chips on images that can be feed into the CNN for its training data set and complete the unbalanced classes.
In ML/DL data preparation phase, new techniques have been proposed to increase the number of samples available thus this will reduce the effect of unbalanced classes such as noise perturbations and flickering of samples. Additionally, labelling based on polygonal selections will be implemented via graphic user interface in order to produce anomalous cases. Due to most of the data being unlabeled, Activate Learning is proposed to optimize labelling tasks; in statistics this is called Optimal Experimental Design. In terms of ML/DL algorithms, the users interact with system models that have been trained on a limited labelled data set. The most difficult (undecided) samples for the latter models are presented by the system to the users in order to be classified. Then the models are retrained with the new user labels closing the loop of Activate Learning. This technique allows models to not get obsolete by very novelty anomalies unseen by the system.
The Q-COLT software aims to become a valuable tool for EO QC engineers in the detection of anomalies of increasingly larger and larger datasets. By itself, it is a useful tool to bring together the relevant information to enable a quality assessment to be performed on a series of data products, and log specific issues found. However, by leveraging the effort and activity already performed by QC engineers to develop and train ML and DL algorithms, it has the potential to increase both their effectiveness and efficiency in identifying anomalous data.

Icebergs are large pieces of ice that calve off glaciers and ice shelves. They are generally composed of compacted snow and are thus of freshwater origin. Once calved into the sea, icebergs are considered a threat to shipping and marine operations. One of the tools that are used to monitor iceberg’s movement is Synthetic Aperture Radar (SAR). Like vessels, icebergs manifest themselves are bright targets in SAR imagery. The SAR response from icebergs is generally complicated due to the random geometric shapes of bergs. To discriminate icebergs from other ocean targets such as sea ice and vessels, automated classification is important. This is particularly important in an age where SAR imagery volumes are immense and difficult to process using the traditional imagery analyst approach. Therefore, a study of polarimetric properties of icebergs is vital to optimize the surveillance of icebergs and enhance iceberg detection in the presence of background clutter. Polarimetric decompositions have been extensively used to study the physical property of scattering behaviour of remotely sensed target for decades. In the case of icebergs, there has not been a systematic study of the utility of various decomposition techniques for iceberg analysis. To look for a suitable technique, several extensively used decompositions have been applied to a dataset of icebergs to differentiate backscatters such as surface, dipole and dihedral. Specifically, Freeman Durden, Yamaguchi, Cloude Pottier and Van Zyl decompositions have been applied on icebergs. All decompositions have unique backscatter measuring techniques and purpose.
Freeman Durden decomposition aims to measure the three elemental scattering mechanisms from the polarimetric covariance matrices of a target: surface, dihedral and volume/dipole scattering. One drawback in the model-based decomposition is the over-estimation of volume scattering. This decomposition classifies all the cross-polarization (HV) backscatter component as volume scattering but some of the cross-polarized returns might be the contribution from the rough surface scattering or terrain effects. Pure volume scattering should be reflected from randomly oriented dipole scatterers which have not been considered. The Yamaguchi decomposition proposed an extended volume scattering model, which discriminates volume scattering between dipole and dihedral scattering structures measured from the cross-polarized HV component. In comparison, Freeman Durden decomposition considers all HV contribution as volume scattering. The Yamaguchi method enhances the double-bounce scattering contributions over the urban areas compared with those of the existing four-component decomposition, resulting from the full utilization of polarimetric information. Another eigen-vector based decomposition, Cloude Pottier has been used. Cloude Pottier decomposition is quite familiar for iceberg studies as it is previously been used for iceberg. Van Zyl is also been applied to the iceberg dataset to compare the backscatters with other decompositions. The intention is to look for the similarities and differences of the trend of decomposition outputs.
This abstract is focused on investigating the trend of the results of iceberg backscatters from all four decomposition outputs. The results obtained from all decomposition techniques have given some clear concepts of iceberg properties. This, analysis have been performed on a validated RADARSAT-2 dataset at C-band. Iceberg targets have been extracted from five RADARSAT-2 images acquired in the Newfoundland and Labrador region of Canada. The analysis results will incorporate to the extensive iceberg studies in future.

Detailed observation products of land information are important for describing and understanding the Earth. New technologies and satellite missions deliver better temporal resolution to improve the model predictions and, to obtain new remote sensing products. However, the reflectance signal is often contaminated with clouds and aerosols, which hampers land parameter estimation. To avoid these problems, we implemented an optimized data assimilation approach on Google Earth Engine (GEE) to combine multi-sensor time series optimally. Smoothed and gap-free reflectance time series were generated and they are used to produce new remote sensing products at continental scales. In this study, we use imagery from the Landsat mission (30 m spatial resolution and 16 day revisit cycle) and MODIS mission from Terra and Aqua platforms (500 m spatial resolution and daily revisit cycle). The methodology takes full advantage of the computational power of cloud computing platforms such as the GEE and allows to process large amounts of data (Landsat and MODIS) very efficiently. We implemented a bias-aware Kalman filtering to generate gap-filled Landsat reflectances at monthly temporal resolution. These new reflectance gap free high spatial resolution data are used along with machine learning models to produce new and robust remote sensing products that provide high spatio-temporal resolution land information at continental scales. Quantitative and qualitative evaluation of the land products confirm the validity of the approach. The new land products generated (e.g. LAI, FAPAR, ...) are of high interest for the scientific community, governmental organizations, and the private sector due to the broad spectrum of potential applications possible: from high precision agriculture and crop and/or forest monitoring, to crop yield prediction, and land use change assessment. They endorse an enhanced understanding of biodiversity and nature of the Earth system with a high level of spatial detail.

In Earth Observation (EO), images and data from satellites are traditionally transmitted to ground for analysis. Recently, Artificial Intelligence (AI), as an algorithmic technique, has gained particular interest and application for analysing the satellite data. While AI models and tools are developed and deployed on the ground, AI solutions are also being enabled on the edge for in-orbit applications via AI-oriented hardware devices onboard the satellites. For instance, ESA will include the Intel Movidius Myriad 2 Visual Processing Unit (VPU) in the Tyvak Mark-I CubeSat, which will carry the HyperScout hyperspectral imager, as well as internal and external environmental monitoring of the International Space Station. The CubeSat is also being considered for maritime vessel recognition, based on the on-board integration of ‘Automatic Identification System’ signals from ships. The availability of AI-oriented hardware devices allows for tasks such as deep learning inference and pre-processing of sensor data in-orbit. The repercussions include potentially reduced storage requirements onboard the satellite and on ground, efficient use of the downlink bandwidth and power onboard the satellite, and reduced cycle time for analysing the satellite data. In this context, it is critical to be able to tune and improve the in-orbit AI models during flight in order to fully realise AI’s potential in space applications.

To address this, this project develops a set of techniques and tools to continuously tune the AI models in-orbit to improve the model’s effectiveness. This will facilitate adaptation of the original model, which was initially built and trained on ground using non-mission sensor data, to the specific and real-time sensor(s) data of the deployment hardware on the satellite, and will allow the model to accommodate to changes in sensor performance over time in a neuro-plastic way. The techniques will also enable performing these in-flight updates efficiently over the limited bandwidth uplink paths. This will be achieved by assessing the weight changes between different versions of the AI models which will then be uploaded to the satellite thus enabling lightweight update and retraining of the models.

Additionally, the tools and infrastructure that are used to perform analysis on satellite imagery and data are typically spread across both in-orbit and ground-based systems. Consequently, the satellite image/data analysis workflow is composed of multiple steps that involve processing performed on a satellite, and detailed analysis, model training and response performed on the ground with multiple systems. Therefore, it is necessary to create a platform for end-users to integrate the steps into a single workflow that is executable via a portal that will allow access to both in-orbit and on-ground systems, and also serve rapid development, deployment and verification of applications on these systems. To achieve this, this project develops a workflow engineering platform that can be deployed on a variety of hardware platforms (lab PCs, HPC clusters, cloud services). This platform will allow instantiation, customisation and maintenance of the workflow for specific use-cases. The platform will also allow interoperability with external systems through services and microservices such as offering shapefiles, events to trigger external workflows, and updates of imagery in GIS tools via plugins for further downstream analysis. This facilitates rapid development of applications that target a wide range of use-cases, target satellites and HPC/data centres.

To demonstrate the techniques, tools and services created in this project, we will also develop and deploy the following use-case: Evidence that Level-0, or a higher level product generated rapidly in-orbit, can produce acceptable results using on-board processing will have a direct and significant impact on the architectures of the next generation of satellites. We leverage the solutions that are developed in this project to assess whether on-board data can be used to detect active wildfires in orbit for both algorithm-based and machine learning-based approaches. When fire detection algorithms can be applied in orbit, the latency on active fire alerts can be reduced from hours to minutes, thus providing a valuable societal benefit. The aim is to take Level-0 data and process it rapidly using data available in the L0 packets and data from pre-flight or in-flight sensor characterisation to produce a quasi-Level-2 product (nrt-L2) for input to fire detection routines. Both algorithmic and machine-learning approaches are possible with the in-orbit Intel Movidius Myriad 2 VPU, and both will be explored here. The minimum level of processing for L0 data required will be investigated so that in-orbit processing can be optimised.

This project is executed by Ubotica Technologies and the Irish Centre for High-End Computing with funding from the European Space Agency.

Super-resolution reconstruction (SRR) is a common term for techniques aimed at enhancing spatial resolution beyond the capabilities of an imaging system. For satellite images, SRR can be performed from a single image or multiple images presenting the same area of interest. Recently, deep convolutional neural networks (CNNs) have been exploited for single-image SRR, rendering outstanding results in terms of the quality of the reconstructed images. On the other hand, multiple-image SRR benefits from information fusion—as a result, it achieves higher reconstruction accuracy than single-image approaches, and to the best of our knowledge, deep learning has not been exploited for this purpose up to date. In this talk, after presenting the latest achievements in SRR attributed to deep learning, we will focus on showing two important research directions that we currently explore, namely: (i) preparing good data for training deep CNNs for enhancing satellite images, and (ii) exploiting deep learning to improve multiple-image SRR.
Deep CNNs effectively learn the relation between low resolution (LR) and high resolution (HR) from a training set of corresponding LR and HR patches. However, as deep networks require huge amounts of training data, they have to be generated automatically—original HR images are degraded and downscaled to obtain LR images. During training, a network is presented the artificially-obtained LR data, and it is learned to minimize the difference between the super-resolved image and the corresponding original HR input. This means that the network learns how to revert the degradation process that has been earlier applied to obtain the LR data, however this process may not fully reflect the actual imaging conditions. Unfortunately, the impact of the procedure employed to generate the training data has been little explored so far, and commonly bicubic interpolation is used for downscaling HR patches. During our talk, based on the results of our extensive experiments, we will argue that the procedure for generating the training data is a crucial factor for obtaining high reconstruction quality, and we will present our conclusions on how these training data should be prepared when dealing with satellite images.
Subsequently, we will report our successful attempts to use deep CNNs for multiple-image super-resolution. We will show that deep learning can improve the capabilities of existing techniques which are based on information fusion. To analyze the reconstruction quality, we employ a variety of measures, which allow for understanding the reconstruction process from different perspectives. Furthermore, we will discuss the performance of the proposed methods for a variety of input images, with particular attention given to Sentinel-2 data.
To conclude our talk, we will outline future research directions which may help in deploying single-image and multiple-image SRR algorithms in practice.

The rapid technological developments in earth observation missions have yielded massive amounts of complex heterogeneous scientific data. Recently, the European Space Agency has become one of the frontrunners in enabling the open availability of Earth observation data thus facilitating open-data science. This includes image data from the Sentinel missions made available through the DIAS platforms. While obtaining raw image data seems straightforward, annotating these data continues to be cumbersome and troublesome task. In particular, the amount of such labeled data is low and limited since it involves expensive procurement process.

On the other hand, deep learning approaches have been surging in many domains facilitating the data-to-discovery process. Their success in tackling various challenges involving image data have put them in the research spotlight of the artificial intelligence community. However, while deep learning methods are able capture and cope with complex patterns present in the data, model construction still remains computationally demanding and data intensive. Their development is largely driven by technological advancements yielding increased computational power as well as the availability of massive amounts of (labeled) data.

Considering these challenges on both ends, we propose to exploit the large amounts of unlabeled data for unsupervised learning of Sentinel image representations with deep neural networks. In particular, using the available large amounts of unlabeled satellite images, we will learn a deep model that will be able to extract visual features by means of deep clustering. We will then utilize the (considerably smaller amount of) annotated ground data to further refine and fine-tune the deep network.

We will demonstrate the ability of the proposed deep learning approach on two practically relevant case studies. The first case study addresses the challenge of measuring water availability, where the task is estimating the water budget from limited amount of ground data. The second case study tackles the task of accurate estimation of land cover and land use. We will show that this approach gives a strong competitive advantage compared to using only labelled ground data for learning deep models. Moreover, the learned deep neural networks can subsequently be used not only for similar image classification tasks from different areas but also as a pre-trained input for more complex downstream tasks.

The constant exploitation of natural resources and the increasing anthropogenic pressure on ecosystems continue to put a strain on and irretrievably threaten global forest ecosystems. Global forest health is declining due to climate change, air pollution and increased human activities. Protecting and monitoring the health of forest ecosystems is a vital resource management function. Standard methods of assessing forest condition is through combined use of ground surveys and remote sensing means (aerial and satellite systems). Remote sensing data and methods have been extensively used for forest health studies, since they provide timely and cost-effective information at different spatial and temporal scales, thus offering insight into the dynamics of stress and mortality patterns caused by different factors.
The need for automation to the remote sensing field prompted the development of automatic algorithms and workflows to reduce complexity and simultaneously increasing accuracy. The main objectives of such approaches are: i) the deep analysis of the contextual relations, ii) the adequate treatment of big data and data variability and iii) the presence of a critical statistical evaluation of the results obtained. Nowadays, the existence of effective tools have boosted the development of automated workflows such as Spatial Models (SMs). The SMs are automated toolkits for building, modifying, and running workflows on geospatial data.
This study presents SMs for time-series forest health monitoring that exploit optical multispectral and Synthetic Aperture Radar (SAR) data. Four different types of SMs were designed: i) SMs that ingest available time-series data and perform pre-processing activities, ii) SMs that calculate timeseries of vegetation, soil and water indices from multispectral optical imagery, iii) SMs that create colored composite images from image algebra and SAR polarizations and iv) SMs that extract change detection maps from time-series SAR data. The main contributions are: i) development of automated solutions ready to feed web-GIS platforms in terms of a Web Processing Service (WPS), ii) user-friendly interface (GUI) of the designed SMs, iii) flexible parameterization of the designed SMs, iv) potential high connectivity of the designed SMs through sub-SMs and v) support of a multi-modal approach, i.e., utilization and process of optical data (multispectral imagery) and SAR data.

The coastal erosion is defined as the loss or displacement of land along the coastline due to the action of rain, ice and atmospheric and marine forcing. In Europe, about 25% of the European coastline for which data is available is currently retreating and erosion up to 20-30 m when severe storm occurs are frequent. In the context of climate change and sea level rise, the coastal erosion issue will be even more relevant in the future, as both the frequency and strength of storm events are likely to increase causing billions of Euros of damages.
Dealing with coastal erosion implies to provide highly precise (sub-metric) geo-located information about shorelines on a yearly to decadal timescale. For this reason, the classical approach for coastal erosion assessment has been based on the analysis of historical aerial orthophotos (enabling the extraction of long-term trends), along with GPS field observations (short-term beach and shoreline changes at the timescale of storm events).
These past 20 years, the use of new technologies for coastal change and shoreline monitoring has significantly increased (airborne lidar topographic surveys, photogrammetry, in situ laser scanning) but their definitive adoption still depends on their cost to effectiveness ratio. In the meanwhile, satellite remote sensing has shown to be adequate and relatively cost-effective for the extraction and monitoring of shoreline change over sandy coasts, and other relevant indicators of coastal erosion (beach and cliff morphological patterns, nearshore bathymetry, tidal flat change in semi-enclosed bays,…).
In the framework of the ESA Coastal Erosion Project, the future “Space for Shore” service intends to unravel the remaining technical issues and to provide a large European end user community spread over 5 countries with prototyped products, that are based on the Copernicus Sentinel-1 and Sentinel-2 missions and, to a certain extent, on Third Party Missions. The challenge definitely resides in the adequation between EO-data ground resolution, satellite product accuracy and the rate of coastal erosion and shoreline change. collectively targeting the retrieval of major coastal indicators designed to support decision-making.
This communication proposes an overview of the first results obtained from the Proof-of-Concept activities run this year, accordingly to the set of products required by European coastal managers in charge of shoreline management.
A soundful validation of these prototyped products is also performed assessing satellite real capability to meet end user requirements in terms mainly of measurement accuracy and temporal revisit. Finally, in conclusion, benefits for the coastal community are addressed along with some perspectives for wide service deployment over Europe.

The Biocoast service is a new-generation tool for biodiversity monitoring and decision making. A full mapping process based on the exploitation of High and Very High Resolution satellite image time series (such as Pléiades, Sentinel-2) and latest machine learning approaches such as Deep Learning has been developed in order to map coastal vegetation habitats. Of course, this has been done accordingly to the existing regulatory typologies (EUNIS, Corine Biotope, Natura 2000) and/or more precise typologies based on phytosociology when required by end users. This service has raised thanks to the support and funding of the French government (Programme Investissement d’Avenir) and of several environmental agencies (Agence Française pour la Biodiversité, Conservatoire du Littoral) and marine protected areas managers.

The core of our mapping process is the articulation of two main pillars:
- On the first side, an iterative discussion with end-users and vegetation specialists allows us to define and fine tune the targeted typology according to the potential of available ground truth data and image constraints. Field campaign measurement is also discussed, designed and adapted to site specifications or thematic consideration with respect to image sampling constraints. This leads to the elaboration of a pertinent and accurate training database,
- On the second side, our image analysis and machine learning strategy is adapted to the problem. A pixel-wise classification of time series image features with RandomForest for various vegetation class mapping or convolutional neural network architectures for very high precision mapping of specific objects can be selected alternatively. Depending on the end user final issue, focus may be more on temporal analysis (vegetation phenology) or either on spatial analysis (such as shapes and textures). Our process allows to investigate and optimize the exploitation of the satellite image capabilities.

The Biocoast mapping process has been applied to produce vegetation maps of five different coastal areas in France with very satisfying results, both in terms of spatial accuracy and typology recognition (an average of 87% of overall accuracy, and an average number of twenty vegetation classes). Since all pilot areas are very different in term of geographical extent, end user needs and issues, vegetation complexity, the robustness and the reproducibility of our approach has been well proven. Note also that very accurate mapping based on convolutional neural network was recently applied to intertidal seagrass mapping and oyster bed mapping as well, with impressive results (when compared to manual digitization) what open doors to concrete applications of deep learning algorithms for coastal issues.

The Biocoast service is now operational with obvious applications being currently performed for Natura 2000 site assessment, for the monitoring of invasive exotic species and of land use in natural protected areas. More is to come and imagine along with the increasing implication of end-users fertilized by the Biocoast results and by big data and machine learning technologies.

European Commission (DG-Agri, DG-Grow, JRC) together with European Space Agency and national stakeholders are currently working to understand the full range of opportunities the Sentinels satellites can contribute to modernize and simplify the Common Agricultural Policy (CAP). Within the World Bank project “Supporting Transparent Land Governance in Ukraine” similar work will be done for Ukraine, taking into account geographical and national specifics and data available. This work is dedicated to verification of the usage of remote sensing data to identify land use, in particular, monitoring of actual land use of agricultural land, as well as cultivation of other land categories in Ukraine for three regions (Kyiv, Lviv and Mykolaiv regions).
For Ukraine, as an agrarian country there is a great need to use remote sensing data for land use monitoring. Every year there are very few violations recorded, although they are actually this number much larger. Such violations include the incorrect crop rotation and unlawful use of land, it very negatively affects the land productivity. Classic instruments have shown inefficiency, which is why land cover and crop types classification maps built on satellite data is considered as the main tool for the transparent land transparency. For this purpose, we propose to use modern land use/land cover (LU/LC) classification techniques for land use assessment based on freely available satellite data (Sentinel-1 and Sentinel-2 data). Our technical approach will be based on the four pillars: (i) best practices of leading international programs and initiatives we are involved in (GEOGLAM, JECAM, Horizon-2020, ESA), (ii) freely available satellite data, (iii) modern achievements of computer science (machine learning techniques) and (iv) specialized open source software (Sen2Agri system, Google Earth Engine platform and Deep learning classification platform).
During the first stage of the project it was demonstrated that proposed methodology is consistent and could be extended to the entire territory of Ukraine or any other country.
Further project improvement could be ensured by automation of in-situ data processing. We plan to use action cameras installed on the vehicles during ground surveys (similar to those used in Google Street images), a stream of images could be collected in such way, classified and incorporated in the generation of ground truth labels. The advantage of this approach is much more reference data could be collected (by our estimates at the order 1-2) and therefore it will benefit generation and validation of geoinformation products.

At the time when a lot of high resolution satellite data are freely available (every 16 days Landsat-8 with 30 m resolution, every 5 days Sentinel-2 with 10 m resolution, every 6-12 days Sentinel-1 with 10 m pixel spacing) and machine learning, in particular deep learning, methods became powerful instrument for solving different problems, providing accurate land cover and land use maps based on remote sensing data for the large territory is important and challenging task [1]. Precise and accurate land cover maps as well as crop specific maps could be used as an input for different other applied applications, like crop yield forecast, crop rotation violation, impact of droughts or for insurance companies and automated validation of land-parcel identification system (LPIS) [2]. Many studies have utilized pixel-based approach for remote sensing data classification, but obtained crop classification maps are contaminated with noise, because such approach considered only spectral and not spatial information. In this paper we proposed filtration methods for land cover and land use classification maps, that obtained based on high resolution satellite images. The main idea is to adapt and utilize state-of-the-art deep learning U-net architecture, that previously has been used for image segmentation, for classification map filtration task and to compare them with traditional rule-based approach. The most complicated tasks in the filtering such maps are to preserve edges, boundaries between different fields and automatically take into account not only single object or its neighbors, but also consider whole surrounded situation. The proposed U-net architecture allows us to overcome these issues and to provide more accurate and precise filtration in terms of accuracies per class. In addition, it shows advantages comparing to rule-based method in successfully dealing with big objects of noise in large fields and at the same time prevent significant changes in small farmer fields that grouped near villages or towns.

1. N. Kussul, et al. "Deep Learning Classification of Land Cover and Crop Types Using Remote Sensing Data," IEEE Geoscience and Remote Sensing Letters, vol. 14, no. 5, pp. 778-782, 2017.
2. F. J. Gallego, N. Kussul, S. Skakun, O. Kravchenko, A. Shelestov, and O. Kussul, “Efficiency assessment of using satellite data for crop area estimation in Ukraine,” International Journal of Applied Earth Observation and Geoinformation, vol. 29, pp. 22-30, 2014.

COMET's LiCSAR is an automatic system performing InSAR processing of Sentinel-1 SAR data. Current LiCSAR platform is developed in three main priority fields: a) achieving knowledge of tectonic displacements over the world, b) a fast response generation of interferograms over areas affected by an earthquake, c) monitoring of activity over world volcanoes. The first priority field is based on exploitation of the huge archives of Sentinel-1 data using current state-of-the-art interferometric techniques reaching maps of average velocity of tectonic motion. The second priority area related to fast response interferograms after an earthquake helps assessing the earthquake parameters by experts. The third area attempts to detect signs of displacements over active volcanoes from generated interferograms, with the help of AI.

On volcanoes, LiCSAR system produces more than 30 000 interferograms on 900 different volcanoes. Machine learning techniques are applied to automatically detect ground deformation signals. Preliminary results could detect 2017 unrest signals at Cerro Azul and Etna. Our final aims are to deliver an early detection system capable to track volcanic deformation in near real-time (with a user interface that enables scientists in volcano observatories to use the data for monitoring purposes). Such system can be further upgraded to monitor other geohazards, for example development of sinkholes.

This contribution will demonstrate our current progress in the data processing approach, results to date, and some of the challenges and technical solutions associated e.g. with reduction of atmospheric delay errors in the interferograms (applying corrections based on GACOS system), deriving large-scale crustal velocity and strain-rate fields at the plate tectonic scale, and the automatic detection of volcanic displacements.

The current goals of the system include deriving InSAR-based displacements products over the whole Alpine-Himalayan Belt and incorporating novel technologies and algorithms to ensure sustainable development of the system for using data of future SAR missions.

Although many actions and environmental regulation are currently deployed, pollution at sea is still increasing and is now major cause of concerns for environmental management and preservation of the marine biodiversity. These sources of pollution are linked to anthropogenic pressure, directly, in the case of marine litters and micro/macro plastics, or indirectly, in the case of micro/macro algae proliferation (e.g. the sargassum now migrating from Western Africa to the Caribbean). The ‘traditional’ detection of such pollution from space (EO) is usually done by using the visible bands of the spectrum (with extension to NIR). Classical examples are i) the Harmful Algal Blooms detection and ii) marine litters (recent advances in Arias et al. 2019 in prep.). All these methods are using radiometric spectral signatures for detection and possible identification and are, the most generally pixel-based. Our research team is investigating the capability offered by the Machine Learning to detect specific geometrical features in addition to this radiometric identification. This gives more reliability to the ‘classical’ detection. Our first investigation has been applied to sargassum detection in the Caribbean area (knowing that the results could be easily exported to marine litter due to the similar shapes – generally non-continuous filaments - of floating accumulation). The proposed technique is done in three steps. First, the EO view (e.g. sentinel 2) is divided into smaller square subsets. Second, a classification is done on these subset by two ML techniques (Random Forest and MPL) to determine the probability of occurrence of five classes (cloud, land, sea, others, no signal). The two techniques are then evaluated (efficiency and computing time). Third, for each subset presenting a significant probability of having interesting features, a new ML technique is applied by screening the subset with a sliding window to detect (after training) possible patches of sargassum. An alternative to this method is the Mask R-CNN applied globally to the subset. Another innovation to this detection of ‘pollution’ at sea is brought by the use of the three Sentinel (1,2 and 3 (OLCI)) together (with similar ML techniques) to reinforce i) the temporal and spatial coverage and ii) the reliability index that one can attach to the detection. These techniques and their efficiency will be presented first with one sensor and, second by using the whole capabilities offered by the Sentinels.

TSAR AI helps emergency and disaster response teams to work quicker, smarter, and more effectively through on-the-fly analysis of geospatial Earth Observation Data provided by satellites. In situations where minutes and hours can make the difference, TSAR AI will improve the existing processes that rely on highly trained human expert analysts.
Our unique approach in the field of AI-driven data analysis is based on know-how obtained through the Technology transfer programme from ESA and developed further through collaboration with the University of Amsterdam. In contrast to alternative approaches, our methods does not require high quality labeled AI training datasets and can rapidly detect known and unknown features through unsupervised ML and models based on inputs from human crowd of analysts.
Our team is also one of active (remaining) teams of Round 3 AI for Good XPRIZE competition (sponsored by IBM Watson) with project entering an active phase, as we enter into the final round.

Monitoring tree cover dynamics in the tropical region can serve multiple purposes such as carbon emission mitigation or conservation. In the recent years, early warning systems have become increasingly needed in complementarity to monitoring and reporting frameworks like the Reducing Emissions from Deforestation and Degradation and enhancement of carbon stocks (REDD+) proposed by the international policy framework on climate change. Such early warning change detection applications can support notably rapid Law enforcement measures on the ground.
The European Copernicus Programme has been providing Earth Observation (EO) data for a few years now. Both optical and radar data can be accessed for free; the latter being of particular interest in the context of tropical forests that are regularly overcast. Concomitantly, large-scale tree cover change products have been released for free (e.g., Global Forest Watch products, or national forest inventory products). Although the accuracy of such products is not always known, they constitute a significant source of information that is worth being considered for calibration and validation purposes.
Several machine learning techniques capable of handling EO data are proposed in the literature. For over a decade now, neural networks, a.k.a. Deep Learning techniques, have opened avenues in the remote sensing arena thanks to new computational capabilities (e.g., advent of Graphics Processing Units) and the availability of large EO datasets. In this study, we explore the potential of specific deep learning algorithms that can handle explicitly the time dimension that is a core feature of the EO datasets considered in this study. The potential of different Recurrent Neural Networks (RNN) configurations will be investigated. As pre-processing step, optical/ radar data fusion techniques will be implemented to densify the EO data time series. The results coming out from the RNN architectures will be benchmarked with established machine learning algorithms like random forests or classification trees.

The Forestry Commission (FC) spends 20-40% of its budget monitoring, administering, and pursuing the establishment of young tree stock by aerial photography and on-site inspections. Capgemini currently works on an improved monitoring system combining the used of Space data processing and machine learning expertise. Indeed, the innovative solution presented here proposes to engineer features from SAR data as inputs and machine learning models from Data Science field. The inputs used are Sentinel 1 data products, which provide a high frequently measurements, and in-situ data form The National Forest Estate (NFE) and the National Forest Inventory (NFI) providing shapefiles. The applied methodology is a SAR-machine Learning framework relying on pre-processing with Sentinel-1 toolbox, reprojection and multi temporal filtering approach, exploratory analysis, feature engineering and statistical machine learning to obtain actionable insights accessible through a web application offering visual analytics.
The solution considered here is to use Sentinel 1 Synthetic Aperture Radar (SAR) data to identify young trees. As opposed to aerial photography, the Sentinel 1 constellation takes SAR measurements frequently: approximately every five days. Analysing this data will enable a much more timely response to violations of agreements and a close-knit monitoring of assumed woodland. Furthermore, SAR backscatter information can be used to distinguish clear-fell or unused land from regions with newly planted trees by tracking the temporal change in height of the land.
The solution proposed here is to engineer features from SAR data along with auxiliary as inputs for machine learning models. For the purpose of this research, three years of SAR data (2015-2017) have been downloaded. Then, the related data set has been pre-processed with Sentinel 1 toolbox from SNAP, a free software provided by European Spatial Agency. In order to filter the speckle, the preferred chosen approach is to remove it with a Multi temporal filtering.
To obtain a categorisation of the various woodlands in regions of interest, two publicly available shapefiles from the FC are used:
- National Forest Estate (NFE): The NFE contains the FC sub-compartment database and gives information for recording, monitoring, analysis and reporting across the entire FC estate.
- National Forest Inventory (NFI): The NFI includes interpreted forest types for all woodlands over 0.5 ha and interpreted open area information for areas over 0.5 ha surrounded by woodlands.
Every SAR image is split to cover individual polygons, and for every such combination of SAR image and polygon, a record in a data frame storing all combined data is created. Due to the complexity of the problem, it has been selected a simple region of interest for the first trial: Kielder (UK).
The aim of our experiment is to distinguish young trees (precisely conifer) from grassland on a polygon. To create a dataset, we used NFI data in order to keep only polygons containing these types of vegetation. A Python script has been developed retrieving a list with features and applies them to the dataset. The related output is a ‘feature data frame’ which contains all the computed features and a unique identifier for every SAR image. The features created together with the ground truth are used to train the model on the training data and this latter is applied to the test data.
The temporal step between two SAR data is currently tested and results of machine learning algorithms such as SVM, Gradient Boosting and K Nearest Neighbour are compared. According to our data, the machine learning algorithm Gradient Boosting has the best performance of all models tested. The model achieves to a classification score of 87% for Kielder region. These scores are very promising and may already be sufficiently high for many applications.
The results then demonstrate that the SAR-Machine Learning framework is able to produce a machine learning model that can identify trees with a high accuracy. It must be applied on other regions to test the global capabilities of the machine learning model. But, it maybe not feasible to build models for mountainous terrains, potentially due to the distortion of DEM. There is also an opportunity for improving results by adding meteorological data coming from UK MetOffice as input for the SAR-Machine Learning framework.
In the long-term vision, the tree identification model may be implemented as a tool to help business analysts with automatic recommended decisions derived from model predictions, as well as give business intelligence and insight for data mining studies.

REFERENCES
[1] J. Bruniquel, A. Lopes, “Multi-variate optimal speckle reduction in SAR imagery”, Int. J. Remote Sens., vol. 18, no 3, pp. 603-627, February 1997
[2] S. Quegan, J. J. Yu, “Filtering of multichannel SAR images”, IEEE Trans. Geosci. Remote Sens., vol. 39, no 11, pp. 2373 2379, November 2001
[3] S. Tufféry, “Data Mining and Statistics for Decision Making”, John Wiley & Sons, 2011

The Earth is a complex dynamic and networked system. For the past few decades, a large number of extreme weather and climate events in Europe and worldwide have occurred which resulted in infrastructure damages and many casualties. Agricultural drought and wildfires are examples of some of the most critical hazards in terms of frequency, severity and impact on livelihoods. Detecting such extremes and anomalies is of paramount relevance to mitigate impacts and incorporating prevention measures. However, the phenomena are difficult to predict as they are complex and depend on many factors.
A vast suite of approaches have been developed to monitor and characterize e.g. agricultural drought, based on either climatic ground-based data, soil moisture data or a variety of remote-sensing drought proxies. A recently proposed Soil Moisture Agricultural Drought Index (SMADI) is a simple and intuitive index that determines agricultural drought events based on key remote sensing indicators: land surface temperature (LST), vegetation indices (e.g., the NDVI) and surface soil moisture (SSM). While indices like SMADI and other alternative hand-crafted indices have been widely adopted and used in real practice yielding good results, they often ignore the complex nonlinear and multidimensional variable relations in the problem. In recent years, statistical machine learning (ML) has played a role in Earth observation data problems with positive results in problems like classification and anomaly detection. Machine learning can actually cope with multivariate and multiple source data sources and allows one to automatically detect anomalies. There is a plethora of ML algorithms for anomaly detection (AD), ranging from simple histogram-based models to more advanced hierarchical density-based clustering algorithms. Each has its advantages and disadvantages, but often need the help of user expertise and process understanding to improve results.
In this work we introduce the application of a hybrid approach based on state-of-the-art ML anomaly detection methods and standard drought indices like SMADI for drought detection. We will illustrate the performance in the Earth System Data Lab (ESDL), a light-weight platform for Earth observation data analysis in the cloud. The ESDL allows to evaluate algorithms in a wide range of harmonized products including more than 40 variables spanning more than 10 years. The included variables account for atmospheric conditions, climate states, and terrestrial biosphere. Extracting anomalous events is possible with the multivariate spatial-temporal information contained within the Earth datacubes using advanced hybrid ML modeling. Algorithms will be compared in terms of accuracy, robustness and computational efficiency in selected examples of droughts in Europe during the last decade.