This paper provides a review of the functional characteristics of raster-based GIS data model and urban Information Systems that can be used for spatial decision support systems to aid building Insurance Management Decision.
A variety of advanced computer based application is available to insurance companies and regulatory bodies in the management of building insurance policies. These applications often concentrate on land information and values of the property in the pre-assessment evaluation. However, considering other post construction variables, the availability of urban infrastructure, socio vices, disasters risk, and third-party damages all pose a cumulative impact on social costs and the value of insurance premium that should be placed on the property.
The computing tools and technologies available to the insurance companies with the integration of raster-based GIS model and urban information system can mitigate the initial often challenges encountered during the decision making for pre-assessment for insurance and damages associated with climate disaster.
This paper introduces a prototype model which provides decision support systems for the insurance documentation process, on the web environment, by integrating a property insurance pre-assessment application with a raster-based GIS data and Urban Information systems. This model will also incorporate raster based climatic disaster prediction model.
KEYWORDS: Geographic Information System, raster data, disaster risk, insurance and
Urban infrastructure system.

Modelling Earth-Observation Assessments and Adaptation of North-Atlantic Coast to Extreme Climate and Ocean Events
1Akande S.O.
Department of Meteorology and Climate Sciences
Centre for Space Research and Applications (CESRA), Federal University of Technology, Akure, Nigeria
soakande@futa.edu.ng
2Osinowo A.A
Department of Marine Science and Technology, School of Earth and Mineral Sciences
Federal University of Technology, Akure, Nigeria
3Jejelola O.F
School of Environmental Technology, Federal University of Technology, Akure, Ondo State, Nigeria
4Olajire O.O
African Regional Cebtre for Space Science and Technology Education-English
Obafemi Awolowo University, Ile-Ife, Osun State, Nigeria
5Apenuwa O.S
Department of Urban and Regional Planning, School of Environmental Technology
Federal University of Technology, Akure, Nigeria


ABSTRACT
The advancement of earth observation technology has greatly improved global research in climate change over the past few decades. Radar and optical remote sensing applications became numerous and the system concepts have been implemented in modelling coastal vulnerability assessments to climate change and extreme events. Coastal ecosystems are considered to be sensitive to different environmental forcings, particularly sea-level rise and an increase in storm intensity. Human produced greenhouse gases (GHGs) concentrations in the atmosphere are currently rising at alarming rates. These gases are the major driver for changing climate scenarios highlighted by an approximate 1°C increase in sea surface temperatures. In addition to driving global warming, GHGs are readily absorbed by the oceans, resulting in changes in seawater chemistry and a decrease in seawater pH (acidification). In addition, the phases of the El Niño Southern Oscillation affect the climatic (e.g rainfall) trends and pattern in the study area. The study was conducted in the North Atlantic coast which stretches from within four (4) continents (North and South America, Europe and Africa). It is home to a large number people, economic and maritime activities with over two hundred (200) million populations. In order to ensure the preservation of preserving natural resources and ecological features, a modified methodological approached was developed to evaluate risk of coastal inundation from sea level rise and storm surge at specific sites, areas of importance for natural, and infrastructural resources. Monthly and seasonal sea surface temperature data were obtained the National Oceanic and Atmospheric Administration (NOAA) and Advanced very-high-resolution radiometer (AVHRR) over the North-Atlantic region between 1959 to 2018 (60 years) were analysed in relation to three El Nino-Southern Oscillation (ENSO) phases. In addition, the derived earth observation data from Sentinel-1, Sentinel-2, and ASTER satellite images, high spatial resolution multispectral imagery and spatial modelling, were processed to generate land use, vegetation, geomorphology, and coastal erosion maps. The tools were developed using the ArcGIS 10.5 and Idrisi Terrset software to enable users conduct risk analyses of coastlines by classifying vulnerable areas, and also to produce outputs for further analysis. The recommendations from this research thus enhance the capacity and knowledge of governments and other policy makers in the region to prepare for, and adapt to climate change and its extreme events. A coastal spatial planning designed and then the Integrated Coastal Zone Management framework for West Africa coastal region was developed. Lastly, results from this research would help to improve weather prediction thus developing the economy as well as saving lives and properties during climate-related hazards such as drought, forest fires and floods. The study therefore reviews the advances of climate change studies based on Earth observation data and provides examples of Earth observation applications in extreme ocean events and climate change research. The study then concluded that the earth observation data would provide scientific and academic users the ability to make choices of use, depending on the (spatial and temporal) resolution types, with a huge volume of various satellite data types for global research and decision making, and policy implementation.The results from this research would enhance the capacity and knowledge of global policy makers to prepare for and provide a sustainable adaptation strategies in line with the United Nation’s Sustainable Development Goals (SDGs 1,2,13 & 14).



Map of the Study Area (Author’s Research, 2018)

Keywords: Risk, Coastal, Inundation, Ecosystem, Projections

The Canary Islands, Spain are considered one of the best hotspots in the world to see wild marine mammals. Of the approximately 90 species in the world, 31 can be found in the Archipelago. EO data has been extensively used over the years for assisting in the management of marine mammal populations either by establishing protected areas where stakeholders’ activity will be reduced, or by minimizing the impact of anthropogenic threats. It is considered a basic and essential tool for the conservation of the species both by researchers and the government. Satellite measurements of ocean colour are the principal remote-sensing tool for measuring ocean productivity and its response to climate change/variability. Consequently, sea surface chlorophyll-a concentrations (measured as ocean colour) are often used as proxy for primary productivity. Remotely sensed environmental parameters have the potential to identify biological hotspots for cetaceans and to therefore establish or better manage areas of marine conservation priority. The ESA project, “EO_MAMMALS”, led by PLOCAN and with the collaboration of GMV and Univ. St. Andrews, will use the Sentinel S3A and S3B data to model cetacean presence. The region of interest is an area heavily targeted by the whale watching industry and centered around a zone of special conservation. The results of analysis will be incorporated in a governmental application used by different types of stakeholders such a general public, research, industry and government, as a tool for the management of marine protected areas. Finally, for future applications, this study could be extrapolated to other Macaronesian areas as well as worldwide.

With the improvement of the sensors hosted by newest EO Satellites Payloads, the processing of data from such sensors in near real-time is becoming a hard task. In order to make the data more valuable to the end user (either a person or another system), it must be provided as quickly as possible. The sooner the data is provided to the end user the more valuable the data is. On the other hand, with the large number of satellites flying nowadays, the amount of data being stored in ground segments archives is becoming unmanageable. As such, the current approach of processing and archiving the data systematically, in order to have the data ready for the user when required, is being replaced by processing on-demand, meaning that the data is processed on-the-fly immediately following a user request. In order to achieve this, the ground segment data processors must be as fast as possible to reduce the elapsed time between the user request and delivery of the data to the user.
GMV, as part of its EO data processing software suite, has developed a library (GMV Processing Toolbox) which provides the processor designers and developers with the tools they need to implement an EO data processor that makes the best use of the resources available in the platform where the processor will run. GMV Processing Toolbox allows many kind of different parallelization approaches depending on the characteristics of the algorithms and the data to be processed. This parallelization has been historically focused on using as many cores in the CPU as possible, so the processing time of a given algorithm can be divided by the number of CPU cores available in the machine.
With the proliferation of general-purpose computing on graphics processing units (GPGPU) GMV decided to upgrade its Processing Toolbox, adding the capability to utilise such HW capabilities. The challenge was to allow the designer to keep the same processor design whilst at the same time enabling a set of algorithms (or parts of algorithms) to be processed in a GPU (if available) instead of the CPU. Additionally, the library was required to be intelligent enough to detect the HW resources in the machine where the processor is running, so a piece of candidate code for execution in a GPU will do so only if one is available, running on the CPU otherwise.
The GPUs are very specialized to perform a set of computations over large amounts of data under very specific circumstances. If the wrong algorithm is selected as candidate to be processed in a GPU, it can lead to worse performance than running on a CPU. This is the reason why all original parallelization approaches provided by GMV Processing Toolbox (based on CPU) and the new GPU processing capability must coexist in the design and implementation of an optimal EO data processor.
GMV has selected OpenCL (Open Computing Language) as the framework for writing algorithms that will run on the GPUs. The main reason why OpenCL has been chosen instead of CUDA is, apart from being an open-source GPGPU framework, because OpenCL is cross-platform solution that is not attached to a specific GPU architecture (CUDA is a proprietary technology from NVIDIA). Code implemented with OpenCL can be run in any of the GPUs built by the major makers (NVIDIA, AMD and Intel) as well as in any CPU, DSP (digital signal processors) and FPGAs (field-programmable gate arrays).
In summary, GMV provides as part of its Processing Toolbox the following capabilities:
 -Earth Explorer Format file handlers;
 -Compatibility with ESRIN PDGS Generic Interfaces;
 -Generic logging functionality;
 -NetCDF4/HDF5 file handlers;
 -Time operations;
 -Wide set of interpolation and resampling functions;
 -XML file handlers;
 -Processing infrastructure to build the skeleton of a processor from scratch (or automatically from an Enterprise Architect model) providing different kinds of parallelization, including the use of GPUs.

Browse processor generates data visualizations called “browses” from Earth Observation satellites data. It provides the user the possibility to visually assess selected data in a very systematic way.

Browse provides the user with the power to configure both the visual settings and data content. It allows to downsample, slice and filter data basing on parameters, dataset itself or another dataset. Finally user may extract statistical information from selected dataset.

At the current stage the supported products are Level 0, Level 1 and Level 2 products of EarthCARE and Level 1 and Level 2 ADM-Aeolus missions. A plugin will be developed, that will enable processing of other missions data, and any other structured data.

Browse processor is designed to works both:
- as a stand-alone application,
- as an autonomous part of the Payload Data Ground Segment.

This processor generates in an automatized way different types of plots that are of user interest. The following types of plots are defined:
• Basic –plots any two parameters against each other. It is a generic X-Y plot. Can be used, for example, for representing the evolution of two parameters along the satellite track. Additionally, the latitude/longitude of the ground track can be represented on the x axis.
• Trajectory – plots ground track of the satellite
• Merged – allowing to over-plot many variables on the same canvas sharing axes between plots
• Adjacent – it is a plot showing multiple parameters, splitting track into segments and marking the data accordingly
• Vertical Curtain - represents the variation of a 2D parameter corresponding to a vertical profile, with a single value for each position along track and each altitude value.
• Vertical Curtain 3D - This is a specific plot type mapping vertical curtain data on a 3D map of an Earth browser like Google Earth. In this case a KMZ or ngEO output file is created, which contains the images and the metadata needed to present the data in a 3D world map
• Surface strips in native projection - This plot represents the variation of a 2D parameter corresponding to a horizontal surface, sampled over a grid corresponding to the instrument native projection covering the whole duration of the product, from its start time up to its stop time
• Geolocated Surface strips – This plot represents the variation of a 2D parameter corresponding to a horizontal surface, sampled over a grid corresponding to the instrument native projection covering the whole duration of the product, from its start time up to its stop time and with information about its geolocation
• Rectified surfaces strips – This browse is similar to the Geolocated Surface strips but the 2 dimensional grid where the parameter is sampled is rectified to a coordinate reference system (e.g. UTM). This means that instead of providing the geo-location information for each single point of the grid it is sufficient to provide latitude and longitude of two corners of the grid.

The Browse Processor is able to generate plots in two different ways:
- Job Order mode: it will receive a Job Order file as input, which contains the location and filename of the input product and the location and filename of the Auxiliary Data File.
- Standalone mode: No Job Order is required in this mode. The processor will receive the location of the input product and the location of the Auxiliary Data File via command line.
The BRPF processor has a modular architecture. The processing chain can be subdivided in three different stages:
• Pre-processing. In this first processing step, the BRPF, depending on the inputs it has received, detects the execution mode: Job Order mode or Standalone mode. Depending on the mode, the BRPF will call the proper modules for reading the input data.
• Processing. In this step, there are two sub-steps:
o Data Extraction. This module gets the requested data from the product. The Data Extraction Syntax provides paths to the specific data subsets. This syntax specifies which data subset must be extracted from the input product, allowing also a basic data manipulation or filtering.
o Browse Generation. This sub-step contains several modules for performing the different tasks involved in the generation of the browses and their metadata.
 Rescaling & Down sampling. The following functions are implemented:
• Application of a down sampling factor on each axis of the extracted dataset
• Rescaling of the image to a certain size in pixels for each dimension (with the option of maintaining the aspect ratio)
 Color Mapping. This module will map the values to pixel color.
 Plotting & Visualization. This module will produce the corresponding plot for the type of input product.
 Metadata Generation. This module will produce, if needed, the metadata associated with the output format required.
• Post-Processing. This step performs the compression and image formatting.

Deep learning is undeniably one of the fastest growing trends in big data analysis, as it has proved to be very effective and sometimes even able to surpass humans.
The big quantity of remote sensing data of different type that has become available in the last few years are considerably boosting the exploitation of such techniques in this field.
Besides being less detailed and more disturbed, remote sensing data present some new challenges for deep learning. While "ordinary" images are either grayscale or RGB, remote sensing often deals with multimodal and multisource data, each one having different characteristics (wavelength, acquisition geometry, etc.) and probing different properties of the observed scene. In addition, many training data are needed in order to be effective, for example for agricultural applications, where areas of interest are wide.
In addition, the use of time series of acquisitions is very important to obtain the required discriminatory power. Recent missions, like the ESA Sentinel constellation of satellites, provide, free of charge, both SAR and multispectral optical data systematically acquired over most of the world with a short (few days) revisit time. The temporal information brings its greatest contribution in modelling the “expected” time evolution of the signal.
Although it is well known that multispectral data are rather sensitive to vegetation phenology and crop characteristics, time series gaps due to clouds or dense haze represent a major challenge in the effective use of such data, especially at higher latitudes. Synthetic aperture radar (SAR) images can positively contribute, because they can easily detect some ground characteristics like humidity and urban areas and because they are almost unaffected by clouds. However, it is hard to exploit optical and SAR data jointly. Some authors attempted to exploit SAR data to derive some vegetation index, but it is most common to merge them in the final output.
Thus, we propose a new deep learning architecture aimed at classifying irregular (i.e. with possible data gaps) time series of images from multiple sources (SAR, optical). The proposed architecture, which we will call ODIN (Optimal Data Imputation Network), exploits a sequence of deep convolutional neural networks (CNN) and a new RNN that is able to deal with missing data. In the proposed architecture, all the data (with their spectral, spatial and temporal information) are used jointly and optimally in the sense that no imputation, regression or interpolation schema are enforced, but the internal weights that give the best classification results are estimated from the data themselves (hence the name ODIN).
In order to test the proposed architecture, we have considered the application to land cover and crop classification, which is one of the uses of remote sensing data that has attracted more interest in the last few years. Among the other things, reliable crop classification is a cornerstone in the upcoming new Common Agricultural Policy (CAP) regulations based on satellite monitoring.
Table 1 and Figure 1 show the very promising results we have obtained, with an error rate below 1%, and a great spatial consistency without loss of map resolution, a characteristic very desirable in this application.
+-----------+------------+---------+----------+
| | Accuracy | AUC | F1 Score |
+-----------+------------+---------+----------+
| Grassland | 99.13 % | 99.45 % | 94.28 % |
| Cropland | 99.04 % | 99.18 % | 94.23 % |
| Forest | 99.88 % | 99.98 % | 99.86 % |
| Water | 99.92 % | 100.0 % | 99.85 % |
| Urban | 99.27 % | 99.76 % | 96.19 % |
| Vineyard | 99.56 % | 99.98 % | 96.71 % |
+-----------+------------+---------+----------+
Table 1 - Validation results on Avignon

The importance of wake detection on SAR images is twofold: to estimate vessel velocity by exploiting the difference in Doppler and to detect ships faintly scattering (very small boats, or small and very fast boats). When visible, wakes are quite evident and their detection by means of photointerpretation is quite an easy task even when they are not straight lines or have irregular shapes.

Historically, this problem has been addressed by looking for peaks in the Radon transform of the SAR image. Such approach suffers from three issues: it assumes straight wakes, uniform appearance and accurate selection of SAR image tiling parameters. Moreover, it is best suited as a post processing step for estimating vessel velocity after ship detection, for it is very expensive to run it on whole images.

Highly non straight routes and irregular sea features like low wind areas and natural slicks are just few examples of how the task of wake detection, which is well accomplished by photointerpretation, can be actually a challenging task for fully automatic algorithms.

In addition to the difficulties of extracting wakes from a complex sea scene, speckle and existence of different types and aspects of wakes in different weather and sea conditions and in different SAR modes and resolutions make the automatic wake detection task even harder.

This is a typical problem where machine-learning techniques can be very effective. In this work, we will show how deep Convolutional Neural Networks (CNN) can be successfully exploited to detect wakes on SAR images. To our knowledge, this is the first time such networks are used to address this task. The advantage is clear: machine learning techniques can naturally deal with non straight wakes and irregular patterns just because the network is trained with real life examples and it does not depend upon the hard assumptions needed by the radon based techniques. Moreover, although the data collection and training phases may last long, the operational phase is very fast and hence well suited for processing big images in near real time applications.

ARCHITECTURE DESCRIPTION

The main idea at the base of deep learning is that, stacking several layers in a neural network creates a model capable to fit strongly non linear patterns.
When applying deep learning to image processing, the use of convolutional layers is, nowadays, ubiquitous. The so called Convolutional Neural Networks (CNN) allow exploiting the context information contained in the images in order to build automatically an optimal set of feature maps through a learning mechanism. Such feature extraction is automatic and optimal because it is obtained as the results of the minimization of the recognition errors over a set of training data.

The neural network developed in this project consists in:
- 6 convolutional layers alternated with max pooling layers, ranging from 4 up to 64 features per layer
- 2 1x1 convolutional layers (acting as dense layers)
- 2 dropout layers for regularization
- ~64000 parameters

For reducing the generalization error the full available dataset was split in three:
- Training dataset
- Validation dataset
- Test dataset
In order to estimate the best set of parameters, a minimization procedure was applied. As very common in deep learning, a variant of the stochastic gradient descent called ADAM was applied. A small subset of data (batch) is iteratively presented to the network, its output computed and derivatives w.r.t. parameters taken. A correction to the parameters proportional to the computed gradient through a “small” factor (learning rate) is applied at each iteration until each sample of the training dataset has been used exactly once (epoch). During this, ADAM tries to adapt the learning rate by exploiting the first and second moments of the gradient between iterations. At the end of the epoch, the network performance is tested against the validation dataset and a new epoch begins. The procedure ends when either:
- A prefixed number of epochs has elapsed.
- A desired validation accuracy has been reached.
- Overfitting occurs (i.e. the training accuracy increases while the validation accuracy reduces).
The objective function of such minimization (also called “loss function”) was the binary cross-entropy, which has the advantage of making the network output a probability. Wake and non wake images are given different weights in order to take into account the relative imbalance in the dataset (wakes are very rare events).
To evaluate the generalization error of the final network, the test dataset is used.
Although the network is trained to give a binary answer (there is / there is not a wake in this clip), it internally builds a map saying where the wake actually is. Such map is exploited, in prediction, to localize the wake.

EXPERIMENTAL DATASET AND RESULTS

To our knowledge, this is the first documented attempt to perform wake detection on SAR images through deep neural networks. Our main achievements were:
1. The first training dataset for SAR wake detection.
2. The first working network with accuracy above 80% in detecting wakes.
3. On the field test of the network in three different acquisition campaigns.
As experimental dataset, we selected 532 COSMO-SkyMed WideRegion DGM images from which we extracted 64000 land free clips 800x800 pixels (12km x 12 km). We than had expert SAR photo interpreters manually classify each clip as “wake”/”no wake”. Because only 4.6% of clips were classified as “wake” (sign that wakes in SAR images are rare events), we applied weights to the samples in order avoid a bias both in the loss function and in the validation metrics.
Table 1 and Figure 1 show the classification results on test dataset after just 100 epochs of training.



+------------+-----------------+-------------------+
| | Threshold = 0.5 | Threshold = 0.456 |
+------------+-----------------+-------------------+
| TPR | 78.8 % | 81.4 % |
| FPR | 14.1 % | 16.0 % |
| Youden’s J | 64.6 % | 65.5 % |
| Accuracy | 82.4 % | 82.7 % |
| AUROC | 0.89 |
+------------+-----------------+-------------------+

Urbanization is among the most relevant global trends related to the human presence on the planet. As such, it poses a major challenge for the current urban population and the well-being of the next generation. To fully understand and properly mitigate the impact of this change, we need precise and up-to-date information of the urban areas. The Urban Thematic Exploitation Platform (U-TEP) focuses on delivering multi-source information on trans-sectoral urban challenges. In 2019, U-TEP functionalities and user-interface of were streamlined, new products and services added. With these developments, the platform entered the operational phase.

The U-TEP is developed to provide end-to-end and ready-to-use solutions for a broad spectrum of users (experts and non-experts) to extract unique information/ indicators required for urban management and sustainability. Key components of the system are an open, web-based portal connected to distributed high-level computing infrastructures and providing key functionalities for i) high-performance data access and processing, ii) modular and generic state-of-the art pre-processing, analysis, and visualization, iii) customized development and sharing of algorithms, products and services, and iv) networking and communication.

Besides the product portfolio, that provides access to the archives of Copernicus and Landsat missions, as well as products like the World Settlement Footprint (WSF), the on-demand processing of new data products and the possibility of developing and deploying new processors, U-TEP allows analyzing and visualizing project-related geospatial content with the VISAT tool. Multiple visualizations (scopes) are available for registered users, but can also be newly set up. One available scope exemplary illustrates the exploitation of the WSF-Evolution dataset by analyzing the settlement and population development for South-East Asian countries from 1985 to 2015 in the context of the Sustainable Development Goal (SDG) 11.3.1 indicator (Fig. 1). Other accessible scopes focus on green public space (SDG 11.7) or the interactive urban heat island modelling by entering new land-use scenarios by the user.

With high-volume, function-rich Web services on spatio-temporal raster data on the rise there is a growing need for adaptive access control. However, while the datacube paradigm has proven suitable for large-scale, efficient raster services it is not yet clear how appropriate access control mechanisms can be provided to service administrators. In this contribution we provide a model which extends the SQL datacube query language, SQL/MDA, with flexible definition and enforcement of array-aware access control.

MOTIVATION

Web services offering scientific data, such as satellite imagery and weather/climate data, are emerging in increasing numbers, based on recent technology progress. While there is a strong movement towards open data in general it is hardly questioned that often access has to be controlled tightly, such as in industrial and security/privacy critical governmental services. Even scientists enjoy protection of their data for some grace period to publish any findings first.

Typically, services consider data sets as a whole, i.e., a “all or nothing” access. However, with multi-Petabytes of Earth data getting online this is no longer adequate; rather, access protection also of spatial and/or temporal regions must be considered. For example, the European Centre for Medium-Range Weather Forecast (ECMWF) offers very long timeseries of climate data. While the long tail of data is accessible for free the most recent two weeks are priced. Further, services tend to evolve from pure data extract¬ion and download to more and more flexible, user-specified processing in the server prior to downloading results, which can become a cost-relevant factor to be safeguarded.

We propose array–aware access control based on a declarative array query language. SQL triggers serve for defining constraints on array access and processing, together with remedial actions executed whenever a constraint is violated. Such triggers are always applied to queries, except when the user firing the query has been exempted by the administrator.

STATE OF THE ART

User privilege management in standard databases is done through Role-Based Access Control (RBAC), although there is no specific support for (partial) array access. Declarative triggers consist of an event, a condition to be evaluated whenever that event occurs, and an action to be executed should the constraint be fulfilled. Triggers form a proven tool, but they likewise do not support fine-grain array access control. The MongoDB NoSQL system offers privileges listing allowed operations, but no fine-grain access control inside its objects. The SciDB Array DBMS defines access control only on sets of arrays which is too inflexible and coarse-grain.

Earth data services (including satellite imagery and climate / weather archives) typically resort to ad-hoc implementations unaware of fine-grain access control, rather relying on file-level (not pixel-level) access granularity. In the widely used OPeNDAP Hyrax server access control is a known open issue. A family of systems, including the Australian Geoscience Data Cube, offers python APIs with access on file basis and without array-specific authorization.

ARRAY-AWARE ACCESS CONTROL

To make SQL RBAC array-aware we enhance the existing SQL RBAC and triggers with array expressions, plus some novel application mechanisms. First, we introduce two predicates, ACCESSED(A) and MODIFIED(A) where A is some array-valued expression. The result is a Boolean array of the same size as A where each cell contains true iff this cell in A gets accessed by the query under consideration, or modified, respectively. Next, we connect triggers to users and roles. As normally a trigger is always checked we need to remove – i.e.: deactivate – it. To accomplish this we enhance the SQL GRANT statement with an option EXEMPTION FROM to exempt a user or role from a particular trigger. This yields a safe model for restrictions: in the first place, constraints always hold; they need to be explicitly and consciously disabled.

So far we have addressed unconditional access restrictions for particular users. Sometimes, however, this should be more dynamic. Quota, for example, should fire only in case some assigned threshold is exceeded. To allow introspection of query properties we define a context object providing information about data accessed, query result volume, and compute resources spent. This can be used to define various quota policies, such as limiting result sizes (see use cases below). The cost estimator of the query engine provides this information ahead of execution. Obtaining accurate estimates is an area of active research in our array query optimization work, to be published in a forthcoming paper.

USE CASES

The following use cases have been spotted initially, although we feel that there are many more to come:
• Protecting along time and space. By specifying one or more coordinate ranges, space/time sections in the datacube get protected. Such predicates may also express an overlap by at least N pixels, X percent of the forbidden area, or some other empirical measure.
• Protecting arbitrary areas. For some vector polygon the clip() operator (not currently SQL standard) is available will for disallowing access to its area:
• Protecting by mask. Any other object can act as a mask (binary or numeric) for pixel access.
• Quota. They can constrain queries several ways, including query result size, disk access volume, processing cycles, etc. Quota can be static or dynamic drawing from some predefined budget).
• Mandatory access control. An administrator can define an unconditional restriction on all objects in their entirety, of which users selectively can be exempted. Further triggers may govern more detailed policies.
• Upfront cost estimation. If allowed by policy cost estimates can be made available to a client in advance so that the client can decide upfront whether to fire off a heavy-weight query.
Some examples may illustrate the syntax. The following trigger protects sub-datacube region S from access, where S might characterize “the last two weeks”:

CREATE TRIGGER Latest_2_weeks_disallowed
SELECT ON ERA5
WHEN MDANY( ACCESSED( ERA5[ S ] ) )
BEGIN EXCEPTION "Error: no access rights on this area." END

Let P be given as a vector polygon over some area, expressed in WKT (Well-Known Text). Then, the following query will disallow access to the P area:

CREATE TRIGGER Protect_by_Area
SELECT ON ERA5
WHEN MDANY( ACCESSED( clip( ERA5, P ) ) )
BEGIN EXCEPTION "Error: no access rights on this area." END

Protection by some binary mask M is expressed as:

CREATE TRIGGER Protect_by_Mask
SELECT ON ERA5
WHEN MDANY( ACCESSED( ERA5 ) AND M )
BEGIN EXCEPTION "Error: no access rights on this area." END

WHEN conditions may contain any predicate allowed in the query language, including combinations of the above mechanisms.

CREATE TRIGGER Quota_on_Access
SELECT ON ERA5
WHEN MDCOUNT_TRUE( ACCESSED( ERA5 ) ) > 1000000
BEGIN EXCEPTION "Error: data access volume exceeded." END

CONCLUSION

For “Big Data” privacy and security, service attack protection, general housekeeping such as quota, etc., propose the “trigger as guardian” concept as a natural enhancement of the proven RBAC model. To the best of our knowledge there is no other approach published which conveys the same power, flexibility, and ease of use.

Array triggers are implemented in rasdaman. Preliminary experiments have shown that they do not impose a particular extra workload. Checking query costs a-priori, something impossible for procedural, interfaces (e.g., the python-based Australian Data Cube), is done by the rasdaman estimator. Declarative languages – like SQL – are known to be safer in this respect and, additionally, allow all sorts of automatic optimization of queries prior to evaluation, including parallelization and distribution.

Currently we are applying triggers in several projects (such as BigDataCube and Landsupport) and with various rasdaman operators from industry and academia to gain further experience.

ACKNOWLEDGMENT
This work is being supported by H2020 LandSupport, H2020 EOSC-hub, and German BMWi BigDataCube.

Online spatial data processing is becoming more and more popular. The availability of environments providing adequate functionality for this purpose is crucial in this type of IT market. One of the systems that give the possibility to integrate spatial data from many sources and process them in an advanced way is the Jupyteo online service by Wasat company. The original version of this system was developed under ESA contract. Jupyteo is an Integrated Development Environment, which gives the possibility to create any algorithms using many programming languages, including Python to facilitate work with spatial and Earth Observation data. In addition, the functionalities contained in it allow for a convenient presentation of the work results in a graphical form.
The purpose of presented work was to develop a methodology and tools for integrating data from various sources to create a spatial database to support touristic mobile application with Augmented Reality interface. The main goal was to acquire and integrate standard spatial datasources available online such as Openstreet, Sentinel and others and convert them to a schema compatible with the database format needed by the online mobile application presenting AR content to the user. Most of the work presented was done with the use of Jupyteo online service. 
Data processing algorithm was programmed in the form of a Jupyter notebook script using Python and SQL. The resulting database was placed on the PostgreSQL server with the PostGIS extension. From there, the data was made available for AR applications as well as for loading them directly in Jupyteo. As part of this work, the Jupyteo functionality was also improved by enabling convenient editing of data on AR points and their updating on the PostgreSQL server. In this way, Jupyteo has become a tool for acquiring, sharing and updating data for the needs of AR application, which originally was not directly related to this system.

The purpose of the work is to create and provide in the form of a web service the results of data processing from satellite radar interferometry. The authors used Jupyteo for this purpose - a new cloud-based Integrated Development Environment created by WASAT under ESA contract . It is based on Jupyter notebooks solution highly integrated with Docker isolation approach. It allows for EO data scripting by building own processing algorithms. Jupyteo integrates most common Earth Observation, geospatial and GIS based tools, software and libraries. User is able to use toolboxes for EO data (SNAP, OTB), vector and raster data processing (GRASS, GDAL, Postgre/PostGIS) and many others. Jupyteo also enables visualization of spatial data, from processing or other services (WMS, WMTS) in suitable form on the map. Jupyteo is highly customizable, user can add or remove any software element. The main programming language here is Python, but one can use other programming languages like Scilab, SQL, Javascript, Bash and others according to installed interpreter.
The conducted research comprised examining the deformation of the terrain surface with the use of radar interferometry. One of the most important aims of the research was the universality and simplicity of created solution. The best approach seemed to be creation of dedicated software. For this purpose the already mentioned Jupyteo system was used. The basis of implemented algorithms was SNAP Python interface. The application was created in the form of Jupyter notebook, thanks to which full control over the created data processing algorithm was obtained. Thanks to this, one can do similar research for any area in the world. In this way, the notebook created in connection with the Jupyteo functionality has become a tool for performing research based on data from satellite radar interferometry. The resulting tool also gives the opportunity to share test results, which after processing into a raster form can be made available as a WMS service available online to other users of the network.

The Sentinel-2 MSI represents a challenge for the Mission Performance Centre to monitor its health and radiometric performance. As there are 12 detectors and 13 spectral channels, the existing approach for monitoring the radiometric performance is a big task. The MSI radiometric monitoring activity should therefore be automated as much as possible. To help with this task, a machine learning system could be developed and implemented that automates the trending and monitoring of MSI radiometric calibration datasets. This machine-learning framework is used to capture the time dependent trends of these datasets through a training phase. Because the time dependent trend for a dataset is highly sensitive to changes above its noise level, outliers that deviate from the existing pattern are detected by comparing the values of a dataset against predictions of its time dependent trend. Anomalies in the machine-learning framework are defined as unexpected data pattern changes, which are quantitatively characterized by dimensionless parameters. These parameters are used in a clustering analysis to separate anomalous datasets from the normal ones. In addition, the machine learning outputs enable the assessment of relative detector data quality by comparing the noise level of a detector with the average behaviour for a given channel. A detector with a higher noise level has significant impact on the radiometric accuracy. Initial results from radiometric monitoring show that many of the anomalies characterized in the machine learning approach cannot be detected through the existing static approach (Li et al. CALCON 2018). The machine learning approach brings fundamental advances in trend and anomaly detection, and it should lead to a more dynamic and autonomous monitoring of MSI radiometric performance.
Using all L0 radiometric calibration acquisitions (sun-diffuser images and dark-signal images) since 2016, this demonstrator of Artificial Intelligence (AI) will detect and classify defective pixels, and compute L1 radiometric calibration coefficients.
The learning phase will use several years of calibration coefficients and lists of defective and noisy pixels from the sun-diffuser images and dark-signal images available from performed analyses at CSSI since the launch of the Sentinel-2 mission. After the learning phase, the AI tool should be able to ensure reliable calibration and classification tasks much faster than human power or conventional algorithmic tools.
This new AI tool should also improve significantly the quality of the radiometric calibration results and should be able to detect the different kinds of defective pixels (noisy, RTS, cosmic impacts, reset spikes), avoiding any human mistake. The demonstrator results will be validated by the optical radiometry expert team at CSSI.
This presentation will describe the complete AI system and will provide the demonstration of improvements in radiometric quality and in processing time.

Extreme sea levels are a major hazard worldwide. While regional and local differences are important factors influencing the translation of a storm surge into coastal inundation, data availability continues to limit our ability to quantify present and future storm surge risks. Tide gauge sea level measurements and predicted tides for tide gauge locations provide important data for extreme sea level analyses. Tide gauge records typically extend back over multiple decades and often more than a century. This makes them particularly applicable for the estimation of extreme sea level return periods, in contrast to satellite altimetry data. However, even in locations with long periods of data records, missing data together with instrument inconsistencies constitute a problem. Data gaps might range from one observational point to days or months, or even several years. Together with systematic missing values coinciding with extreme sea level events, data gaps may influence the estimation of extreme sea level return periods. The extent of this influence has not been widely investigated, however. The present paper aims to evaluate the potential of a machine learning approach to time series data gap filling and the effect of doing this on standard extreme sea level analyses.

First, extreme sea levels are estimated using the state-of-the-art Skew Surge Joint Probability Method (SSJPM). Extreme sea levels are estimated for tide gauges located along the UK East Coast, including the tide gauge of Immingham, for return periods based on a 30-year data record (1988-2017). The upper tail of the extreme value distribution for the respective record is extrapolated with a Generalized Pareto Distribution (GPD). In a second step, an artificial neural network (ANN) is trained. Different algorithms are assessed to train a backpropagation feedforward multilayer neural network with data from two input vectors of tide gauges of neighbouring locations. In a third step, the neural network is applied to fill the data gaps, with a subsequent sensitivity analysis of the estimated extreme sea level return periods.

Satellite synthetic aperture radar (SAR) missions are important instruments for wide area monitoring thanks to their ability to provide all weather day and night imagery. Thanks to developments in SAR technology and the greater availability of SAR images, it is now possible to exploit a huge amount of data in order to use Deep Learning (DL) techniques. Almost all the state of the art techniques for classification use DL methods. In particular, in the field of image classification, the most used techniques are based on supervised learning approaches, where the training set is composed of pairs consisting of input and ground-truth images.
The main problem of Supervised Deep Learning techniques in SAR tasks is the need of big labelled datasets in order to capture significant features and learn a good model avoiding overfitting (i.e. the problem of fitting noise and disturbances contained in the training data set, thus reducing generalization and predicting capabilities). This issue usually happens when the network has a number of parameters relatively bigger than the size of dataset.
Unfortunately, in these tasks, labels are often difficult and costly to obtain, and a different procedure is required to overcome the problem.
Another relevant problem in deep network is the “vanishing/exploding gradient”, due to the exponentially dropping or increasing in magnitude of the gradient in the backpropagation algorithm in networks with more than three layers.
To overcome the mentioned problems, we propose a new semi-supervised approach, consisting of a three-phase procedure:
- Pre-training phase using a Convolutional Autoencoder (CAE) on unlabeled data.
- Extracting the Encoder part of the CAE and add some layers to perform classification.
- Training phase using a Deep Neural Net on labeled data.
This approach let us overcome the small labeled dataset problem while exploiting the huge number of unlabeled data in the pre-training phase by changing net parameters in order to assume a better configuration and start the minimization problem in a point closer to the optimal global minimum.
In addition to that, other techniques are used, like Data Augmentation to create more data, Dropout and Max-pool layers to avoid overfitting, Convolutional layers that extract features through filters and ReLU activation functions that increase the non-linear properties of the network.
We have applied the proposed technique to ship classification, an important issue in maritime applications, in particular for ocean pollution monitoring and illegal fishing monitoring. The tests performed shows very promising results and confirm the validity of the proposed deep learning approach.

In order to support the analysis of infrastructure and building stability, it can be very useful to exploit outlier-detection machine learning techniques, with the purpose of reliably and automatically detect anomalies in the deformations of infrastructures and buildings measured by synthetic aperture radar interferometry (InSAR).

InSAR is a technique that processes a stack of SAR images, detects the structures that have a stable radiometric backscattering (Persistent Scatterers, PSs), and measures their displacement (along the line-of-sight of the satellite) and 3D position, with millimetric and metric precision respectively. The result of the typical InSAR processing is a collection of tens of thousands of PSs per square kilometre (and for each PS, tens of displacement measurements corresponding to the satellite acquisition times), which describe the displacement of the area with a dense but sparse collection of measurements. This wealth of information is extremely meaningful, but at the same time requires expert knowledge to obtain a concise description of the underlying phenomena, and to detect interesting spatial or temporal trends that can be useful for stability early warnings. Therefore, it is very useful to develop techniques that can support the decisions of the end user through automatic information extraction from the PS measurements.

The detection of anomalies in InSAR time-series is generally performed with empirical techniques, such as those described in [1] and the works cited therein. In this work, we explore the possibility of using AI techniques for anomalies detection, in order to improve the performance achieved.

Outlier detection is a thriving field of study, yet our field of application is peculiar, insofar the identification of early warnings for infrastructure stability depends on the detection of differential displacements within the same structure, which in turn depends on both the detection of spatio-temporal outliers and their connection with phenomena that could affect infrastructure stability. Furthermore, usual approaches to outlier detection in time-series data suppose data available at regular intervals, while in the typical InSAR scenarios the acquisitions are irregular. For this reason, among the different outlier detection techniques, we investigate the use of the Multi-scale Anomaly Detection approach, which overcomes the need for a regularly acquired displacement data.

Another peculiarity of InSAR series outlier detection is that infrastructures are affected by some phenomena (such as the seasonal thermal expansion) that are normal and do not constitute an anomaly and a source of early warnings. Therefore, we need to include in the techniques some models for physically permissible displacement evolutions in order to improve the detection of actual phenomena that are dangerous for the infrastructure stability.

Early results are promising and confirm the validity of the proposed approach.

[1] M. Costantini et al., “Automatic Detection of Building and Infrastructure Instabilities by Spatial and Temporal Analysis of Insar Measurements,” IGARSS 2018 - 2018 IEEE International Geoscience and Remote Sensing Symposium, Valencia, 2018, pp. 2224–2227.

The Qinghai Tibetan Plateau (QTP) is the highest and largest plateau system of the world. It provides important ecosystem functions for the climatic and hydrological system of east-Asia. Over the last 30 years land-use changes have widely occurred on the QTP. Some consequences include an easily observable growing infrastructure and accelerated vegetation dynamics. Depending on the location on the QTP, the latter can include both gain and loss in biomass and vegetation cover as well as shifts in species compositions. Understanding vegetation changes on the QTP and their drivers is of crucial importance to ensure the sustainability of the QTP.
Many earlier attempts to explain the vegetation trends across the QTP were based on satellite remote sensing data to cover its vast extent. Satellite observations allow to reconstruct vegetation dynamics across at least the last 30 years using time-series of well-known, reflectance-based proxies for vegetation abundance such as the Normalized Difference Vegetation Index (NDVI). While the trends can hence be re-constructed quite eastily, the most important drivers for the satellite-observed vegetation dynamics are still not fully understood and widely debated in the literature. Here, I use the multi-faceted environmental variables provided by the Earth System Data Cube (ESDC) of the Earth System Data Lab (ESDL) to explain the occurrence of negative vegetation trends across the complete QTP. The vegetation trend product was derived from annual Landsat NDVI time series at a pixel size of 30 m (see Fig. 1) as well as from a corresponding product derived from MODIS data at 500 m pixel size for the time period 2000-2018. To relate the the environmental variables provided through the ESDL with the vegetation trends, fractional covers of degrading areas were calculated for a 0.0083 ° grid covering the complete Tibetan Plateau and matching the ESDC’s highest available resolution. Then, random forest models between the %cover of degrading areas were calculated for both, a global model covering the complete Tibetan Plateau as well as many random forest models implemented in a spatially-adaptive approach with a moving window that repeatedly calculates models for local subset of the Tibetan Plateau. The results of preliminary analyses show that the most important environmental variables explaining the variation in %cover of degrading area vary widely with scale (both extent and grain). This is expected but was so far hardly discussed in the scientific literature and has important consequences for the derived conclusions concerning suitable mitigation strategies to prevent further grassland degradation on the Tibetan Plateau.

High resolution push-broom satellite sensors are very sensitive to vibrations, in particular thoses due to momentum wheels responsible of satellite orientation. These vibrations cause geometric distorstions on images. In addition, these distorsions cause georeferencing issues. Because of their frequencies, theses vibrations are poorly detected by onboard stellar sensors. They need to be corrected. Usual methods to detect vibrations are based fourier tranform to detect harmonics or dense correlation. These method are efficient for fixed shot (star shots...) but demonstrate their limits in nominal shots, confusing platform vibrations and landscape scrolling. In this study, we propose two methods using deep learning in order to detect and correct such vibrations. The first method tries to try to correct these vibrations using a single neural network. Different architectures of networks have been tested such as CNNs and GANs. These deep neural networks are trained to directly learn the mapping between distorted images and corrected ones. They are trained on data from different sensors : Pleiades, Venus and Sentinel-2. The second method splits this approach in two different steps. In the first step, a CNN is trained to produce shift maps. These shift maps represent the x and y shifts for every pixels of the input image. Then, these shift maps are re-used to produce time-referenced data. A recurrent neural network (RNN) is then trained on these data. Finally, this RNN is used to predict futur vibrations with which we can correct distorsion.

We present a robust algorithm for detecting changes in a time series of images. The proposed paradigm makes uses a sequence of satellite images to improve the reliability of a given pairwise change detection method. The approach is based on calculating the pixel-wise minimum change magnitude over a set of pairwise change comparisons. The idea makes use of the transitivity assumption of pixel values during a constrained time period, that is, x ≤ y and y < z implies x < z, where x, y and z represent pixel values of three sequential images.

The process is designed for earth observation data to decrease the number of pixels with high change magnitude in pairwise comparisons that do not represent actual changes on the ground. These differences may be caused by, for example, differences in the atmosphere, weather or season.

As a preprocessing step, the following procedure is proposed to remove cloudy images. First, the area of interest is split into geographically defined rectangular blocks in a specific coordinate system. Then, satellite data is acquired from the time period of interest and cropped into these blocks, producing a time series of satellite image tiles for each block. All tiles for which the data coverage is less than 100% are discarded.

The next phase involves the discarding of abnormal tiles for each block, that is, tiles that are significantly different from the majority of tiles in the time series. This is executed by studying the distributions of the values of the satellite data bands for each pixel. The purpose of this phase is to remove the images that contain clouds, shadows of clouds or other elements that could interfere with the change detection.

Using the AutoChange algorithm as a change measure, the time series process is applied to Sentinel-2 satellite images from 2015-2018 to detect changes in Finnish forests. The results are verified by the Natural Resources Institute Finland, suggesting that the precision of the proposed fully automated processing chain is up to 99.4% for clear-cuts and 91.7% for thinnings.

Due to these results, the Finnish forest law enforcement executed by the Finnish Forest Centre was permanently altered. Since 2018 the forest law enforcement related to loggings has been based on satellite change detection.

We present results of the Wavelet analysis of local manifestation of Halloween geomagnetic storm (from 29 to 31 October, 2003) in the middle latitude using in situ geomagnetic and TEC minute data. The case study analyses are performed for three geomagnetic observatories (GOs) Ebro EBR (Spain), Surlary SUA (Romania), and Beijing Ming Tombs BMT (China) as well as for closely located to them permanent IGS sites EBRE, BUCU, SOFI, and BJFS. By means of the wavelet coefficients, we have carried out a comparative analysis of the variations of the planetary geomagnetic index Dst, H-component of the geomagnetic field registered at three GOs, and VTEC values obtained from GNSS observations in the IGS sites. Wavelet Analysis permits to distinguish very clearly the ionospheric from the ring current influence on the geomagnetic DS index in the high frequency domain. This result which will be pursued in future research seems to be related to soliton-like waves (with periods about 40-60 min) in the ionosphere. Belashova and Belashov have theoretically modeled a generation of soliton-like wave “forerunners” with periods ~40–60 min for geomagnetic latitude = 450 (see Belashov, V. Yu, E. S.Belashova, 2015. Dynamics of IGW and traveling ionospheric disturbances in regions with sharp gradients of the ionospheric parameters Advances in Space Research, Vol. 56, Issue 2, 333-340).
Soliton-type plasma structures with periods lasting from 30 min up to 3 hours can be associated with group structures on the wavelet analysis plots for the DS index. Similar phenomena of IGW and TID are extensively studied in the last several decades by many authors: see e.g. the papers Heisler, L.H., 1959. Occurrence of giant travelling ionospheric disturbances at night. Nature 183, 383–384; Hunsucker, R.D., 1987. The sources of gravity waves. Nature 328, 204–205; Hocke, K., and K. Schlegel, A review of atmospheric gravity waves and traveling ionospheric disturbances: 1982–1995,Ann. Geophys.,14, 917–940, 1996.

We may conjecture that short-period phenomena identified on the wavelet graphs may be explained by the dynamics of 2D-solitary nonlinear internal gravity waves (IGW), as well as traveling ionospheric disturbances (TID) of the electron density excited by them at heights of the ionosphere F-region.
We provide also analysis of the same parameters during a quiet phase. It is remarkable that one still observes non-negligible indications for soliton-like waves having smaller periods. Let us note that the availability of fast wavelet transforms is in-line with the Big Data paradigm and permits the consideration of arbitrary high frequency data.

This research has been partially supported under the grant DH-02/13 “Modern Mathematical methods for Big Data and applications” with Bulgarian NSF.

Sea surface temperature (SST) is an environmental variable of interest for climate monitoring and oceanographic modelling. The Copernicus Marine Environment Monitoring Service (CMEMS) aggregates data from multiple satellite sources to provide a daily, global SST map at a spatial resolution of 1 km. This resolution is derived from current satellite instrumentation and is well-matched to the typical length scale of SST variations in the deep ocean. However, in coastal regions the SST can vary substantially on the sub-km length scale as a result of various phenomena including upwelling [1] and river discharge [2]. These variations are not captured by the coarse, daily SST map, yet such fine detail can provide important information to industries such fisheries as well as for ocean current modelling and sea rescue efforts.

The instrumentation onboard Sentinel-3 includes thermal infrared bands to measure SST. These bands have a spatial resolution of 1 km and produce image series with a temporal resolution of 2 days for mid-latitudes. Landsat 8 carries the Thermal Infrared Sensor (TIRS), capable of recording SST with a spatial resolution of 100 m and temporal resolution of 16 days. No current or planned satellite is able to combine high spatial and temporal resolution for the measurement of surface temperature.

The objective of image super-resolution is to recover high-resolution images from low-resolution images. For thermal satellite imagery acquired over land, the TsHARP algorithm [3] has proved successful because of its operational simplicity and effectiveness. However TsHaRP is not applicable to regions containing water (i.e. seas). The advances in deep learning convolutional neural network (CNN) architectures have also been applied to the super-resolution problem [4], with the Super-Resolution CNN (SRCNN) showing superior performance over established, “classical” methods [5].

In spite of the wide scientific and commercial interest in the ocean, there are few published articles focused on super-resolution of sea scenes, compared to land scenes. In this submission we address this deficit by investigating resolution enhancement of cloud-free thermal images of the coastal waters around the Maltese Islands. Due to the small land mass of the country, and the various features of the highly structured coastline, accurate coastal SST variations are not captured by CMEMS products. A large thermal image dataset was constructed from near-coincident passes of Sentinel-3 and Landsat 8 satellites. We show preliminary results from training a SRCNN to learn an end-to-end mapping between the low resolution Sentinel-3 and the high resolution Landsat 8 thermal images. This process yields a series of SST images with the unique combination of spatial resolution of 100 m and temporal resolution of 2 days. We also show the image pre-processing steps which enhance the super-resolution performance, including destriping and denoising [6]. It is hoped that our work will lead to unique SST datasets of direct use for ocean and climate modelling tasks thus demonstrating the promise of modern super-resolution techniques for enhancing satellite-derived sea images.

References
[1] D. A. Gagliardini et al., “Detection of small-scale coastal oceanographic processes through Landsat-TM/ETM+ images: implications for the study of biological processes along the Patagonian coasts of Argentina”, Gayana 68, 2 (2004)
[2] C. J. Gleason and L. C. Smith., “Toward global mapping of river discharge using satellite images and at-many-stations hydraulic geometry”, PNAS, 111, 13 (2014)
[3] W. P. Kustas et al., “Estimating subpixel surface temperatures and energy fluxes from the vegetation index–radiometric temperature relationship”, Remote Sens Environ, 85, 4 (2003)
[4] Z. Wang et al., “Deep learning for image super-resolution: a survey”, arXiv:1902.06068v1 (2019)
[5] C. Dong et al., “Image Super-Resolution Using Deep Convolutional Networks”, IEEE Trans. Pattern Anal. Mach. Intell., 38, 2 (2016)
[6] D. T. Lloyd et al., “Noise and Striping Suppression in Landsat 8 Thermal Infrared Sea Images“, accepted for presentation at: GHRSST XX Meeting (2019)

Forest ecosystems are under severe degradation because of population pressure amplified by climate change. Remote sensing data and geographic information systems (GIS) analysis techniques are recognized as a powerful approach to obtain information regarding mapping and monitoring forest species . In this work, Sentinel-2A MSI data and field survey data were used to map forest species in the Azrou forest situated in the Central middle atlas of Morocco. Four classes (holm oak, cedar forest, bare soil, and others-unclassified) were selected; they represent, respectively, 27, 11, 24, and 38% of the study area.The overall accuracy of classification was estimated to be around 99.72%. Thus, mapping land cover via Sentinel-2A might be sincerely a good tool of sample-based National Forest Inventories at regional scale.

In the last few years, many private or public investors encouraged the use of the growing amount of satellite imagery in order to respond to global needs. Urban planning, environmental monitoring and other applications in socio-economics studies are of interest.
More precisely, road and building extractions represent a key to understand human societies as they describe both the structure of human exchanges and the population distribution. However, the methods used to identify and classify roads, buildings or houses were inaccurate until a few years ago and mostly limited to particular cases. Buildings shading other buildings are problematic for these methods and require new and improved treatment processes.
Exploiting satellite imagery for building extraction is a real challenge, that Deep Learning algorithms might solve. One issue remains as unsupervised learning for building detection has not been explored in-depth yet. We must train our neural networks with labeled datasets, i.e., using supervised learning. Still, labelling large images such as satellite imagery is a tedious and expensive task. Challenges such as “DeepGlobe” competition provided labeled dataset allowing supervised training for Convolutional Neural Networks (CNN). Previous results using neural deep neural networks tend to be data specific with the lack of generalization. It means that the CNN trained on a specific dataset (satellite imagery with a particular resolution or specific spectral bands) tend to weakly extract building footprints on data completely different from the training dataset.
We propose to modify existing neural networks for building extraction, TernausNetV2, DAN and others, to obtain a neural network which can output an accurate binary mask (i.e. each pixel is building/non-building) on various datasets. To this end, we only use RGB channels with different resolutions (30 or 50 cm resolution) as input features. Generalizations to lower resolutions (GSD of 1meter) will be presented, using super-resolution techniques to artificially increase the resolution of the satellite images. Aggregation of predictions with two or more neural networks will be shown.

Super Resolution, a classical problem in computer vision in image processing science aims to recover a high-resolution image from a single low resolution one. The problem is clearly ill-posed since several solutions exist for a given low resolution pixel.
Increasing resolution, clearness and reducing noises of satellite imagery with digital filtering methods are one of the most important research axes in satellite image processing since the last decades. Classical methods have solved this problem with digital filters and engineering methods.
Nowadays, new super resolution methods based on deep neural networks are using parameters that are learned from a training set of examples by minimization methods of a loss function.
These methods have the advantage of aggregating the information stored in millions of free parameters of a large number of examples on the selected images from a learning database of images.
Super resolution methods have proved their effectiveness on photographic and medical images. In order to apply them in satellite image processing, the physical properties of satellite images must be taken into account when developing the deep learning neural network.
We show that Deep Generative Adversarial Networks (or GAN) are suitable candidates for increasing the spatial resolution of SENTINEL-2 images from a set of pair images of SENTINEL-2 (high resolution) and PLEIADES (very high resolution).
GANs are deep neural network architectures composed of two networks (Generator and Discriminator) pitting one against the other (thus the « adversarial »). GANs are class of AI algorithms used in Unsupervised Machine Learning, however the discriminator part of the network belongs to the category of supervised algorithms.

The novelty of this method is to use a target image set (PLEIADES) from a different satellite than the input image (SENTINEL-2). The main goal will be to be able to reproduce images with PLEIADES’s resolution from SENTINEL-2 images.
For this, it is important to consider the physics of the input data to be in agreement with the use of the algorithm in terms of contribution of adjacent environment. After training the network, we studied the error distribution and the sensitivity and robustness of the model by different measures of image quality in order to have a common reference for the comparisons of the performance of the different methods.

SAR Interferometry (InSAR) allows the generation of digital elevation models (DEMs) useful in numerous areas of Earth sciences and human activities, such as remote sensing, geology, water works, agriculture, mining, urban runoff projects, assessment of risk and environmental impacts, among others.
However, depending on the wavelength and processing method used, the final InSAR-generated DEMs can be different from a digital surface model (DSM) such as those generated by photogrammetry, and from a digital terrain model (DTM) such as those obtained with light detection and ranging (LIDAR). Additionally, presence of noise in the radar signal makes phase unwrapping not trivial, and atmospheric effects can also lead to degradation of the quality of the generated models.
A quality enhancement in InSAR-generated DEMs can be achieved from product adjustments that better approximate the reality that the DEM tries to represent, that is to say, the terrestrial surface under investigation. In this way, products can be obtained at relatively low operating costs in relation to field surveys using Global Navigation Satellite System (GNSS) receivers or aerial photogrammetry. However, it is of fundamental importance to determine the final application of the DEMs, in order to determine the suitable SAR acquisition mode to use (e.g. StripMap or SpotlLight) and the optimal tolerances in accuracy and precision.
In this work, we study the precision, accuracy, goodness and representativeness obtained in the adjusted and validated DEMs with GNSS data measured in the field. The proposed methodology consists in a workflow including georeferencing, adjustment, validation, calculation of accuracy and precision, analysis of the result of the model and its associated errors. Georeferencing consists of a linear displacement in X, Y and a linear displacement in Z. Subsequently, the differences between the true data (i.e. GNSS data) and the measured value (i.e. InSAR DEM) are calculated to construct an adjustment surface (non-linear deformation in Z).
Our input data are InSAR DEMs generated based on ASI’s COSMO-SkyMed images using ESA’s SNAP software, and a network of control points obtained by field surveys with double constellation receivers (GPS + GLONASS) and double frequency (L1 / L2) model Trimble R6 and processed with proprietary software. It is to be considered that, for the purposes of this work, GNSS control points with negligible error are considered. In addition, the network of control points is separated into adjustment points (70% of the total) and validation points (30%). For the processing of adjustment the use of the Julia programming language is proposed.
The final purpose of this work is to make a step-change in the quality of InSAR-generated DEMs delivered to the final users, such as surveyors and engineers. On the other hand, the adjusted DEMs can be used in the future to contrast DEMs obtained using other radar wavelengths (e.g. C-band Sentinel-1 and L-band SAOCOM data) or generated by other surveying methodologies.

Land use classification and crop identification from remote sensing images have been a recurrent topic for the last years. With the availability of open and freely available medium resolution images such as Sentinel 2 and Sentinel 1 there are unprecedented opportunities for the production of reliable maps of crop type. However, what is currently still missing is sufficient training data of crop type labels. This research proposes an approach of Machine Learning (ML) which aims at automatic labelling of pictures such as geo-tagged photos (e.g. LUCAS, Mapillary, Flickr) to derive reference datasets.



In recent years Deep Learning DL has shown to perform very well and achieve high recognition rates, given that it possess a powerful capability for extracting features directly from large image datasets. One of the particularities of DL is that it surpresses, to some extent, the need for human invovement -e.g. feature extraction, avoiding some bias introduced by the change of the feature space. One of the widely used DL architecture is the Convolutional Neural Network (CNN), because of its successful results on image recognition and classification by using a sequence of feed-forward layers. Results of the DL based picture labelling and the associated accuracy will be presented. Initial results show that high accuracy can be achieves for 10 major crops. Further application of the reference dataset will be discussed.

Land use classification and crop identification from remote sensing images have been a recurrent topic for the last years. Due to its complexity, studies have revealed that the same crop has different spectral resonses at various stages of its growth and at the same time different crops may have the same or a similar spectral resonse, due to seasonal changes. There has been a big hipe in the recent years on the potential of applying Deep Learning (DL) algorithms, given that they possess a powerful capability for extracting features directly from large image datasets, and also they achieve high recognition rates. One of the particularities of DL is that it suppresses, to some extent, the need for human invovement -e.g. Feature extraction, avoiding some bias introduced by the change of the feature space. One of the widely used DL architecture is the Convolutional Neural Network (CNN), because of its successful results on image recognition and classification by using a sequence of feed-forward layers. The DL algorithms will be applied to different picture archives namely lucas, mapillary and other geotagged crop type pictures.

The rise of waters can be caused both by the rise of sea level and by the subsidence of Earth surface. The combination of these two factors is the "Relative Sea Level rise".
The rise in sea level is expected to considerably aggravate the impact of coastal hazards caused in the coming years. Potential impacts are expected to be significant, for example the increase in the frequency of floods, the acceleration of coastal erosion, as well as the imminent slump of coastal areas from salty waters. The impact depends not only on the intensity and extent of coastal changes, but also on human response and adaptability to hazards including the type of exposure.
Coastal areas with very low altitudes are particularly vulnerable to the relative sea level change. In particular, for coastal urban centers the risk is increased due to high exposure. Coastal protected areas are also of great interest due to the high conservation and protection of the special characteristics of the natural environment. Land settlement on a local scale can make a significant contribution to the relative sea level rise, so it is necessary to understand the rhythm and spatial distribution of potential surface deformation in coastal areas.
The current project aims at the implementation of interferometric techniques such as PSI and SBAS based on SLC SAR images from Sentinel 1 A & B satellites. The interferometric results will be combined in a GIS environment with the rates of sea level changes derived from Altimetry, morphology (using a high resolution digital embossed model), and land use types for study area.
The scope of the project is to create a scenario of future relative sea level rise through cartographic imaging with the aim of spatially-time assessment of vulnerability of coastal areas.
The pilot coastal area of Thermaikos Gulf at northern Greece is presented which is characterized by the presence of the city of Thessaloniki (established as the second most important urban centre of Greece), an extended industrial zone, locally tourism infrastructure, an international port that constitutes the centre of merchant shipping for the Balkan countries, the Axios river delta and the smaller Gallikos river delta. Their deltaic plains, with a bird-foot shape, indicate the dominance of fluvial over marine process in forming the most active part of the Thermaikos shoreline.
The project, to which academic advisor is Prof. Issaak Parcharidis, is in progress with a duration of 15 months is co-financed by Greece and the European Union in the framework of the Operational Programme: Human Resources Development, Education and Lifelong Learning. The contribution of this research to scientific knowledge is primarily the determination of land deformation using SAR Interferometry (from 1992 till today), which combined with existing altimetry data, will allow the relative sea level to be calculated in three pilot cases: the of city of Alexandroupolis and the Evros Delta, the coastal zone of Thermaikos Gulf and the coastal area of Killini-Araxos which bring together the characteristics of interest.

The rate at which global climate change is happening is arguably the most pressing environmental challenge faced today. The United Nations Framework Convention on Climate Change (UNFCCC) provides the vehicle for multilateral action to combat climate change and its impacts on humanity and ecosystems. In order to make decisions on climate change mitigation and adaptation, the UNFCCC requires a systematic monitoring of the global climate system.

The objective of the ESA Climate Change Initiative (CCI) programme is to realise the full potential of the long term global Earth Observation archive that ESA together with its Member States have established over the last forty years, as a significant and timely contribution to the climate data record required by the UNFCCC. Since 2010 the CCI programme has contributed to a rapidly expanding body of scientific knowledge on 21 Essential Climate Variables (ECVs), provided by 450 scientists spanning 178 European institutions under the programme.

AI has played a pivotal role in the production of these CCI climate data records. Eleven CCI projects - Aerosol CCI, Cloud CCI, Fire CCI, Greenhouse Gases CCI, Ocean Colour CCI, Sea Level CCI, Soil Moisture CCI, High Resolution Landcover CCI, Biomass CCI, Permafrost CCI, and Sea Surface Salinity CCI - have applied AI in their data record production and research, or planned specific AI usage in their research roadmap.

The use of AI in these CCI projects is broad and varied, for example using Machine Learning (ML) to detect burned areas in the Fire CCI project, neural networks to retrieve dust Aerosol Optical Depth from thermal infrared spectra in Aerosol CCI, as well as pixel classification through a custom-built AI tool in Ocean Colour CCI.

Furthermore, cross-ECV science challenges potentially benefiting from AI have been identified by six CCI projects. These challenges include the use of ML for the analysis of multi-decadal records for extreme events across different ECVs, and the estimation of other variables such as photosynthesis by combining various different ECV datasets within an ML algorithm.

Moreover, the CCI community has identified potential future challenges in applying AI to their science work. These include leveraging GPU clusters, as well as the use of CCI data standards to prepare data for cross-ECV research.

ESA Φ-week 2019 - Abstract Submission
9 – 13 September 2019 Frascati, Italy [Rome]

ABSTRACTION / SUBMISSION

THEME:

"AGRICULTURE: A NEW FRONTIER FOR EUROPEAN SPACE POLICY"

Access to data has become a central issue for the agricultural sector. And could benefit from
European space programs.

From field to space, there is only one step. With the development of precision farming, access to data is becoming more and more central to the agricultural sector.
The aim is to improve the knowledge of farms through new technologies and ultimately to adapt production techniques. While some of this data is collected by terrestrial tools (drones, sensors), it is also produced on a larger scale by the many satellites around Earth, including those financed by the European Union as part of its policy Space.

Towards a new common agricultural policy

The post-2020 review of the CAP is part of a new thirty-year cycle, as was the one in 1992 marking a fundamental change with the policy defined in 1962. We then moved from a development-oriented CAP to remedy the dependence with a policy of limiting overproduction and focusing on rural development. We are now entering a new period that will have to respond to new challenges, which will be those of the 21st century: feeding 10 billion people, replacing renewable energies with fossil fuels and improving the impacts of human activities on the climate. Agriculture can play a vital role and these challenges are opportunities for European farmers. The developments in the CAP since its inception, particularly since 1992, have been based on rules governing the means. The success of the new objectives will have to be based on a new deal based on an obligation of results. Creativity must take precedence over procedures. In the "toolbox" at our disposal, the data and services derived from space systems offer credible solutions to meet the future needs of farmers. This is the direction in which I oriented my intervention by emphasizing three points: expertise, forecasting, risk management. It is essential to strengthen the ability to digitize expertise through the development of artificial intelligence-based tools, and space-based data can contribute to this. Our needs do not stop there: spatial data can also help forecast yields, stabilize markets and set up production strategies. They can contribute to the evaluation of the environmental impact and thus to the definition of virtuous strategies generating profits. They will be necessary for the evaluation of the risks, for the measurement of the impacts of climatic and environmental hazards and consequently for the granting of financial compensations.

Galileo and Copernicus: services already available

Galileo and Copernicus are operational and can already provide very high quality services. It was emphasized at the symposium that the data and information provided by these programs are free, transparent and accessible to all. Galileo is a satellite positioning system developed by the European Union. The expected accuracy is unparalleled. Copernicus is a European Earth monitoring program. Together they represent a real opportunity for the deployment of new strategies.

Pierre Delsaux in his speech focused on synergies between space and agriculture. He said that while the potential of space solutions has not yet been fully exploited, the space is already providing very high quality services to the agricultural sector.
EGNOS (European Complementary Geostationary Satellites Navigation Service) and Galileo already provide precision improvements: for tractor navigation, for locating livestock and for measuring surfaces. Copernicus data provides essential information on the state of the crops and the forest sector.

Delsaux also pointed out that space is not the only source of information available to farmers since hybrid systems combining spatial data with data from different sources (drones, data processing, etc.) can also be used. . The Galileo and Copernicus programs will soon offer new possibilities. As far as Galileo is concerned, the European Commission has agreed with the Member States to move to 20 cm of measurement accuracy by early 2019. For Copernicus: DIAS will be set up for summer 2018 to meet the massive amount of data and information collected. In addition, the European Commission plans to expand the program to address post-2020 climate change issues.

Concrete applications

Satellite imagery can facilitate the monitoring of information reported by farmers in the context of the CAP by integrating combined Galileo and Copernicus data into agricultural equipment to achieve better technical, economic and environmental performance. In his speech, Maurice Borgeaud presented some services that ESA offers in Europe and around the world, such as crop monitoring in Ukraine or the assessment of rice growth in Vietnam. Gérard Dedieu for his part identified the challenges related to the increased amount of data and information collected by satellites. How to identify the owner of the data and what types of property rights are related to the data. What is the availability of long-term data? What are the possible aggregations of spatial data, meteorological data with data from drones and farmers? How can pre-processed data be used for service delivery? David Hello presented Pixagri services he offers farmers, and Œnoview for growers, both based on satellite imagery. TerraNIS acquires images from various sources, processes them and derives the situation on the ground. It can thus provide decision-making services to farmers and winemakers. The Commission is willing to promote such services at European level, so does our independence of decision. It plans to allow the use of all available resources to enable SMEs to develop new services.

The new agricultural deal at the heart of space strategies for Europe

Agriculture is a pivotal sector to address global challenges. It is at the epicenter of the major issues of tomorrow. Developing our capacity to feed the planet, contribute to the energy transition and mitigate the human impact on the climate will be essential tomorrow. Data from space represents an important potential to accompany us in the changes that have become unavoidable. There is a need to identify barriers to the use of space services for agriculture, to guide future regulations to enable a positive contribution of spatial data to new agricultural challenges, to foster innovation and to provoke meetings on the occasion of dedicated events, space and agriculture. Taking agriculture into account in space policies will not do everything, but it is part of the solution. Tomorrow, feet in the ground and a little head in the stars that can be built our future.

cordially

Dr. LEBEAU PEMHA THINA: Chief Executive Officer and Founder AIPEA / AIPIA / ACAEPB and LAUREAT: SCIENTIFIC POSTER at the HIGHROC International Science Conference from 7 to 9 November 2019 in Brussels, Belgium.

Contact: 6 rue Ambroise Croizat 93200 Saint Denis, France
Phone: +33 7 67 94 53 21 or +33 9 50 85 90 60
Site: www.aipea-aipia.com

From 2012-2016 CubeSat launches grew from 23 to 230 per year; testing new sensor technologies, imaging and measuring the Earth, and providing universities a platform for training and inspiring the next generation of engineers. All of those platforms are collecting and transmitting important data, but where is it all going? There has been discussion and progress in the industry about commodity hardware and the democratization of space, but not as much conversation around the ingest, archive, and processing systems that make that data available to end-users. This talk explores the implications of the growth in SmallSat and drone remote sensing and advances in scalable, cloud based approaches to data integration and access.

Element 84 has been working on a number of systems to fill the gap and offer small and CubeSat teams powerful, standardized options for dealing with the full processing pipeline of their remote sensing data. Many startups, academic institutions, and government agencies are building SmallSat hardware and often don’t have the data expertise or desire to handle the ingest, archive, processing, and distribution of the data. By leveraging open source components that have supported NASA’s Earth Observation System, we’ve developed FilmDrop and SatCAT–our cloud based data pipeline processing engines built to handle both permanent and ephemeral archives of remote sensing data. FilmDrop is an end-to-end system designed to provide ingest, archive, processing, and distribution for more permanent archives. It can act as the foundation for larger systems and expand to handle numerous data products and formats. SatCAT is specially designed for ephemeral archives of convenience that can be spun up and down, costing almost nothing when they’re not running. SatCAT can also be used in situations where a user may need a temporary archive or is experimenting with data and doesn’t need to build a permanent data source for future work.

Beyond the direct use of these stacks by SmallSat operators, we have begun applying them in the disaster response space. SatCAT is capable of provisioning a ruggedized, field deployable archive of convenience hosting critical data for low-to-no communication in-situ environments on an AWS Snowball Edge. FilmDrop can be provisioned onto the Snowball Edge allowing for in-situ collection and distribution of drone and responder captured data. Integrating publicly available data such as Sentinel and Landsat with high resolution, near real-time information and familiar tooling provides improved situational awareness, communication, and coordination.

As more Small and CubeSat hardware becomes commoditized and readily accessible to a larger swath of the engineering community, data systems need to be developed to keep pace. Element 84 is working with government and industry partners to bring these systems to life and improve the data management of an expanding satellite industry.

Identifying the physical aspect of the earth’s surface as well as how the land is exploited by man (i.e. Land cover and Land use) is a challenging problem in environment monitoring and many other subdomains. This can be done through field surveys or analysing satellite images. While carrying out field surveys is more comprehensive and authoritative, it has obvious limits such as being expensive and taking a long time to update (i.e. lack of real-time classifications).
With recent developments in the Space industry and the increased availability of remote sensing data (both optical and radar), satellite images due to their map like format and repetitive coverage, are providing an effective source for land cover classifications exploiting Machine Learning algorithms. There exists already a reasonably rich literature in this context. ([1], [2] and [3])
In our approach to implement a land cover classification, we have chosen the Protected Area Doñana National Park, Spain, which as an Area of Interest has the particularity to consist of a wide variety of land cover types, from urban fabrics to rice fields and forests, we came across the issue that temporal changes in land cover (from bare land to harvesting crops) plus natural dynamism of the region, may introduce an additional dimension to our dataset which is both a constraint and a source of additional information. Our methodology can be applied to other areas of interest and use different sources of ground truth.
The present study aims at answering the following question: shall we involve a time component and treat our data as a time-series by adding an explicit chronological order between observations, i.e. a time dimension? In fact, our problem fits into the characteristics of time-series observations, having both land cover seasonal and cyclic variations/oscillations as well as trend variations.
Starting from the beginning of each observation year, we have setup an incremental classification procedure: every time a new image is available (whether Sentinel-1 or Sentinel-2) over the AOI, a model is trained by taking into account the acquisition date of the last-added image and associating a weight to each of the previously available images. The other consideration regards the optimal weight assignment, which results to be a nonlinear function of time, in the sense that higher weights shall be assigned to most recent entries as well as to those satisfying a cyclic interval (i.e. annual or seasonal). In order to analyse the benefits of our approach we compare the results with the baseline methods where acquisition time is not considered as an input parameter.
The processing chain is designed to be deployed on the Cloud using the Ellip Platform (https://ellip.terradue.com/) and using the Orfeo ToolBox (version 6.2) supervised Machine Learning classifier algorithms, while for training the models, we used the EUROSTAT LUCAS - Land use and land cover survey which provides hundreds of classified samples over the selected AOI.
References:
[1] Jordi Inglada, Arthur Vincent, Marcela Arias and Claire Marais-Sicre. Improved Early Crop Type Identification by Joint Use of High Temporal Resolution SAR and Optical Image Time Series
[2] Amanda Veloso, Stéphane Mermoz, Alexandre Bouvet, Thuy Le Toan, Milena Planells, Jean-François Dejoux, Eric Ceschia. Understanding the temporal behavior of crops using Sentinel-1 and Sentinel-2-like data for agricultural applications
[3] Nicola Clerici, Cesar Augusto Valbuena Calderón & Juan Manuel Posada (2017) Fusion of Sentinel-1A and Sentinel-2A data for land cover mapping: a case study in the lower Magdalena region, Colombia

In this study, we address the problem of change detection monitoring using Sentinel-2 data. The continued development of infrastructures such as buildings and roads has increased the attention toward the soil consumption. The latter occurs often in low density urban areas, at European level, as well as in developing country, which are subject to rapid urbanization processes. A detailed knowledge of such changes across time is highly demanded since it can be useful to understand the urbanization process and make proper urban planning management.
The launch of the Copernicus Sentinel-2 satellites allows us to address several remote sensing problems by use of multi-spectral data with a maximum spatial resolution of 10 m, 13 bands, wide acquisition coverage and short time revisits. The availability of the Sentinel-2 huge amount of open data and the recent development of novel machine learning techniques allowing to improve the accuracy of change detection applications.
In this work we seek to address the problem of identifying artificial changes altering the natural landscapes, using the Sentinel-2 products and Convolutional Neural Networks (CNNs). The latter have become a state-of-art solution for several computer vision problems, such as image classification, object detection, semantic segmentation, etc. CNNs can learn hierarchical image representations from the input data by successively abstracting higher-level features from the previous low-level ones. In the last years, this type of neural networks has been used to solve several tasks in different domains. In particular, CNNs have demonstrated excellent performances on different problems in remote sensing field, such as pixel-based classification, target re- cognition and image pan-sharpening.
CNNs have also shown great capability in the extraction of image’s features, the utiliza-tion of which for processing remotely sensed imagery is relatively recent.
The proposed method is based on a Fully Convolutional Neural Network with a siamese architecture optimized through a local similarity loss function. Siamese neural network [1] is a class of neural network architectures that contain two or more identi-cal subnetworks. More specifically, the term siamese means that the subnetworks have the same configuration with the same parameters and weights. The parameter updating is mirrored across both subnetworks. Each subnetwork takes one of the two input images, which here represent two Sentinel-2 products, acquired at different time, considered to extract the artificial changes over the images.
In general, in change detection applications the images acquired at different time are concatenated at the input layer of the net. The proposed approach using the siamese network in the encoder part as fusion strategy, allows us to combine the features extracted from the two input images in more suitable way. Indeed, the capability of the convolutional layers to extract images features are exploited for the two input images separately.
The study of Daudt et al. [2], has already shown, with a simple Unet architecture [3], that this kind of approach performs better than a typically fully convolutional deep network, since the siamese strategy fusion guarantee a better optimization. To further improve the results, it would be necessary to create a more deeper neural networks to extract more information from the data. Anyway, it is empirical demonstrated [4], how at the network depth increasing, the accuracy gets saturated and then degrades rapidly. In the present work we studied a solution to this issue demonstrating how the use of the recent ResNet architecture [4] in a siamese model ensures a better convergence capability during the training phase.
The change detection algorithm is developed considering that, in a siamese architecture, each subnetwork essentially produces a representation of its input. If the inputs represent the same context, it makes sense to use similar model to process similar inputs. Therefore, we exploited a local similarity loss function to force the siamese encoder to learn how to map, into a dimensional space, each local region of a Sentinel-2 image. If no changes occurred, the siamese encoder extracts similar features, otherwise the features representations of the inputs are locating distant in the dimensional space and are consequently classifying as changes.
To evaluate the proposed method, we used the Onera Satellite Change Detection dataset [5] (OSCD). The experiments demonstrate the effectiveness of the proposed technique with respect to other consolidated methods.

[1] Sumit Chopra, Raia Hadsell, and Yann LeCun. 2005. Learning a Similarity Metric Discriminatively, with Application to Face Verification. In Proceedings of the 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05) - Volume 1 - Volume 01 (CVPR '05), Vol. 1.

[2] Caye Daudt, R.; Le Saux B & Boulch, A. 2018. Fully Convolutional Siamese Networks for Change Detection, IEEE International Conference on Image Processing (ICIP). Athens, Greece.

[3] Olaf Ronneberger, Philipp Fischer & Thomas Brox. U-Net: Convolutional Networks for Biomedical Image Segmentation.Medical Image Computing and Computer-Assisted Intervention (MICCAI), Springer, LNCS, Vol.9351, 234--241, 2015

[4] He, Kaiming, Xiangyu Zhang, Shaoqing Ren and Jian Sun. “Deep Residual Learning for Image Recognition.” 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2016): 770-778.

[5] Urban Change Detection for Multispectral Earth Observation Using Convolutional Neural Networks, R. Caye Daudt, B. Le Saux, A. Boulch, Y. Gousseau. IEEE International Geoscience and Remote Sensing Symposium (IGARSS’2018). Valencia, Spain. July 2018.

Launched in 2013, SARAL/AltiKa is the first satellite radar altimetry mission that operates in the Ka band. The Ka band is known to be more sensitive to the atmospheric water content than the historical Ku band. As a result, rain cells can affect the measurement accuracy.

Margins on the mission's link budget mitigate the impact of rain on the data availability rate, and the geophysical data record delivered to users comes with a rain flag based on the waveform trailing edge variability. However this flag is known to be too severe and unable to separate atmospheric water effects on the waveform shape from other sources of trailing edge variability, in particular those originating from ocean surface characteristics within the radar footprint such as sigma blooms for example.

In recent years, machine learning and artificial intelligence techniques have demonstrated their ability to extract valuable information from high dimensional earth observation data and solve difficult classification problems. However these techniques have rarely been applied to radar altimetry measurements where classifications generally rely on an empirical human-based expertise.

In this study, we experiment with different machine learning techniques for the detection of rain cells in SARAL/AltiKa's radar altimeter measurements and compare their performances. The performance estimation relies on standard statistical metrics (false alarm rate, F-score) but also on dedicated calibration and validation metrics such as cross-overs, which are widely used in the radar altimeter community.

The training/test dataset is based on short delay colocations with a SSMI based rain rate product, and features are extracted from radar altimeter retracking outputs and onboard microwave radiometer measurements.

Results show that all algorithms considered in the study perform well and even simple algorithms are able to do a better job at rain cell detection than the current product flag. This suggests that there is a relatively easy way to separate populations in the data.

While these preliminary results should undergo a complete validation, they show that machine learning techniques should be investigated to address emerging issues in radar altimetry data processing, such as high rate data editing or error detection on future Ka band missions.

Change analysis in combined time series of multilook polarimetric Sentinel 1 (S1) SAR data with Sentinel 2 (S2) optical data is a challenging and promising topic. This is one interest for the French CNES, which is financing this R&T study, for the definition of new products for automatic change detection for flood monitoring compatible with the incoming SWOT(5) mission. We propose here a new mathematical framework based on an omnibus likelihood ratio test statistic (LRT) for the equality of several variance-covariance matrices. We introduce two approaches; i/ the extension to a combined S1+S2 covariance matrix, including normalisations of the covariances for the two sensors; ii/ measurements of T-conorme (Martin, 2005 (6)) to take a posteriori decision considering the two change detection performed using each sensor.


In a second step we compare our results of the monitoring of flood that occurred in the southern part of Tunis in September 2018 with the results of the Copernicus Emergency Monitoring Service(7). We show how both the resolution of 10 to 20 meters and the use of the combined S1+S2 images to monitor the flood event may enhance the monitoring of this disaster.

Finally, we show how the identified “changed patterns” may be characterized (land, water, vegetation..) using the multi-variate (spectral + radar) distribution to provide automatically, in the future, a rolling archive of comprehensive maps of short time-scale (from one week to one month) of changes of the land coverage, base onto the incoming SWOT and S2 data.

 
Figure 1: Flood monitoring (in red) using 3 images S1A GRD VV&VH 20m resolution, & 3 images S2A 20m resolution (detection threshold at 99%), of the flood event in Tunisia in 2018.

(5) SWOT: Surface Water and Ocean Topography
(6) Martin, A. (2005). La fusion d’informations. Polycopié de cours ENSIETA-Réf, 1484, 117.
(7) COPERNICUS EMERGENCY MONITORING: https://emergency.copernicus.eu/

Crop production is a complex of different agriculture systems and depends on various factors i.e. Irrigation, Fertilization, seed, Insects/pests measures. Robust and correct real time decisions are necessary at field level for optimum crop output. In preview of smart agriculture and precision agriculture systems, near real time access to spatial and synoptic information and crops monitoring is a prerequisite for quick and reliable Assisted Decision Support System for farmers.
In most of the developed countries, the boundaries level data is available either in the form of digital land records or through aerial survey campaigns. In Pakistan, land registry records are still not fully digitized and a recent campaign for digitization of the records is slow and lacks the accuracy. Moreover, in terms of crop health monitoring and yield modeling a generalized boundary of crop type is more important. The problem encountered in the goal of agriculture field boundary markings are
• Pakistan, being a large country, its agriculture is spread over a vast area.
• The agriculture itself is dynamic and a range of crops are grown with seasonal, yearly and long-term changes in their locations and extents.
• There is no stringent enforcement of policies regarding dates of sowing or crops allowed in a region.
• The individual fields are often mixed together and cultivated/harvested in unison. The boundaries are often complicated to identify due to outcroppings, tree lines or overlapping features at field boundaries.
• There has not been a high-resolution aerial or lidar survey carried out for the entire country.
• The official records are in the form of mouza maps and lack the digital field level geographic information.
• Generally farmers lack the up-to-date scientific knowledge-base for field challenges for better crop production.

Availability of high-resolution satellite imagery makes it now possible to carry out huge surveys within a very short span of time. These images are used to discriminate crops, estimate crops acreage and crop health monitoring. The extended scope of high-resolution satellite imagery enables to map and monitor farmer land parcels, developing farmer's geospatial database and monitoring of crops growth profiles at field scale.

In the study, using stacks of multi-temporal satellite imagery, field boundaries are identified based on changes in the cropping patterns across seasons and our project proposes to apply AI to test deep learning methods for automatic delineation of agriculture field boundaries based on a temporal analysis of satellite images in a heterogeneous cropping country of Pakistan. Thus providing a non-existent layer of boundaries for field level analysis, monitoring, evaluation of crops in Pakistan.

The above is an abstract of the ongoing masters dissertation at UCL which will be completed by 30th Aug, 2019 to be ready for poster presentation at the conference.

Searching anomalies in training and validation data when solving the problem of crops or land cover classification on the basis of satellite data is very important and thus a non-trivial task. The difficulty lies in the fact that it is possible to detect anomalies associated with incorrect marking of fields or confused classes only when applying a time series of satellite images. In this work, time series of satellite images Sentinel-1A were used for experiments. When training artificial intelligence models for agricultural crops classification, incorrect samples can very negatively affect classification result, and incorrect samples in the test sample provide an inadequate model quality assessment. To solve detecting anomalies or errors in the in-situ samples problem , 5 of the most popular methods of anomalies detection were used: SOM ANN, unsupervised SVM, isolation forest, robust covariance and local outlier factor. The best method was chosen according to the criteria: speed of learning-detection, accuracy of detection, resistance to noise and anomalies in the sample. After this, the classical method of threshold screening of the sample based on the quantitative characteristics of the anomalous pixels in the sample was replaced by a new technique. The new technique consists in detection anomalies in morphological features of anomaly map based on the learned convolution neural network. This model take as input anomaly mask that contain 3 possible values: -1 if pixel classified as anomaly, 0 if no data for pixel and 1 if pixel is not anomaly. This model can be trained using any in-situ data dataset and training samples of anomaly and normal features can be generated with adding uniformly distributed noise or by providing anomaly detection classification by model trained for other crop classes. Main advantages of this approach are: good separation of incorrect marked features and features with high noise level on radar images because use of convolution neural network can analyze noise form and distribution, approach contain use of unsupervised anomaly detection method and supervised method for which training data can be easily generated, approach have high speed in use and high accuracy on real SAR data time series tests.

The proliferation of deep learning approaches, coupled with the recent progress on unmanned aerial vehicles (UAVs), have allowed the development of new methods that combine unprecedented spatial resolution as well as exceptional accuracy. Regardless of the usability of the introduced remote sensing modality, the gain in resolution (in comparison to satellite data) is inherently linked with a loss in scalability. As a result, the community is facing the challenge of developing a paradigm that would include both UAVs and satellites in an optimised setup.

Specifically in agricultural applications, such parallel sources of knowledge feeding each other could mimic the farmer-agronomist relationship at a larger scale. The farmer assesses the whole area of a field and flag specific regions of interest, before an agronomist provide a locally-informed opinion that if is generically applied would optimise the outcome. Equivalently, satellite imagery typically has the resolution and extend to provide a generic understanding of an area but is not always fitted to base real-world decisions and applications, while UAV imagery can focus in a region of interest to produce high-accuracy top-quality insight, even though the deployment of UAVs over a large area may be highly impractical.

Following the above correspondence, we have developed a number of solutions in which the UAV and satellite imagery processing are working in tandem. In this paradigm, from a technical point of view, the UAV is producing an in-depth analysis of the preliminary conclusions that are derived from a (coarser resolution) satellite imagery. This coupling is also interesting from an AI point of view, since it links, in a multi-scale approach,classification and regression, detection and identification, segmentation and instantiation, etc.

Examples of the satellite/UAV synergy include: (a) the detection of gaps in vegetation from high-resolution satellite imagery before an UAV sensor analyses the soil characteristics (b) the satellite-based detection of weeds in pre-emergence fields, leading to a UAV survey to provide a more accurate estimation of the weed extent and possibly the weed types present in the field (c) the classification of soil types in satellite imagery of fields in pre-emergence before UAV images are used to monitor growth during the season.

Applications of this multi-scale and multi-platform paradigm, along with the machine learning approaches associated to each of the stages, are going to be presented in this talk. The technical content is going to be further explored, along with potential uses of the same technology in similar remote sensing and earth observation domains. Finally, the directions of future development as well as inherent limitations are going to be presented in detail.

The major topic of this abstract is the design and implementation of a Cloud-Based Land Cover Monitoring Front Office application for execution on the C-DIAS Mundi Web Services. The overall objectives of this application are to enable Users in developing countries to process and analyze Sentinel-2 data, and to stimulate the uptake of Copernicus data in the framework of Earth Observation (EO) based forest monitoring in tropical countries. The target group are Stakeholders in developing countries implementing the United Nations Framework Convention on Climate Change Reducing Emissions from Deforestation and Degradation (UNFCCC REDD+) policy, as well other Users which need to have capacity to operate National Forest Monitoring Systems (NMFS). The application provides cloud-based processing of dense time series of Sentinel-2 data for Land Cover monitoring in a realistic IT environment for Users in developing countries.
The typical use case is the creation of a Land Cover Map by the User in the framework of a REDD+ Monitoring, Verification and Reporting (MRV) system in a developing country. The User is provided with a complete mapping workflow, which consists Project Definition, Pre-processing, Land Cover Classification and Validation.
The workflow is implemented in a web-based, Front End which allows the use of the Front-Office without the need to install any software on the computer. The User starts his mapping project by defining the AOI and time frame as well as the maximum cloud cover allowed for the source imagery. The Pre-processing of all images fulfilling the search criteria includes atmospheric correction, cloud-masking, the assembly of a time series image stack and calculation of time features (statistical temporal metrics of the time series). Time features are powerful input features for various classification or regression tasks and are particularly useful for the classification of land cover categories with a distinct phenology such as deciduous forest types or the Dry Forests in southern Africa (Valero et al., 2016). The land cover classification into the six standard IPCC Land Use classes (Forest Land, Cropland, Grassland, Wetlands, Settlements, Other Land) is done by the supervised Machine Learning algorithm LightGBM (Ke et al., 2017). The user selects the training data in the Front Office or loads them into the application from an external dataset. Classification results can be viewed instantly, and the post-processing in terms of a User-defined MMU filtering is applied. The land cover map is validated based on an external reference dataset, which fulfils the requirements of the User’s REDD+ MRV system on independence, sampling design and response design. The analysis design is implemented fully in the Cloud-Based Forest Monitoring Front Office, producing an accuracy matrix which is used to assess overall accuracy, class-wise producer’s and user’s accuracies, as well as estimating the area uncertainty based on the accuracy results and the area proportions of the different land cover categories (Olofsson et al. 2013).
The Front Office application relies on a fully cloud-based processing Back End, which runs on an OpenStack Infrastructure. The Back End is currently rolled out on the Open Telecom Cloud and contains all the processing steps, which the User operates from the web-based Front End. The Back End is designed in a micro service architecture, which contains the different processors in modular and portable units using container (Docker Inc.) technology. This technology allows the developed application to be light-weight following the concept of being modular and portable. The processors within the Back End cover the full processing chain including image download, preprocessing, thematic processing, post processing, validation and output. Every service is completely scalable by using a Kubernetes Cluster deriving an automated orchestration and deployment of the needed services on the cloud infrastructure. The cloud-ready architecture utilizes Analysis Ready Data (ARD) in a scalable, modular preprocessing pipeline and the final data is indexed and ingested in a set of Data Cubes with a Web Coverage Service interface to the S3 Object Storage. Multiple requests on the Data Cubes enable the upscaling of the Time Feature Calculation Service and the Classification Services.
Besides the typical use case, the application offers the opportunity for an iterative data science approach for a thematic exploitation of the ARD. An expert user can run a classification model building service based on training samples which is iteratively validated to a set of ground-truth-points. After each model building and validation, the user gets the feedback of the preliminary accuracy (via confusion matrix) of the model and he gets the opportunity of changing the training samples, the used time features and the Machine Learning Algorithm (LightGBM, Random Forest, Support Vector Machine, etc.). If the user reaches an acceptable level of optimization, the Back End transfers the model to a Batch Processing Service which enables the creation very large coverage Land Cover Maps limited only by the size of the used cloud infrastructure hardware.

References
Ke, G., Meng, Q., Finley, T., Wang, T., Chen, W., Ma, W., Ye, Q., Liu, T. (2017). LightGBM: A Highly Efficient Gradient Boosting Decision Tree. Advances in Neural Information Processing Systems 30 (NIPS 2017), pp. 3149-3157. https://papers.nips.cc/paper/6907-lightgbm-a-highly-efficient-gradient-boosting-decision-tree.pdf
Olofsson, P.; Foody, G.M.; Stehman, S.V.; Woodcock, C.E. (2013). Making better use of accuracy data in land change studies: Estimating accuracy and area and quantifying uncertainty using stratified estimation. Remote Sens. Environ. 129, 122–131.
Valero, S.; Morin, D.; Inglada, J.; Sepulcre, G.; Arias, M.; Hagolle, O.; Dedieu, G.; Bontemps, S.; Defourny, P.; Koetz, B. (2016) Production of a Dynamic Cropland Mask by Processing Remote Sensing Image Series at High Temporal and Spatial Resolutions. Remote Sens. 2016, 8, 55.

Disruptive small satellite technology has given an opportunity to various aerospace entities to join the space activities and opened a new breadth of applications. The biggest portion of the CubeSat missions nowadays is dedicated to Earth analysis - collection of imagery in different spectrums, information about vehicles mobility or weather conditions detection using radio occultation to name a few. These technologies have proven extremely useful for humanitarian and development work, from monitoring disaster damage to poverty estimation. Nevertheless, access to information and connectivity is one key challenge separating nowadays the most vulnerable from access to opportunity and choice and monitoring connectivity remains today outside the capacity of satellite-based observation. The current work aims to analyze the approach of using CubeSat technology as a tool for space-based radio frequency (RF) mapping. The information about RF signal intensity distribution can serve as a valuable source for analysis of the economic situation in different regions, disaster monitoring, evaluation of connectivity in remote areas. Focusing on these applications, the UNICEF research group in collaboration with Small Satellite Project Group (SSPG) of Space Generation Advisory Council has examined the small satellite technology readiness level.
With the present work, UNICEF and SSPG are addressing the question of the feasibility of the CubeSat technology utilization for the radio frequency geospatial detection. In other words, the current work evaluates available technologies, which can fit in the CubeSat form-factor, to detect and localize radio signals from space. The technology analysis is based on an overview of existing antennas and receivers, their capabilities and limitations. Once possible receiver configuration is defined, the paper provides a review on different CubeSat architectures and their possibilities from power, mass, size, and attitude control perspectives to carry the corresponding payload. Besides that, the analysis of possible satellite data use and fusion with other data sources is conducted in order to map the opportunities of the proposed approach. Finally, the feasibility of the proposed concept is supported by the case study, which is based on gathering data about connectivity in remote areas and understanding how disconnected these areas are.

The aim of this study is to present MARRE project, a research project which will focus in the Greek Seas in order to provide new local EO datasets to monitor the marine environment. The project will utilize the Sentinel 2 and Sentinel 3 remote sensing data provided by Copernicus as well as, in situ measurements from existing monitoring networks, like the HCMR Poseidon System, and new research cruises that will be carried out in the Aegean Sea for the specific needs of MARRE. Using these data, new local algorithms will be developed in order to provide validated and accurate information about the status of the marine environment in the Aegean Sea.

The products that will be generated and operationally monitored and assessed by MARRE will include chlorophyll concentration, Posidonia Oceanica meadows mapping, Potential Fishing Zones, and marine environmental parameters like salinity and turbidity. All these parameters are very important for a wide range of scientific areas including marine ecology, fisheries management and physicochemical monitoring of the Greek seas.

The methodology for estimating Chlorophyll concentration, will focus on the combined use of satellite data and the in situ measurements and will evaluate and improve existing empirical algorithms as MEDOC4 and MEDOC4me. As a result, the derived products will be adapted to local conditions and the specific features of Greek seas. Importance will also be given to the monitoring of the marine biodiversity by the mapping and the monitoring of the Posidonia meadows. For a high-resolution analysis, the Sentinel 2 data will be used. Overall, the satellite processing will be based on 4 pillars: a. Automatic data acquisition, b. Satellite data preprocessing, c. Object Based Image Analysis (OBIA), d. Accuracy assessment.

The potential fishing zones (PFZ) of the Greek seas will be extracted using the remotely sensed datasets of Sea Surface Temperature and Chlorophyll-a which are provided by Marine Copernicus. Those parameters will allow us to identify the corresponding oceanic fronts which are used to estimate the spatial distribution of pelagic fishes. Finally, a principal component analysis and a simple multi-criteria analysis will be applied to produce the daily Potential Fishing Zone maps based on the near real time satellite data.

All the products of the MARRE system will be publicly available through a web portal. The operational use of the system will be supported by an IT platform which will consists of: a. A central database, which will store all the data produced during the project. b. Web services, to access data and information derived from project. Since most of the project data are spatial data, the services that will be provided for spatial data management will be based on open standards such as those defined by the OGC: WMS, WFS, WCS, WPS and the INSPIRE directive. c. Applications, for the dissemination of the project’s results, which will be developed using open source software like GET SDI PORTAL, for the geoportal, and CKAN, for the open data portal.

More information about the progress of the project can be found on the official website at http://marre.gr/

The project is financed by the Operational Programme Competitiveness, Entrepreneurship and Innovation (EPAnEK, ESPA 2014-2020) of the Greek Ministry of Economy & Development, and co-funded by the EU (European Regional Development Fund) and by national funds.

The advances in remote sensing technologies accompanied by the extraordinary developments in hardware and software capabilities has leveraged the earth data collection, distribution, and analysis for a better capitalization of the latest AI techniques. The massive amount of high resolution remotely sensed data acquired with a temporal frequency from the entire earth is apparently a rich source of information about the earth states and its continuous changes. However, the full exploitation of data and the automatic extraction of such information is still a challenge.

Land use and land cover are the core knowledge about the earth surface for any earth monitoring, development, controlling and planning task. On the other hand, collection of such information on a temporal basis from extensive ground assessments is very expensive and challenging. Therefore, the focus of most research works aiming at employing AI on remotely sensed data is to reach an automatic land cover/use classification with the highest accuracy.

The development of brain-inspired deep learning techniques during the past decade has shifted all the expectations from the AI into a new level. Being beyond the human performance, deep learning techniques have proven the incredible capability of machine in understanding and analysis of complex big volume data. Multispectral images acquired by the optical remote sensors are intrinsically complex and rarely used by machine learning techniques in research. The format of such data passes all the requirements to employ deep learning techniques, with possibly some customization and data prepossessing. Moreover, the rapid advances in optical sensors and expansion of spectral channels with rich information about the earth surface, is an absolute evidence of how important the exploration of such data is.

This review work compares and evaluates the available multispectral datasets developed for machine leaning purposes, and discusses the techniques and approaches employed in the literature that use such datasets for land use/cover classification. At the end, the results obtained by different approaches will be compared, and the open ground for improvements and future works will be introduced.

Physical radiative transfer models (RTMs) of leaf and canopies with sufficient realism enable the quantitative retrieval of biophysical variables from imaging spectroscopy through numerical inversion. They can serve as tools for precision farming application, e.g. for mapping the health status of the crop. However, advanced RTMs are computationally intensive, which hampers practical applicability of inversion schemes against remote sensing images. To bypass the computational load of such RTMs, it has been proposed to approximate these models by means of machine learning, i.e. emulation. The principle of emulation is, once an accurate emulator has been developed, it can approximate the inputs-outputs relationships of the original RTM - and this at a tiny fraction of the original speed - and be readily applied in tedious processing applications. This technique opens a variety of new opportunities in using complex RTM-like emulators in Earth observation applications. For instance, with respect to vegetation properties mapping, the hypothesis is that when the original RTM is replaced by its emulated counterpart, then numerical inversion can go very fast and can become again a truly attractive and generally applicable retrieval method. Moreover, when obtaining an accurate emulator of a computationally expensive RTM with high realism, then in principle the inversion scheme does not only run fast, but it can also infer vegetation properties more accurately.
In this contribution we will give an overview of latest advances in the usage of emulators for approximating leaf, canopy and atmosphere RTMs. The benefits of emulation will be illustrated by means of global sensitivity analysis (GSA) studies for leaf-canopy and atmosphere RTMs. Traditional GSA has considered expensive brute-force approaches based on permutation or Sobol indices. The use of Gaussian process (GP) emulation offers more efficient and amenable alternatives rooted on principled Bayesian inference. We plan to evaluate and compare several methods in terms of accuracy and speed, e.g. Monte Carlo approximation, Automatic Relevance Determination, GP sensitivity analysis and Bayesian GP-LVM. Comparison of GSA results of original RTMs vs emulated counterparts will demonstrate the physical consistency of the emulation method.
Finally, an overview of other emulation applications will be given, including retrieval applications, synthetic scene generation and atmospheric correction strategies. This presentation will close with guidelines how emulation can be easily applied into tedious image processing applications.

Palm oil is nowadays the most common vegetable oil used globally. As the demand grows, the expansion of Palm Oil plantations in the tropical regions will remain one of the main drivers of deforestation, having a considerable impact on biodiversity and disrupt the carbon cycle. Although the current hot spot of production is located in South-East Asia, future expansion is expected for African and South American tropical regions. The monitoring of its current extent and future expansion is crucial for validating the recent efforts towards a sustainable production on a global scale.
One of the main difficulties in mapping palm oil on a global scale is the regular cloud coverage over tropical regions and its spectral similarity to natural forests in the optical domain, which both hinders its applicability on a global scale. In this study we overcome both limitations by using Sentinel-1 C-band SAR, ALOS-2 Palsar-2 L-Band SAR and SRTM elevation data. Traditionally L-Band SAR data is considered to be most efficient in mapping forested areas due to its long-wavelength signal that is capable of penetrating the canopy. In contrast, the strength of our method capitalises mainly on the dual-polarised VV/VH C-Band SAR data from Sentinel-1, which is particularly well suited to distinguish palm trees from other forest types.
In addition, the 12-day revisit cycle of Sentinel-1 allows for the creation of dense time-series. The temporal behaviour of the backscatter is then captured by using the timescan method, which depicts the usage of descriptive statistical parameters for every pixel in the full time-series. This helps in reducing the influence of environmental conditions and adds further strength in discriminating palm oil plantations. Simultaneously, the amount of data used for classification is reduced when compared to the full time-series and the number of input channels into the subsequent machine learning algorithm is standardised.
The drawback of this method is the huge amount of data pre-processing. To overcome this obstacle, we utilise the online platform Google Earth Engine that allows for the planetary-scale application of Earth Observation data in reasonable amounts of time.
About 6000 collected ground truth samples are then used to train and tune a Random Forest classifier, which is applied to the full dataset. We further elaborate an independent validation of the map using very-high-resolution data.

Policymakers rely on Land Use and Land Cover (LULC) maps for evaluation and planning. They use these maps to plan agriculture policy, improve housing resilience (to earthquakes or other natural disasters), and understand how to grow commerce in small communities. A number of institutions have created global land use maps from historic satellite imagery. However, these maps can be outdated and are often inaccurate, particularly in their representation of developing countries.

We found that Global Land Use and Land Cover maps were wholly inadequate for this analysis and little high-quality data existed to reproduce the process. To overcome this, our Machine Learning and Data Teams partnered to create a new methodology for machine learning in data-sparse environments and used it to create an accurate and timely settlement map for Ethiopia. At the core of our approach is a two-step machine learning process to map urban settlement and track changes. To overcome a few challenges we were facing in Ethiopia, we created an ML pipeline:
combines the strengths of a small group of human mappers with machine learning, and
combines the strengths of both unsupervised and supervised ML methods.
Our final prediction of the urban area still had a few false positives. Bare soil is sometimes misclassified as built-up/urban in the desert and steppe terrain in the Western and Southwest parts of Ethiopia. This is in part because the urban settlements in these parts of Ethiopia are rarely paved, meaning that there is a high composition of bare soil in the pixel. These subtle distinctions are still best made by humans. We undertook the final analysis of the candidate areas, using the image, surrounding context, and supplemental data sources like OpenStreetMap. The post-processing has two purposes in our workflow:
Excluded bare soil areas in Western Ethiopia where desert and steppe terrain is dominant;
Removed the detected individual or small villages that aren’t the urban settlements.

Our urban settlement and the changes mapping represent a detailed, granular, and accurate map of urban dynamics than is currently available from OSM or other sources. This a scalable and repeatable approach. By automating this approach, we believe we can create vastly more accurate results for land cover maps in most countries.

Accurate data about school locations is critical to provide quality education and promote lifelong learning. The UN’s concerted effort to address global challenges through the Sustainable Development Goals (SDGs) places both inclusive and equitable access to education (SDG Goal 4) and the reduction of poverty worldwide (SDG Goal 1) as two of their top priorities. However, in many countries, educational facilities and school location records are often inaccurate, incomplete or non-existent.

An accurate, comprehensive map of schools - where no school is left behind - is necessary to measure and improve the quality of learning. Such a map, in combination with connectivity data collected by UNICEF’s Project Connect initiative, can be used to reduce the digital divide in education and improve access to information, digital goods and opportunities for entire communities. In addition, understanding the location of schools can help governments and international organizations gain critical insights around the needs of vulnerable populations, and better prepare and respond to exogenous shocks such as disease outbreaks or natural disaster.

Despite their varied structure, many schools have identifiable overhead signatures that might make it possible to detect them with modern deep learning techniques applied to high-resolution satellite imagery. This paper presents the results of a Convolution Neural Networks school classifier for high-resolution satellite imagery. This classifier is based on the pre-trained models, Xception and MobileNetV2, from ImageNet, with a validated and cleaned dataset of schools in Colombia.

We selected the best-performed model with 0.94 area under the ROC curve and 9% of the false positive rate from our nearly 200 training iterations. A large-scale model inference process across Colombia and the Eastern Caribbean islands was implemented over 52 million DigitalGlobe Vivid imagery tiles. A group of five expert mappers was able to validate 73,000 predicted school tiles from our machine learning inference within eight working days. We added about 11,000 schools to the map in Colombia and the Caribbean islands, and around 7,000 of them were unmapped schools.

Our study showed that current deep learning and inexpensive cloud computing can assist humans to detect schools at scale in a rapid, rigorous manner. This provides the first object-based detection model for schools. A complete and accurate school facility map can further reduce the digital divide in education and improve children’s access to information, digital goods, and opportunities, and make the best use of limited educational resources.