Optical Land Observation from Space has meanwhile evolved from traditional LEO-Satellites like Landsat (USA) or SPOT (FR) to high performing next generation satellites with a continuous global mapping. The acquired visual and near infrared spectral data are used for a wide range of applications e.g. for geo-physical as well as for bio-physical products. An impressive example is the digital farming, which uses the “chlorophyll” images from satellites to assess the vegetation growth status, by which the fertilization of each GPS-spot of a field is optimized. The European Union built this next generation of satellites called "Sentinel". The key objective of the Sentinel-2 mission is to image continuously and systematically world-wide all land areas between -56° (Patagonia) and +84° (Greenland) latitude with a twin satellite constellation. It is a major breakthrough in the area of optical land observation, achieved by a very large Multi-Spectral Instrument (MSI), since it combines for the first time:
• continuous and systematic acquisition of all land surfaces and coastal waters world-wide
• high spatial resolution (10m / 20m / 60m) with very large swath of 290km
• large number of spectral bands (13 bands in VNIR-SWIR domain)
• High re-visit frequency of 5 days with two Sentinel-2 satellites
• High precision on-board attitude and position knowledge (CE95 = 3.2 m)
• Accurate image geo-location (16m without Ground Reference Points)

Such a huge amount of data produced on-board challenges the design of the on-board data handling in several performance aspects. Each Sentinel-2 satellite collects per day 1.5 Terra Byte. This “Big Data” needs to be stored temporally on-board and down linked to Earth for further on ground processing to generate a variety of bio-physical and geo-physical products.

This presentation gives an overview on the Sentinel-2 satellite data handling design challenges and their solution.
• Constellation:
Flying a constellation of 2 satellites in the same orbital plane and spaced at 180°.
• High data acquisition rate:
Acquire 13 of spectral bands in parallel and process the image data at real-time using on-board a state-of-the-art wavelet compression. This results in a more less continuous data stream of 110 GByte per Orbit.
• Large and fast on-board storage:
All acquired data need to be stored on-board until the next ground station is available and the data can be down linked. This is done by (currently) the largest Mass Memory and Formatting Unit (MMFU) flying in space. This MMFU can handle high speed recording in parallel to high speed formatting and downlinking. This MMFU is based on flash technology and provides an on-board storage capacity of 6 Terra-bit (BoL).
• High speed downlink:
Two parallel X-band down links at 2*260 Mbps = 520 Mbps are used to get the images down to ground. The visibility of ground stations is in LEO orbits limited to max. 15 minutes. This Sentinel-2 down link capacity is augmented by an additional Laser Communication Terminal (LCT/OCP) allowing longer periods (about 30 minutes) of data transmission by LASER first to a geostationary relay satellite e.g. EDRS and then down to Earth.
• Autonomy:
The Sentinel-2 satellites are designed for high autonomy allowing for pre-programming of the operational schedule for 15 days in advance without interference from ground. This is done by in total 15000 commands split into Time-Tagged commands as well as by Orbit-Position-Tagged commands.
• Availability:
The on board autonomous Fault Detection Isolation and Recovery (FDIR) scheme and the related redundancy concept guarantee for each Sentinel-2 satellite an in-orbit availability of the instrument data greater than 97%.

1. SATELLITE OPERATIONS
1.1 Orbit
Sentinel-2 is a polar orbiting satellite in a sun-synchronous, low-Earth orbit with a mean altitude of 786 km. The global acquisition of the land surface of almost the entire earth is achieved within 5 (with 2 satellites) or 10 days (with 1 satellite), respectively. The stable orbit conditions are ensured by the selection of a so called frozen orbit with a fixed repeat cycle of 10 days and a cycle length of 143 orbits. This results in an orbit period of about 100 minutes (corresponding to 14+3/10 revolutions per day) and a ground track velocity of 6,6 km per second.
To achieve an optimum sun illumination for the image acquisition a mean local time at descending node has been selected for 10h30, i.e. each orbit the satellite crosses the equator from North to South at 10h30 local time. Imaging at higher Northern latitudes approaches noon time thus providing for optimum illumination conditions.

1.2 Space to Ground Interfaces
The Sentinel-2 Ground System encompasses a Flight Operations Segment (FOS) operated by ESA/ESOC in Darmstadt (Germany) and the Payload Data Ground Segment (PDGS) managed by ESA/ESRIN from Frascati (Italy).
The telecommand uplink and the downlink of high priority, real-time and playback satellite telemetry is performed by S-band TM/TC Stations. The mission image data and the related satellite ancillary data are transmitted via X-band to the Core Ground Stations. In addition all satellite housekeeping data are also available via X-band. Furthermore, Local User Stations can be supported.
Data transmitted via X-band can also be transmitted via an optical link between the Sentinel-2 OCP and a Laser Communication Terminal on-board a geostationary relay satellite for subsequent transmission to ground in Ka-band.
In routine phase, two S-Band station passes from Kiruna will be scheduled every day for real time TM downlink and tracking/ranging activities as well as all necessary TC uplink. The recorded satellite HK TM will be also downlinked once per orbit in X-band to a Core Ground Station (under PDGS control) and forwarded to the FOS.
The X-band Core Ground Station network comprises Svalbard, Prudhoe Bay (Alaska), Matera and Maspalomas. Alice Springs, Sioux Falls and Cuiaba have been considered as Local user Stations for mission analysis.

1.3 Data Acquisition, Downlink and Latency
The time between the Sentinal-2 mission data acquisition and downlink to the Core Ground Stations is called data latency. This time is mainly driven by the acquired data volume, the available Core Ground Stations, and the assigned downlink priorities.
Due to the required image quality (i.e. compression) the data volume acquired per orbit repeat cycle is deterministic and varies only over season. Due to the illumination conditions and global land surface distribution, a higher data volume is acquired during summer at the northern hemisphere. Since also the Core Ground Station network is fixed, the data latency is controlled by the assigned downlink priorities, i.e:
 Real-Time (RT) Data are acquired and immediately downlinked in the visibility area of a ground station.
 Near-Real-Time (NRT) Data are related to mission products that will be sent to the user within 3 hours after acquisition. NRT data have priority over Nominal data and are downlinked in a FIFO order
 Nominal Data – neither RT or NRT data. They are downlinked in a FIFO order as well and associated products shall reach users no later than 24 hours after data acquisition.

In case of nominal acquisition, the latencies are below 4 hours for almost all areas. Areas with the lowest data latency mainly correspond to Canada, the North-Eastern part of Russia and New Zealand, with latencies below 1 hour. Most areas over Europe and Russia show latencies below 2 hours.
The latencies over dedicated areas can be reduced by acquisition of Near-Real-Time data, e.g. over Europe. This results in latency figures below 1 hour for Europe but increases the data latency for other areas.

Acknowledgement:
The Sentinel-2 satellites are build by Airbus as prime contractor under an ESA contract with the financial contribution of the European Union that are gratefully acknowledged.

The increased availability of a large amount of spatio-temporal images is driving the need for adequate tools to process, analyse and extract actionable information from the data. Machine learning (ML) and, in particular, deep learning methods have become the state-of-the-art in many vision, language, and signal processing tasks, due to their ability to extract patterns from complex high-dimensional input data. Classical ML methods, such as random forest and support vector machines [1, 2] have been used in many EO applications to analyse temporal series of remote sensing images. On the other hand, convolutional neural networks have been employed to analyse the spatial correlations between neighbouring observations, however mainly in single temporal scene applications [3]. In this paper, similarly to the work of [4], we investigate a deep learning architecture capable of simultaneously analysing the spatio-temporal relationships of satellite image series. In this work we demonstrate its application on the land cover classification of the Republic of Slovenia, using annual Sentinel-2 satellite images for the year 2017.

METHOD
We present a Temporal Fully-Convolutional Network (TFCN) that extends the FCN architecture (a.k.a. U-Net [5]), which is currently the state-of-the-art in single scene semantic segmentation. Similar to [4], the architecture exploits spatio-temporal correlations to maximise the classification score, with the additional benefit of representing spatial relationships at different scales due to the encoding-decoding U-Net structure.
The input tensor of the time series has a shape of [T, H, W, D], where T is the number of time frames, and H, W and D are the height, width and number of bands (i.e. depth) of the input image tensor. The algorithm performs a 3D convolution in the spatial as well as the temporal dimension. By default, max-pooling is performed in the spatial domain only. As the target land cover labels are not-time dependent (i.e. one label per pixel is available for the entire time-series), 1D convolutions along the temporal dimension are performed in the decoding path of the architecture to linearly combine and reduce the temporal features. The schematic of the model architecture is shown in Figure 1. The output of the network results in a 2D label map which is compared to the ground-truth labels. The architecture performed three encoding and decoding steps (i.e. three max-pooling and three deconvolution layers), with a bank of two convolution layers at each encoding and decoding scale. The number of convolutional features was set to 16 at the original scale, with a factor of 2 applied at each deeper level, with a kernel width of 3. The Adam optimiser (learning rate 0.001) was employed. The TFCN model is implemented in TensorFlow. The code to reproduce the entire pipeline will be released at the time of the conference.

EXPERIMENTS
To evaluate the TFCN, we present an end-to-end workflow to generate a land cover map for Slovenia. We do this by using eo-learn, a framework to handle multi-temporal and multi- source satellite data, both in raster and vector format. The framework allows splitting the area-of-interest (AOI) into smaller patches, that can be processed with limited computational resources and allows automation of the processing pipeline. The framework consists of open- source Python packages and is designed to facilitate the prototyping and building of Earth Observation applications. We use this framework to process the data and train the ML models to predict classes of land use and land cover.
The framework was used to generate a land cover map of the Republic of Slovenia for the year 2017. The inputs to the framework are a shape-file defining the geometry of the AOI, the Sentinel-2 L1C images for the entire year, and a set of training labels. Avoiding to download and process entire tile products (e.g. Sentinel-2 granules) provides flexibility and facilitates an automation of the processing pipelines. A pipeline in eo-learn is defined as a connected acyclic graph of well-specified tasks to be performed on the data. eo-learn supports parallelisation of operations, such that the same workflow (e.g. data preparation for land cover classification) can be run in parallel for the smaller patches constituting the AOI. Logging and reporting allow to monitor and debug the execution of the processing pipeline. The automated pipeline for predicting the land-cover labels follows the following setup:
The boundary of the AOI was split into 25 × 17 equal parts, which resulted in about 300 patches of about 103 × 103 px2 at a 10 m resolution. The splitting choice depends on the amount of available resources. The output of this step is a list of bounding boxes, covering the AOI.
2. The bounding boxes were used to create patches, or containers, where the information for the corresponding area is stored. Given the time interval of interest (e.g. from 2017-01- 01 to 2017-12-31), eo-learn used the sentienlhub-py package to download the desired band data. The s2-cloudless cloud detector was subsequently used to obtain the cloud probabilities and cloud masks. The following Sentinel-2 features were input to the model (bands B02, B03, B04, B08, B11, B12, their euclidean distance NORM, and indices NDVI and NDWI).
3. Ground truth information was obtained in a vector format, which was then rasterised over the given bounding box and added to the patch. The ground-truth is openly available through the Slovenian Ministry of Agriculture, Forestry and Food, and is mapped to the following classes: cultivated land, forest, grassland, shrubland, water, wetlands, artificial surface, and bareland.
4. A random spatial sampling of patches was performed to select time-series of pixels which are used for ML model training. The random sampling is uniform throughout the patch and independent on the labels.
5. Due to the non-constant acquisition dates of the satellites and the irregular weather con-
ditions, missing data is common in EO. To deal with these, the mask of valid pixels in the time series is used to interpolate the values and fill the gaps due to the missing data. Linear interpolation was used, after which the time-series was sampled at uniform temporal dates (with a 16-day interval) to unify the dates among all the patches. This normalisation generalises the model to different AOI and time intervals.
6. The data was split into the train, cross-validation and test sets (60:20:20) and provided to the deep learning model, which was trained on an Amazon Web Services (AWS) instance.

RESULTS
The trained model was used to predict the labels on the test sample and the obtained results were then validated against the ground-truth. An overall accuracy of 84.4% and a weighted F1 score of 85.4% were achieved. In general, poor prediction was achieved for under-represented classes such as wetlands and shrubland. These results represent preliminary work on a prototype architecture which was not optimised for the task at hand. Despite this, results in line with previously reported work (e.g. Inglada et al. [1]) were achieved. Optimisation of the architecture (e.g. number of features, depth of the network, number of convolutions) and of the hyper-parameters (e.g. learning rate, number of epochs, class weighting) is required to fully assess the potential of TFCN.

DISCUSSIONS AND CONCLUSIONS
In this paper we have presented an approach to a land-use-land-cover classification using deep learning. While the architecture of the presented TFCN model was not fully optimised, it shows promising results, with the overall accuracy and weighted F1 score reaching levels of about 85%. A thorough and extensive evaluation comparing the TFCN with state-of-the-art methods will be presented at the symposium. Insights into good practices for training, evaluating, and deploying machine learning and deep learning models for EO applications will also be discussed. Source code to reproduce the framework and the models will be made available at the time of presentation to foster research and development in the exploitation of spatio-temporal data in EO applications.

REFERENCES
[1] Inglada J., Vincent A., Arias M., Tardy B., Morin D., and Rodes I., “Operational High-Resolution Land Cover Map Production at the Country Scale Using Satellite Image Time Series“, Remote Sensing 9(12), DOI: 10.3390/rs9010095, 2017.
[2] Karakizi C., Karantzalos K., Vakalopoulou M., Antoniou G., “Detailed Land Cover Mapping from Multitemporal Landsat-8 Data of Different Cloud Cover“, Remote Sensing 10(8), DOI: 10.3390/rs10081214, 2018.
[3] Helber P., Bischke B., Dengel A., and Borth D., “EuroSAT: A Novel Dataset and Deep Learning Benchmark for Land Use and Land Cover Classification“, arxiv: 1709.00029, 2017.
[4] Russwurm M., and Koerner, M.,“Multi-Temporal Land Cover Classification with Sequential Recur- rent Encoders“, ISPRS International Journal of Geo-Information 7(4),DOI: 10.3390/ijgi7040129, 2018.
[5] Long J., Shelhamer E., and Darrell T., “Fully convolutional networks for semantic seg- mentation“, 2015 IEEE Conference on Computer Vision and Pattern Recognition, DOI: 10.1109/CVPR.2015.7298965, 2015.

Through its Sentinel satellite missions, the Copernicus Earth Observation programme provides more data in a higher quality than ever before with superior temporal, geometric and spectral resolution. At the same time, fast evolving IT technologies like big data, cloud-based processing as well as machine learning and recently in particular deep learning increasingly effect the Earth Observation sector and land monitoring applications. As a very recent example, GeoVille has utilized artificial intelligence, machine learning and big data analytics to process the entire data archive from the Copernicus Sentinel-2 mission to generate a new high-resolution map of vegetation across Earth’s entire landmass. The new map depicts for the first time global vegetation dynamics in high spatial resolution and gives unprecedented insight into land productivity around the globe.
GeoVille’s landmonitoring.earth processing engine is a system for retrieval of high resolution global land cover monitoring information, designed along Space 4.0 standards. The highly innovative framework provides flexible options to automatically retrieve land cover based on multi-temporal data streams from the Sentinel-1, Sentinel-2 as well as third party missions. Dedicated methods for data pre-processing, including automated quality flagging of all input data (based on a GeoVille-developed Sentinel-2 metadata database), sophisticated time series processing methods, smart interpolation methods, highly automated stratified machine learning classification approaches assure a robust and repeatable workflow.
The new global land cover dynamics layer is a four-dimensional datacube, which indicates for each location the abundance of biomass, its annual variability and the date when vegetation greenness reaches its peak. Besides information on the quantity of biomass, also the quality and the dynamic of the biomass across the year are detected.
At the heart of the new global vegetation dynamics database is a new multi-dimensional Land Dynamics index cube. This index cube inherits a new generation of vegetation phenology layers, providing time series metrics characterising the seasonal dynamics of vegetation phenology. In its current published version, the index encompasses only one seasonal phase derived from Tasseled Cap coefficients and seasonal metrics derived over the observation period, such as the sine and cosine curve, the constant value or offset of the curves, a linear, overall trend of the time series and the root mean square error to the data. The combination of these layers provides information on the amplitude and phase of the time series at the level of individual pixels, helping to minimize misclassifications through seasonal effects and changes.
In our presentation we will address different variants of such multi-dimensional Land Dynamics Indices to DeepLearning models and evaluate the performance improvements for fully automated land cover monitoring approaches in various operational contexts. This will cover a number of use cases, such as the Copernicus Land Monitoring service delivered on pan-European scale, applications in the context of SDG reporting as well as related mapping services for the international development sector.

According to the IPCC, the Global South will seriously suffer from climate change. As traditional rainfall patterns shift, accurate rainfall information becomes crucial for farmers to optimise food production and to ensure a sustainable future. The scarce rain gauge distribution makes rainfall analysis difficult in these regions. Satellites could offer a solution to this problem but present satellite products do not account for local characteristics and perform poorly, especially in (West) Africa. For example, for our pilot area around Wa in the Upper East Region of Ghana, the TAMSAT rainfall product correlates very poorly with ground measurements, grossly overestimates the number of rainy days, while underestimating the amount of rainfall per event.

SaS aims to overcome the lack of ground data and accurate satellite rainfall products through Citizen Science and advanced Machine Learning. The new satellite rainfall product uses an innovative combination of aerosols, cloud top temperature, and soil moisture products derived from diverse sensors onboard ESA’s Sentinel satellites (S1, S2, S3 and S5P), as well as MSG’s Aviris. These data will be fed into a Convolutional Neural Network (CNN) to estimate precipitation. Training and validation of the algorithm will be done through Citizen Science-based rainfall monitoring carried out at schools in Ghana, where the Trans-African Hydro-Meteorological Observatory (TAHMO) is active. A side benefit of the project will be the creation of a uniquely dense rainfall dataset in Ghana.

The designed algorithm first pre-processes the satellite data in order to create a combined data cube that can be fed to a CNN, which will output estimated rainfall rates. Data pre-processing is done using the Sentinel Application Platform (SNAP) software package and the CNN will be developed in Python, using the TensorFlow and Keras Machine Learning libraries. Literature suggests that aerosols are partially to blame for the poor performance of satellite rainfall products over Africa, which is why, in addition to traditional products such as cloud top temperature and soil moisture, TROPOMI (Sentinel 5P) data are included. All data and software used are open source.

On the ground, Citizen Science-based rainfall measurements will be carried out using inverted soda bottle rain gauges and rainfall sensing umbrellas, gathering the data with the mobile application Open Data Kit. TAHMO weather stations and disdrometers will provide data to cross-validate CS-based measurements. These technologies have been developed at TU Delft and tested in Africa and Nepal.

The first phase of the project development, in June 2019, will be to create the ground network in the study area in Northern Ghana. For this, a terrain assessment will be done taking into consideration factors such as safety and maintenance of the disdrometers and school locations. Workshops will be given in the area during which schoolchildren will be taught how to build and read their own rain gauges.

The presentation will cover the project design and first results as well as a comparison between resulting citizen science data, disdrometer and satellite data. Furthermore, insights into within-pixel variability and its effect the use of satellite rainfall estimates for farmers' decision-making on the ground will be discussed.

Once operative, the SaS rainfall product will guide agricultural extension agents and support crop insurance. Ultimately, SaS will contribute to economic growth and food security in the Global South.

Automating the IMGeospatial Data Pipeline

Company Information
IMGeospatial is a bootstrapped start-up currently without VC investment, whilst recently being valued at £4,000,000 in a seed investment round in November 2017.

They use machine learning, AI and multi-spectral, remote sensed data to provide Automated Business Intelligence to increase effectiveness and enhance the customer experience across a wide range of industrial and commercial sectors. IMGeospatial ‘Digests, Distils and Disseminates’ data to enable industry to achieve more whilst driving efficiency and cost-effectiveness.
For example Insurance sector users may choose their Land Use Classification, and Rebuild Calculator tools, whilst those in the water utility sector can call upon the Vegetation Proximity Index, or Unmetered Water Use tools. Their automated data pipeline then sends the results to those that need it, when they need it and in the best format for full integration with their systems.
By producing Automated Business Intelligence in this way, IMGeospatial can “CHANGE the way you see YOUR WORLD”, supplying critical answers to a user’s questions and allowing them to work smarter and do better things, more efficiently. With successfully completed demonstration projects for The World Bank, Affinity Water, the State of Nevada and Allianz UK to name a few, IMGeospatial is making significant improvements to many people's lives in communities across the globe.


ESA funded data pipeline project introduction
In February 2019 IMGeospatial embarked upon a major ESA funded project: Scaling AIMEE, their Automatic Intelligent Multi-feature Extraction Engine for Countrywide/Continental Deployment, in partnership with Affinity Water, Anglian Water and the World Bank.


The objective of this project is to develop a fully-automated data pipeline, through which they will provide customers and users with a genuinely game-changing opportunity to access automated business intelligence that is available in weeks not months or longer, which can be renewed in either 3, 6 or 12-month cycles, and represents savings in both time and expenditure of up to 60% compared to current survey methods.
Further, customers will benefit from IMGeospatial's technical development philosophy that facilitates a modular, bespoke approach, providing solutions to each individual company’s needs and wants every step of the way from initial request to delivery of data - automatically.


For the first time, Alexis Hannah Smith, IMGeospatial’s Founder and CEO will present a unique overview of this ground-breaking project, the first of its kind, explaining:

“IMGeospatial changes the way you see the world, by changing the way the world uses data. We deploy artificial intelligence in unconventional ways to digest, distil, and disseminate data efficiently, regardless of how imperfect or vast that universe of data becomes.
“We give clients across multiple sectors access to a structured pipeline of data that would have otherwise been too complex, inaccurate, or expensive to tap. By anticipating and preparing for a world where business intelligence is truly automated, we help you redeploy human resources to maximise return on investments in data technology.”


Project Objective
The objective of this project is to develop a fully-automated data pipeline, through which IMGeospatial will provide customers and users with a genuinely game-changing opportunity to access actionable business intelligence that is available in weeks not months or longer, which can be renewed in either 3, 6 or 12-month cycles, and represents savings in both time and expenditure of up to 60% compared to current survey methods. Further, customers will benefit from IMGeospatial’s technical development philosophy that facilitates a modular, bespoke approach, providing solutions to each individual company’s needs and wants every step of the way from initial request to delivery of data - automatically.

This ability is IMGeospatial’s unique selling proposition: no competitor can currently provide resilient intelligence that accurately reflects up-to-date features in a landscape.



Business Opportunity
Overview
IMGeospatial offers a commercial, autonomous platform to provide actionable business intelligence via DaaS (Data as a Service) currently to the insurance, power utility and water utility sectors, and NGOs, using their proprietary platform AIMEE (Automatic Intelligent Multi-feature Extraction Engine), coupled with EO data. They will also develop AIMEE’s application range to allow them to exploit further markets.
USP
AIMEE allows for the automatic extraction of multiple features from any given landscape, along with attributes associated with those features such as tree height, canopy size and probable root-ball mass, all achieved with no human intervention whatsoever.
This is in marked contrast to the current state-of-the-art provision which relies upon data analysts using huge datasets to ‘teach’ programs how to identify distinct features in a landscape such as roads, hedges, structures and bodies of water. AIMEE can learn the differences between such features autonomously, then extract them from the landscape along with their attributes, presenting the user with an easy to digest distillation of intelligence relevant to their individual needs.
Justification
IMGeospatial’s DaaS (Data as a Service) model is based on the concept that the service can be provided on-demand to the user regardless of geographic or organizational separation. Additionally, the actual platform on which the data resides is irrelevant. This is because customers do not buy a product, they subscribe to a service.

• There is no comparable system available from competitors (all competitors must supervise feature extraction)
• AIMEE can be configured with a range of modules to suit a wide variety of users within multiple industry/commercial sectors
• Global market can be exploited (developed and developing world can benefit from IMGeospatial technologies)
• Agnostic supply of satellite data
Strategy
By 2021, the global actionable business intelligence market is forecast to be worth 73.91 Billion USD.
Income from products from the Insurance sector worldwide is €560 million, and it is the intention to position IMGeospatial for an exit at an enterprise value (EV) of €224 million+ in 3-5 years



Target Sectors

1. Insurance Companies
Departmental Users
Commercial, underwriting, catastrophe and flood modelling, property risk modelling and pricing.

To be able to accurately assess a land parcel’s insurable status, insurers need to understand the nature and use of structures (residential/HMO, commercial/industrial/agricultural) and their attributes (roof type, footprint, total area, number of storeys, boundary and accessibility via transport networks).
IMGeospatial estimates insurance companies they have canvassed lose up to £12 million in revenue per annum through lack of accurate, resilient business intelligence on entities insured by them. Furthermore, the potential cost savings for utilising our technologies to address this problem represents an additional 200% ROI at our technology’s current development status.




2. Water Utility
Departmental Users
Optimisation, billing, leak detection, asset management, smart metering and catchment management.

To control cost, maximise revenue and provide value to customers, water utility companies must be able to understand a residential/commercial/agricultural building’s use and attributes. Knowing the occupancy probability for a residential building, for example, will help assess likely water service requirement and whether unauthorised water consumption is occurring.
The total estimated lost revenue due to the lack of reliable business intelligence is at least £1m per annum for a water company serving 1 million customers. The potential cost savings for utilising IMGeospatial’s technology to solve these problems represents an additional 300% ROI.

3. Energy Utility Companies
Departmental Users
Asset management, network maintenance and gas pipe surveying
For energy supply companies that IMGeospatial are in contact with, they estimate the total lost revenue due to the inability to accurately assess their operating environment for hazards is at least £2 million per annum.
4. NGOs
According to a new WHO/UNICEF Joint Monitoring Program (JMP) report , 2.1 billion people worldwide lack access to safe, readily available water at home, and 4.4 billion lack safely managed sanitation.
As a fundamental part of their remit, NGOs need actionable business intelligence to help monitor population migratory patterns, especially in times of war or humanitarian crisis. With plans and projections based on this intelligence, they can influence decision-makers in local and national government to create legislation for the improvement of citizens’ lives, creating stable local economies and bringing improved prosperity to their peoples and nations.
5. Service Providers
Within local and national government, housing staff, land management staff, environment officers, development agencies and water management departments will benefit from subscribing to IMGeospatial’s services which will offer them unrivalled access to easily digestible intelligence and assist in critical decision-making.

This work explores Deep Learning (DL) approaches to perform trend change detection in InSAR time series. Synthetic Aperture Radar (SAR) images are used in the Earth Observation (EO) domain to measure millimeter-scale ground deformation. The temporal evolution of surface deformation phenomena, such as subsidence, can be observed by using Interferometric SAR (InSAR) techniques to process a sequence of SAR images, acquired over the area of interest with a time interval that mainly depends on the satellite’s revisiting period. The output of this process consists of millions of displacement time series, each spanning the time period covered by the SAR image stack. One of the major efforts of the EO community is to automatically extract and classify the patterns contained in these time series in order to identify changes in trend or, more generally, anomalous behaviors. The applications for this are numerous, including environmental monitoring, urban planning, and emergency response. Statistical approaches have successfully been employed to accomplish this task, nevertheless, these kinds of methods are slow and do not scale with the vast amount of data acquired every day. Therefore, the need for fast while accurate automatic analysis tools has become increasingly important.
In this context, Artificial Intelligence and Machine Learning provide powerful methods that allow understanding the information contained in the data, by learning directly from the observations of a specific phenomenon. In particular, we consider Long Short-Term Memory (LSTM) networks, because of their ability to capture the temporal correlation among samples. These recurrent models are very effective in tasks for which a temporal analysis is fundamental, such as automatic translations, speech recognition, music generation, and many others. The power of this technology relies on the memory component, which allows understanding the information at a certain time, based on past observations. LSTMs are based on a system of learnable gates, which allow updating the features extracted by the memory cell as the sequence is processed. Given that the InSAR sequences clearly show a temporal dependency among the observations, this kind of architecture is perfectly suitable to identify the trends in this data.
This work aims to develop a specifically designed recurrent neural network (RNN) that makes use of multiple LSTMs to capture the temporal patterns in time series. The considered architecture performs a bidirectional analysis of the input sequence by considering both the past and future information to take the decision about the change point. This is accomplished by fusing the features extracted at each layer by two memory cells, one per direction, before being analyzed by the next recurrent layer. The extracted features are then used to solve a classification problem in which an additional classification network is put on top of the recurrent model to discriminate between normal and change points, i.e. the point belongs to a trend (class 0) or the point is a transition point between trends (class 1) respectively. A further adaption to the network allows for non-uniformly sampled time series, as for example InSAR sequences characterized by a variable lack of observations through time. This is accomplished by adopting a modified version of the original LSTM cell, known as Time-Gated Long Short-Term Memory (TGLSTM), in which additional time gates have been added to let the model learn to adapt the current understanding to the sampling rate.
Our goal is to show the results of our first exploratory experiments conducted on real InSAR acquisitions provided by TRE ALTAMIRA, our partner company in this project, giving an overview of the model and the adopted training procedure. Even if the proposed work consists of a preliminary analysis of the impact of Machine Learning techniques on the trend change detection problem, the results are promising and create the basis for future developments.