Satellite-based persistent scatterer SAR interferometry is used to monitor the deformation of the earth’s surface or objects on it. With the increasing number, shorter repeat intervals and higher resolutions of wide-swath SAR systems data volumes are exploding. The deformation behavior over time of this large number of coherent radar scatterers can be estimated to gain deeper insight. Especially in the built environment this behavior may vary significantly between objects, also scatterers may have different qualities.

We developed a method aimed towards modeling the deformation behavior for each individual scatterer, the Individual Scatterer Model Learning method (ISML). This method uses machine learning algorithms which efficiently lead to significant improvements in data interpretation and parameterization as well as in correctly estimating the quality of the estimators.

A case study on a dataset of TerraSAR-X data over the Netherlands shows that the method is able to separate relevant classes of phenomena more efficiently. In the analysis the large amount of data is first reduced in dimensionality while preserving as much information contained in the dataset as possible. Next, scatterers which show similar behavior are grouped together allowing for the analysis of these groups as a whole. This reduces the computational burden of the analysis, facilitating the model estimation for each individual scatterer and obtaining a more realistic quality estimate of the results.

Figure 1 shows one of the results from the case study. Figure 1A shows an aerial photo of the observed Eemcentrale power plant. A top-down view is shown in figure 1B. The top-down view shows the location of the scatterers which show the high spatial density of scatterers. The red points are located on the roof of the building, where the blue points are located on the structure on top of this roof. This structure is an air inlet filter mounted on a steel frame which transfers air from outside to inside the power plant. The color of the scatterers corresponds to the output class of the ISML method and the corresponding time series are shown in figure 1C and figure 1D. The time series clearly show the different behavior of different (groups of) scatterers which are in close proximity of each other. A linear subsiding behavior with a slop of -1 mm/year is observed for the red point. This linear behavior is also observed for the blue points, but these points also show a seasonal dependence with an amplitude of 4mm. The daily average temperature is shown as the black line in figure 1D which after visual comparison gives rise to the hypothesis that the deformation behavior of the blue points is temperature dependent.

Figure 2 shows the scatterers corresponding to a different output class of the ISML method. The dataset in reduced dimensions is shown in figure 2A. The blue points in this figure are considered to be a separate class as determined by the ISML method based on their deformation behavior. The spatial location of the corresponding scatterers form linear structures as can be observed in figure 2B. These linear structures are in fact electricity posts as shown in figure 2B and figure 2C. This is an example of the output of the ISML method where the scatterers are spatially distributed but show a similar behavior. This is a significant results as the conventional InSAR analysis approach does not capture this kind of phenomena where scatterers show similar behavior but are spatially distributed.

The case study shows that the ISML method increases information extraction from large datasets of InSAR data by estimating an optimal parameterization of the temporal displacement time series for each individual scatterer in an unsupervised and automated way. This leads to significant improvements in data interpretation, parameterization and quality of time-dependent deformation assessments. The ISML method performs independent of the temporal complexion and spatial location of the data, leading to the possibility of performing a categorical deformation assessment thereby augmenting traditional spatial and temporal analysis.

The plethora of new collection platforms for Earth observation data from space, air, and ground is contributing significantly to the resurgence of machine learning (ML) and AI via deep learning, data mining, and statistical analysis-- or collectively being referred to as data science. Moreover, future collection system payloads are expected to increase in their capacity to acquire denser spatial, spectral, radiometric, and temporal information. Data assimilation (e.g., calibration and validation), and trans-disciplinary model development represent significant challenges facing data science.

A fundamental goal of the National Land Imaging Program (NLIP) of the U.S. Geological Survey (USGS) is to ensure public availability of a primary data record about the current state and historical condition of the Earth’s land surface. A significant portion of the NLIP resources are dedicated to supporting the ground segment of the Landsat Program via the Earth Resources Observation and Science (EROS) Center. The EROS Center manages over 40PB of free and open Earth science data, and conducts world-class remote sensing science. Future NLIP/EROS ambitions include advancing the Sustainable Land Imaging (SLIP) program to meet future user needs, as well as exploring the potential for integrating government and commercial space system data (large and small sat), and Unmanned Aircraft System (UAS) data, to support an operational Earth data science strategy. An overarching goal of this strategy is to advance our understanding of interconnected Earth processes via the implementation of big data analytic constructs that can integrate heterogeneous observables captured by in-situ and remote sensors, to better model and predict change over time.
In this paper we provide an overview of Earth observation data initiatives at USGS that are postured to support the needs of data science. Topics discussed include challenges and approaches to: 1) sustainable land imaging needs from space borne assets; 2) remote sensing data standards for storage, processing, and distribution to support operational and science needs; and 3) collaboration on domestic and international Earth observation collection and data management policy.

Differential Synthetic Aperture Radar Interferometry (DInSAR) is a remote sensing technique with important applications for geological and entropic processes studies by benefiting of a very spatially dense ground displacement maps with centimeter to millimeter accuracy.
In particular, starting from an appropriate sequence of DInSAR interferograms, it is possible to study the temporal evolution of the detected displacements through advanced approaches. An effective way to investigate this temporal behaviour is via the generation of deformation time-series. In this context, one of the most difficult task for the retrieval of displacement information is represented by the Phase Unwrapping (PhU) step, especially in presence of fast deformation phenomena or in area affected by decorrelation phenomena [1].
In this paper we present an innovative technique to correct the phase unwrapping errors in a generic deformation time-series retrieval process based on a guided scanning by means of a genetic algorithm (GA) [2]. This algorithm permits to manage sets of integers, representing the multiples of 2 pi that have to be added to the wrapped phase value, resulting in a solution coherent with the discrete nature of the phase unwrapping itself.
Once the unwrapped interferograms are inverted to generate the corresponding displacement time series, then the wrapped interferograms can be easily reconstructed from the time series itself. The phase difference between the wrapped interferograms and the reconstructed ones is higher for interferograms where a phase unwrapping error has occurred.
The presented algorithm considers a first inversion step based on the L1-norm minimization [3], which guarantees a more sharp and accurate identification of the interferograms that may need a correction with respect to the most common L1-norm inversion. For every single considered pixel, the algorithm compares the unwrapped interferograms reconstructed from the L1-norm inversion with the starting phase unwrapping solution. When the difference between these two solutions is higher than a fixed threshold, the corresponding interferogram is considered in the correction procedure as a gene of the GA chromosome.
The GA creates a large number of chromosomes, starting from the first one evaluated from the L1 inversion taken as reference. The rest of the chromosomes are creating by mutations of single genes and crossing over among two parent chromosomes. The former is based on a Poissonian distribution, since the allowed values of the genes are integers to be considered as 2pi multiples for the phase unwrapping. Every generation ranks and saves the most performing chromosomes from whom a new generation is created until convergence. Convergence is reached when a chromosome keep on staying in the first position of the rank for a certain number of generations, whose value depends on the chromosome length[4].
The objective function for the rank of the GA procedure is given by the temporal coherence [5] of the pixel, which is evaluated after the L2-norm inversion of the new unwrapped solution of the GA procedure. We realised that a pixel with an already low value of temporal coherence, which entails an high presence of noise and PhU errors, is not so sensitive to be corrected, or, in a worse case, can be corrected by GA procedure toward an algebraic solution with no physical meaning. So we decided to apply the GA only to pixel with a temporal coherence higher than a fixed threshold. In other words we consider pixels where the number of wrong unwrapped solutions is not so high.
In principles, the presented algorithm can be applied to more generic procedure of unwrapping, once the fitting function for the algorithm is identified.
In our case, we apply the correction technique to a solution obtained by the Small-Baseline Subset (SBAS) DInSAR which retrieve the deformation time series by relying on a least squares inversion (so based on a L2 norm) of an overdetermined system of equations representing the interferometric pairs involved in the processing [5].
The phase unwrapping is carried out through the well known EMCF (Extended Minimum Cost Flow) approach [6], based on a temporal and spatial phase unwrapping procedure. The above mentioned correction is applied at the end of the whole phase unwrapping procedure.
The parameters of the GA have been tuned according with the SBAS DinSAR technique, after a long testing phase, which included also old reconstructions and simulated data.
We recommend a redefinitions of the parameters for the application of the GA to different cases.
The results we obtained show that in case of strong deformations the algorithm maximise the temporal coherence resulting in an increasing of the deformation dynamics, according with the known effect of underestimation risk for the phase unwrapping algorithms [7].
Another important result is given by the preservation of spatial consistence of the corrections estimated by GA procedure, according with the phase unwrapping error behaviour. It is important to highlight that the space consistency is not intrinsically required by the GA procedure, but it emerges as a consequence.
To conclude, we observe that, comparing our correction technique with respect to a standard L1-norm corrected solution for retrieving phase unwrapping errors, the algorithm reduce the false alarm probability from 8% to 2%, based on simulated data.
In addition, the developed algorithm has been build up with a parallelization strategy characterized by a very high granularity level (pixel level) thus entailing an high scalability of the algorithm itself and corresponding reduced processing time.
The flow chart in figure 1 summarizes the correction procedure of the algorithm.

Figure 1. Description of the correction procedure starting from the solution of the EMCF phase unwrapping. Third, fourth and fifth step are depicted in separate blocks because the flow parallelization separates the set of points to be analysed in different subsets according with the calculation power.


References

[1] H. A. Zebker and J. Villasenor. Decorrelation in interferometric radar echoes, IEEE Transactions on Geoscience and Remote Sensing, vol 30, issue 5, pp. 950-959, September 1992.
[2] Mitchell, Melanie. An Introduction to Genetic Algorithms, Cambridge, MA: MIT Press, 1996. ISBN 9780585030944.
[3] Qin Lyu Zhouchen Lin, Yiyuan She, and Chao Zhang. A comparison of typical lp minimization algorithms, Neurocomputing, vol. 119, pp. 413-424, 2013
[4] Holland, John. Adaptation in Natural and Artificial Systems. Cambridge, MA: MIT Press. ISBN 978-0262581110.
[5] P. Berardino, G. Fornaro, R. Lanari, and E. Sansosti. A new algorithm for surface deformation monitoring based on small baseline differential car interferograms, IEEE Trans. Geosci. Remote Sent., vol. 40, no. 11, pp.2375-2383, 2002
[6] A. Pepe, R. Lanari. On the Extension of the Minimum Cost Flow Algorithm for Phase Unwrapping of Multitemporal Differential SAR Interferograms, IEEE Trans. Geosci. Remote Sent., vol. 44, no. 9, pp.2374-2383, 2006
[7] D. Just, R. Bamler. Phase statistics of interferograms with applications to synthetic aperture radar, Applied Optics, Vol. 33, Issue 20, pp. 4361-4368, 1994

GPS-enabled oceanic drifters are useful remote sensing tools employed worldwide, to build networks of measurement nodes on the sea surface, which provide valuable information about several relevant physical quantities (such as position, current and wind speed, salinity, etc.).
A wide array of such devices exists, ranging from industrially-produced and high-precision ones, providing high resolution telemetry and differential-positioning, to lower-budget ones, built with off-the-shelf components and meant to exist as fully disposable units.
A common problem faced in the research work, when employing any of such devices, stems from the task of receiving the data sent by them, and conveniently storing it, either for quick operational tasks (for example, while retrieving units stranded on segments of the coastline) or for further, more in depth processing.
We introduce a novel software platform, which is designed to collect and store data, sent from virtually any GPS-enabled oceanic drifter.
The software platform can be run either locally or on a distributed environment.
It allows for a custom number of access points to be defined: each access point is configurable so to be able to receive data sent through Hyper Text Transfer Protocol (HTTP) methods, or through low-level sockets, in virtually any text-encoded format. It is configured to support generic National Marine Electronics Association's (NMEA) -type strings as well as Sanav proprietary encoding.
The application stores the data in a PostgreSQL Database as a JavaScript Notated Object (JSON). JSONs are unstructured, flexible and readable by humans, thus removing the need for any specialized database language.
Once the data is stored, the application allows for exporting it by means of a search-and-filter graphical interface, into several common file formats (like CSV, XML, HTML, CSV, etc.).
The application also supports being extended by means of plugins. Plugins are small software units which can be enabled and allow for quick specific tasks to be performed on the collected data. An example of working plugin is provided with a real-time map visualization plugin.

Neural Network classification approach gives a very good results in task of land cover classification and crop classification. Using this approach [1,2] Space Research Institute of NASU and SAU produce every year crop classification map in country scale for Ukraine. This crop classification maps and land cover maps are more accurate then global products and have better spatial resolution 10 meters. Although this approach is one of the best, however, the result, as in all other classification algorithms, heavily depends on the training data. We use in-situ data collected during expeditions throughout Ukraine, therefore, we can judge them the most reliable, but with this approach to the creation of a land sample there are problems associated with errors in data collection, problems in the fields themselves and satellite data problems. For automatization of land cover classification we going to provide anomaly detection method for training data validation and correction. This approach we are testing and implementing in automate crop classification for Horizon2020 Era-Planet projects [3,4] in which we are use crop classification maps and land cover maps for crop state analysis and land degradation estimation [5].
In this study we have analysed 5 most popular methods for anomalies detection: SOM ANN, unsupervised SVM, isolation forest, robust covariance and local outlier factor. Anomaly detection methods usually build on clustering principles so trained model work as two classes classification model which output equal to 1 if input is normal sample or 0 if input is anomaly. Anomaly score for every polygon can be calculated as a percent of pixels that classified as anomaly in this polygon. We provide training of model for every class in training set. For training we select random pixels from each polygon in count of the minimum number of pixels among these polygons to balance the effects of all training samples.. After training model can be used for anomaly score calculation for every polygon. Anomaly detection algorithm for satellite data must pass few criterias: it must be fast, low false positive score [6], stability to the noises and anomaly data in training. The most important criterion is the last one because this approach must be used in automatization of training data creation and validation so a minimum human effort should be used to teach the model and the model should be resistant to possible noise and problems in optical, radar satellite data and mistakes in in-situ data. By this criteria winners are isolation forest and SOM ANN. Use of isolation forest is the most optimal because it ensemble method in which every tree in forest grove by random subset so it gives low false positive score, it very stabil to the anomaly data in training data and very fast. SOM ANN is accurate enaf too, but it as all artificial neural networks approaches is very dependent on hyperparameters tuning and sometimes it is very hard to tune this network for perfect result, but it is any way gives a very good result.
We provide this two anomalies detection approaches to detect anomalies in training data for agricultural classes in time series of Sentinel-1 and Sentinel-2 [7] satellite images. For testing we use our collected in-situ data set. This data and anomaly fields were used in training and validation. For fields with incorrect labels SOM ANN and isolation forest gives very high anomaly score in most cases, but some errors such as wheat marked as barley are quite difficult to find. Another kind of anomalies that we added in training set is incorrect fields edge detection, it’s polygons with threes, roads, grass , another artificial objects and polygons that include 2 or more fields. In this case anomaly score depends on the size of the anomalous part of the polygon, so these errors are very well detected. Third anomaly type that correctly detected related to the data - clouds in optical data and artifacts in SAR data very well detected by this methods too.

References
1. Kussul, N., Lavreniuk, M., Skakun, S., & Shelestov, A. (2017). Deep learning classification of land cover and crop types using remote sensing data. IEEE Geoscience and Remote Sensing Letters, 14(5), 778-782.
2. Lavreniuk, M., Shelestov, A., Kolotii, A., Vasiliev, V., Shumilo, L., & Kussul, N. (2018, April). Automated System for Crop Mapping in Amazon Web Services based on Sentinel Data. In EGU General Assembly Conference Abstracts (Vol. 20, p. 16629).
3. Shumilo L. , Kolotii A., Lavreniuk M. , Yailymov B. (2018 July). Use of land cover maps as indicators for achieving sustainable development goals. In IGARSS, Valencia, Spain. – 2018. – P. 830-833
4. Kussul, N., Lavreniuk, M., Sumilo, L., Kolotii, A., Rakoid, O., Yailymov, B., ... & Vasiliev, V. (2018, July). Assessment of Sustainable Development Goals Achieving with Use of NEXUS Approach in the Framework of GEOEssential ERA-PLANET Project. In XVIII International Conference on Data Science and Intelligent Analysis of Information (pp. 146-155). Springer, Cham.
5. Kussul, N., Kolotii, A., Shelestov, A., Yailymov, B., & Lavreniuk, M. (2017, September). Land degradation estimation from global and national satellite based datasets within UN program. In Intelligent Data Acquisition and Advanced Computing Systems: Technology and Applications (IDAACS), 2017 9th IEEE International Conference on (Vol. 1, pp. 383-386). IEEE.
6. M. Lavreniuk, N. Kussul, A. Shelestov, B. Yailymov, T. Oliinyk, and A. Kosteckyi,“Validation methods for regional retrospective high resolution land cover for Ukraine,” IEEE International Geoscience and Remote Sensing Symposium (IGARSS), pp. 4502-4505, 2016.
7. Lavreniuk, M., Kussul, N., Meretsky, M., Lukin, V., Abramov, S., & Rubel, O. (2017, May). Impact of SAR data filtering on crop classification accuracy. In Electrical and Computer Engineering (UKRCON), 2017 IEEE First Ukraine Conference on (pp. 912-917). IEEE.

ExtremeEarth is a H2020 project that aims at developing Extreme Earth Analytics techniques and technologies that scale to the petabytes of big Copernicus data, information and knowledge, and applying these technologies in two of the ESA TEPs (Food Security and Polar), demonstrating two highly societal and environmental relevant use cases.
Opening up the TEPs and DIASs by extracting information and knowledge hidden in the data, publishing this information and knowledge using linked data technologies, and interlinking it with data in other TEPs and DIASs and other non-EO data, information and knowledge can be an important way of making the development of downstream applications easy for both EO and non-EO experts.
Contrary to multimedia images, for which highly scalable Artificial Intelligence techniques based on deep neural network architectures have been developed recently, similar architectures for satellite images, that can manage the extreme scale and characteristics of Copernicus data, do not exist in Europe or elsewhere today. Training datasets consisting of millions of data samples in the Copernicus context do not exist today and published deep learning architectures for Copernicus satellite images typically run using one GPU and do not take advantage of recent advances like distributed scale-out deep learning.
The ExtremeEarth technologies will be demonstrated in two use cases with societal, environmental and financial value: The Food Security use case will develop high resolution water availability maps for agricultural area allowing a new level of detail for wide-scale irrigation support. The maps will be available as linked data together with other geospatial layers (e.g., OpenStreetMap, field boundaries, crop types etc.) and made available to farmers. The main objective of the Polar use case is to produce high resolution sea ice maps from massive volumes of heterogeneous Copernicus data. The maps will be made available as linked data and will be combined with other information such as sea surface temperature and wind information for informing maritime users.
The technologies to be developed will extend the European Hops data platform to offer unprecedented scalability to extreme data volumes and scale-out distributed deep learning for Copernicus data. The extended Hops data platform will run on a DIAS selected after the project starts and will be available as open source to enable its adoption by the strong European Earth Observation downstream services industry.
The technical approach of the project as well as the results of the first user workshop will be presented. ExtremeEarth is funded under grant number 825258. Its start is in January 2019.