Satellite data analysis for monitoring of territory, infrastructure and mobility

According to the European Association of Remote Sensing Companies (EARSC), the Earth Observation (EO) sector is continuously growing together with the number of EO satellites launched every year. In the next future, the capability of producing more and more data of the Earth should be supported by a capability to download and process all this data as fast as possible. Just providing raw data isn't enough, the sector needs to develop new algorithms and tools to turn Big EO data into information for users. Based on current market demands and considering the rapid development of the sector, the main objective of this research work is to develop new methodologies and algorithms for the analysis of EO images that can contribute to turn EO data into actionable insights for users. The presented research is divided into two parts, the first related to the combination of space-based Synthetic Aperture Radar (SAR) data with information of other variables that could influence the events under observation the prediction of future ground movements by means of a neural network. The second concerns the analysis of SAR and optical images directly on board the satellite for near real time monitoring applications. In the first research work Interferometry SAR Persistent Scatterer Interferometry (InSAR PSI) data are used for monitoring and measuring ground deformation. InSAR PSI enables the precise measurement of surface deformation rates with millimeter-level accuracy and can be used to monitor infrastructures and natural areas. Some scientific works present the potentiality of statistical and analytical methods to make prediction on future movements of the territory. However, when the phenomena complexity increases analytical methods lose precision. The objective of this research work is to explore new methods to improve the capabilities of predicting future movements by combining Artificial Intelligence (AI) with the fusion of different types of data. I this research has been proposed the methodology called Context-Based InSAR Prediction System (CIPS). It aims to train a Long Short-Term Memory (LSTM) Recurrent Neural Network (RNN). It has been specifically designed for capturing long-term dependencies, combining inSAR space-based measurements with additional information regarding other variables that contribute to describe the context of the event. As an example, this general methodology has been tested in a real scenario of mudslides, where InSAR displacement measurements have been combined with meteorological rain data taken in the same periods, resulting in better predictions of future territory movements. The second part of the research work has been stimulated by the need to detect and monitoring extreme events in quasi-real time, while currently this capability is limited by long revisit time and hence unacceptable latency. In address this point, new algorithms for edge computing have been developed to analyze optical and SAR images directly in Space, to extract the requested information before in-orbit and download only the results of the elaborations instead of the full raw dataset. The proposed algorithms make use of Convolutional Neural Networks (CNN), and their implementation has been performed in nanosatellite-compatible hardware, resulting in the design, implementation and test of the called Space Edge Computing Workflow (SEC-W). The SEC-W makes possible to obtain a scalable way to produce CNN models able to perform image segmentation directly in space, permitting to prioritize the optical and/or SAR images to be downloaded and to provide to the end users with ready-to-use, nearly-real-time information. SEC-W showed the capability to perform the segmentation of raw images with a very significant reduction of the data volume to download. This makes it possible to downlink a ready to use segmented image to the ground in few seconds, allowing a faster response to sudden and critical territory changes.