Machine Learning Methods for SAR-derived Time Series Trend Change Detection (MATTCH)

The MATTCH project – Machine Learning methods for SAR-derived Time Series Trend Change Detection – aims to apply Machine Learning techniques to InSAR (Interferometric Synthetic Aperture Radar) derived surface deformation measurements, with the goal of identifying, among the huge number of measurement points (MP) identified by advanced InSAR algorithms, the ones exhibiting displacement time series characterized by a change in trend or, more generally, an “anomalous behavior”. This data screening step is extremely important to support the End Users Community in the exploitation of frequently updated (every few days) and highly populated (millions of MPs) information layers resulting from advanced InSAR analyses over large areas.MATTCH aims to identify whether and how a Machine Learning approach can be applied successfully to the “data screening and data mining” step (with a particular emphasis on the detection of changes in trends), relying on the experience in SAR data processing of TRE ALTAMIRA and the extensive knowledge of POLIMI (Politecnico di Milano – Dipartimento di Elettronica e Informazione e Bioingegneria) about Machine Learning algorithms and their applications.To capture the temporal dependencies in the long displacement time series, the main Deep Learning architectures proposed for the analysis are Long Short-term Memory (LSTM) and Gate Recurrent Unit (GRU).The main objectives of the project are:Making SAR-derived surface deformation products more user-friendly and effective, supporting the analysis and the exploitation of InSAR-derived data, through the generation of a reliable layer of information driving the attention of the final users on a set “hotspots deserving special attention”;Enhancing the SqueeSARTM processing chain, via the implementation of a Machine Learning approach for time series trend detection, which is expected to improve the reliability and reduce the computational cost with respect to the statistical procedure currently in use;Increasing the knowledge about Machine Learning techniques applied to Earth Observation Big Data in both TRE ALTAMIRA and POLIMI groups, strengthening an effective cooperation between industry and academia in this relatively novel research field;Increasing the knowledge of Graphic Process Units (GPU) and cloud-based services to perform high throughput data processing and flexible scale-up;Improving the exploitation of ESA Sentinel-1 data, by creating innovative solutions, spurring new services to end-users and hopefully increasing the Earth Observation market