WP5

The InSAR phase deformation signals dataset will be stacked with DEM-derived morphological layers and quality control layers, and will be used to inform a deep learning model for object/feature detection. Signal amplitude (i.e. backscatter) and coherence will also be analysed to capture and correlate changes related to surface roughness and its dielectric constant. The library will train a Deep Learning based object detector (Convolutional Neural Network, CNN).

The procedure will be iteratively trained, validated, and used on DInSAR products covering different temporal periods, provided by WP3 and WP4. The setup of the CNN will require:

• to investigate different filters in the preparation of DInSAR products (WP3) to better highlight signatures in interferometric phase patterns.

• to define a proper structure of the CNN including number and type of layers, bounding box sizes, and stack preparation (e.g. stack size, resampling, and scaling)

• to refine hyperparameters tuning with overfitting mitigation strategies. This procedure will be iteratively trained, validated, and used on DInSAR products covering different temporal periods, provided by WP3 and WP4.

The main purpose will be to recognize signals related to specific classes of mass movements in new interferograms with the aim of automatically and rapidly identify and classify processes.