Automated detection of active mass movements in SAR interferograms using Deep Learning

Slow mass movements are widespread players of slope dynamics, with different mechanisms depending on involved materials and geomorphic settings. Alpine para/periglacial environments are extensively affected by slow rock-slope deformations, deep-seated rock and debris slides, and active periglacial features, while fluvial-dominated mountain ranges are typically affected by rapid rockslides and long-lived earthflows. These processes exhibit different deformation patterns and rates, threatening lives and infrastructures in different ways. Mapping and monitoring slow mass movements is thus essential for civil protection, land management, and disaster risk reduction, requiring capabilities to rapidly map and classify processes over large areas.

Current regional-scale approaches to capture mass movement activity rely on geomorphological techniques supported by remote sensing. These approaches are accurate but time consuming and difficult to update. Such gaps could be filled using artificial intelligence techniques, currently mostly based on the interpretation of optical imagery or multitemporal InSAR data. Nevertheless, mass movements are often too fast to be captured by multitemporal InSAR and too slow for optical or amplitude SAR image analysis. Dual-pass satellite DInSAR products offer a valuable alternative to study these intermediate processes by the analyses of interferometric fringes, yet they suffer from noise, artifacts, and unwanted signals due to atmospheric disturbances.

We propose a deep learning model to automate the detection and classification of different types of mass movements in different geological and geomorphological settings through the interpretation of deformation fringes in DInSAR interferograms. To this aim, we use a YOLO, a convolutional object detector, aimed at interpreting routinely available wrapped interferograms. To mirror the interpretative process carried out by a human expert, input data include interferograms, a compound measure of the reliability of the interferogram, and a composite layer of geomorphological and morphometric information.

To train our net, we developed a geomorphologically constrained methodology to construct libraries of labeled expert-interpreted InSAR phase signal, corresponding to different mass movements recognized in two large (10³ km²) test areas in the Central Alps (Lombardia) and Apennine (Emilia-Romagna) of Italy, representing diverse processes and geological settings. The model is tested with sets of routinely generated SAR interferograms, to produce automated maps able to detect and classify mass movements over different timescales. This approach promises to streamline the rapid generation and update of active landslide inventories, to support local-scale landslide monitoring plans and civil protection actions, and improve the integration of data into landslide modeling efforts.