Improving Ground Deformation Classification by Integrating InSAR Time Series With Geospatial Information

Classifying ground deformation processes, such as landslides, subsidence, deep-seated gravitational slope deformations (DSGSDs), and mining-induced deformations, is key for large-scale hazard assessment and national land-use management. Earth observation provides heterogeneous data over the same geographic region, including interferometric synthetic aperture radar (InSAR) time series (TS), multispectral imagery, and terrain products. However, effectively integrating such spatiotemporal information from multimodal datasets remains a major challenge. To fully use the rich information contained in the TS and to exploit the complementary strengths of spatial and temporal data, we propose a dual-branch deep learning (DL) approach that integrates InSAR ground deformation TS with geospatial information for classifying slow-moving ground deformation processes. To validate the approach, we construct a ground deformation dataset containing over 26 000 active deformation areas (ADAs), labeled into four deformation types: landslide, subsidence, DSGSD, and mining. Results demonstrate that our model achieves an overall classification accuracy exceeding 90% on both ascending and descending test datasets, though confusion remains between certain classes, such as landslides and DSGSD. Explainable AI (XAI) analysis indicates that spatial and morphological features contribute more significantly to classification performance than temporal deformation patterns, with clearer distinctions for subsidence and mining, but more overlap between landslides and DSGSDs. This work highlights the strength of multimodal data fusion method to classify ground deformation processes, while setting the stage for future research.