Title : Optical–SAR and GIS fusion with deep learning for InSAR coherence enhancement in deformation mapping
Abstract:
The integration of artificial intelligence and remote sensing data into Geographic Information Systems (GIS) is transforming the capabilities of Earth observation and spatial analysis. Within this context, accurate deformation monitoring using Differential Interferometric Synthetic Aperture Radar (DInSAR) remains challenged by spatial and temporal decorrelation in interferograms.
Recent studies (e.g., Zhou et al., 2023) have shown the potential of deep learning to mitigate atmospheric and topographic disturbances in InSAR time series. In this work, we propose an innovative workflow that combines optical–SAR data fusion, a hybrid U-Net/Transformer architecture, and full GIS integration, aiming to generate enhanced coherence maps and extract reliable deformation time series.
The processing chain includes precise geometric co-registration, extraction of optical features (NDVI, NDBI, texture indices), multimodal patch construction (SAR amplitude + optical bands + indices + raw coherence), and a dual-branch deep model that predicts coherence through nonlinear inter-modal correlations, inspired by the DeepInSAR framework.
Furthermore, self-supervised learning and Physics-Informed Neural Networks (PINNs) are incorporated to reduce dependency on labeled data and to enforce physical consistency (phase periodicity, signal energy conservation). These mechanisms improve model generalization across different sites and extreme events, in line with recent developments in physics-informed geohazard modeling (Moeineddin et al., 2023).
The predicted coherence maps act as quality indicators for guiding phase unwrapping and DInSAR inversion, significantly improving deformation accuracy and reliability. Final products are integrated within a GIS environment (GeoPackage / WFS / WMS) for multi-layer spatial analysis (enhanced coherence, displacement, DEM, land cover), vulnerability zoning, and decision support in land and infrastructure monitoring. This unified approach, supported by recent advances in multimodal data fusion and AI-based geospatial analytics (Sensors, 2025), illustrates the convergence of Earth observation, deep learning, and geoinformatics for more accurate and operational deformation mapping.
Keywords: DInSAR, InSAR coherence, Optical–SAR fusion, U-Net, Transformer, Self-supervised learning, PINN, GIS, Deformation mapping.
