DANI-NET: A Physics-Aware Deep Learning Framework for Change Detection Using Repeat-Pass InSAR

• Published in: IEEE Transactions on Geoscience and Remote Sensing
• Main Authors: Costa, Giovanni, Virgilio Monti Guarnieri, Andrea, Parizzi, Alessandro, Rizzoli, Paola
• Format: Article in Journal/Newspaper
• Language: English
• Published: 2025
• Subjects: Block structure; Change Points; Definitive Changes; SAR interferometry (InSAR); Synthetic Aperture Radar (SAR); Temporary Changes; X-AI; Iceland
• Online Access: https://hdl.handle.net/11311/1283849; https://doi.org/10.1109/tgrs.2025.3542179

Repeat-pass Interferometric SAR (InSAR) is widely used for a variety of application scenarios, such as terrain displacement and subsidence monitoring or measuring the state of infrastructures. In this context, the development of effective algorithms to detect temporal and spatial changes in the radar targets becomes of paramount importance. Typically, state-of-the-art methods only return the spatial, temporal, or both locations of the occurred changes without any information about the causes. In this paper, we present a novel change detection method able to infer not only whether a target has changed and when but also the reason why a change is detected, defining the concepts of definitive and temporary changes. This is done by jointly exploiting four radar amplitude images and the corresponding six interferometric coherences computed at different temporal baselines. To this aim, we propose a new deep learning-based framework based on a fully-convolutional neural network (CNN) called DANI-NET (Deep Analysis for Non-stable InSAR targets Network). The network design and training strategy are driven by explainable AI (XAI) principles. Here, we rely on the development of fully synthetic training and testing datasets by following a robust statistical derivation, which allows for a full understanding of the network outcomes. We evaluate the DANI-NET performance on an independent synthetic dataset and we compare it to the state-of-the-art Permutational Change Detection (PCD), a nonparametric statistical approach, achieving extremely competitive results. Moreover, we also provide a feature analysis on the prediction explainability using the SHAP method. Finally, we apply DANI-NET to two real-case scenarios, by considering a Sentinel-1 repeat-pass dataset acquired over Iceland during the 2023-2024 Sundhnúkur eruptions and a TanDEM-X multi-temporal stack acquired over an open-pit mining site. We validate the method over the Iceland dataset, where we compare the predicted lava field extension with external reference ...