Change detection of slow-moving landslide with multi-source SBAS-InSAR and Light-U2Net

Interferometric Synthetic Aperture Radar (InSAR) techniques are commonly used approach for identifying Slow-moving Landslide (SML). However, most SML boundary identification with deep learning are based on single-source InSAR data, which cannot fully explore the dynamic process of destabilization, and are inefficient due to high model complexity. Meanwhile, research on automatic procession with multi-source InSAR data is few. To enhance efficiency in geohazard monitoring, this paper proposed an automatic framework for Boundary-Changed Slow-moving Landslide (BCSML) detection by integrating multi-source Small Baseline Subset InSAR (SBAS-InSAR), Convolutional Neural Network (CNN), and change detection methodologies. Firstly, surface deformation was estimated using multi-source SBAS-InSAR. Then, a novel and effective Light-U²Net was constructed with decreased complexity to identify Significant Deformation Zone (SDZ) and locate SML candidate. Finally, BCSMLs were identified using a change detection approach based on newly defined geometric measurements. Two study areas were selected to test the model’s performance: Zayu County and the Nu-Lancang River parallel flow (NLPF) area (in China). The proposed Light-U²Net model achieves high Precision (80.1 %), Recall (80.2 %), and F1-scores (80.1 %) in Zayu County. Additionally, the model’s complexity has reduced by 42.4 % without compromising identification accuracy compared to the original model. The pre-trained model was then applied to the NLPF area, and a total of 273 BCSMLs were detected, with 176 identified as expanding and 97 as shrinking. BCSML identification accuracy can reach to 90.47 %. The results have proved that the proposed framework with the Light-U²Net model is effective and practically potential in landslide disaster prevention.