Deep learning models integrating multi-sensor and -temporal remote sensing to monitor landslide traces in Vietnam

Abstract

Landslides pose significant threats to lives and public infrastructure in mountainous regions. Real-time landslide monitoring presents challenges for scientists, often involving substantial costs and risks due to challenging terrain and instability. Recent technological advancements offer the potential to identify landslide-prone areas and provide timely warnings to local populations when adverse weather conditions arise. This study aims to achieve three key objectives: (1) propose indicators for detecting landslides in both field and remote sensing images; (2) develop deep learning (DL) models capable of automatically identifying landslides from fusion data of Sentinel-1 (SAR) and Sentinel-2 (optical) images; and (3) employ DL-trained models to detect this natural hazard in specific regions of Vietnam. Twenty DL models were trained, utilizing three U-shaped architectures, which include U-Net and U-Net3+, combined with different data-fusion choices. The training data consisted of multi-temporal Sentinel images and increased the accuracy of DL models using Adam optimizer to 99% in landslide detection with low loss function values. Using two bands of the Sentinel-1 could not define the characteristics of landslide traces. However, the integration between Sentinel-2 data and these bands makes the landslide detection process more effective. Therefore, the authors proposed a consolidated strategy based on three models: (1) UNet using four S2-bands, (2) UNet3+ using four S2-bands, (3) UNet using four S2-bands and VV S1-band, and (4) UNet using four S2-bands and VH S1-band for fully detect landslides. This integrated strategy uses the capabilities of each model and overcomes model result constraints to better describe landslide traces in varied geographical locations.