A deep Learning-based approach for landslide dating from time-series of SAR data

Landslides are a serious geologic hazard common to many countries around the world.  They can result in fatalities and the destruction of infrastructure, buildings, roads, and electrical equipment. Especially rapid-moving landslides, which occur suddenly and travel at high speeds for miles, can pose a serious threat to life and property. Landslide inventories are essential to understand the evolution of landscapes, and to ascertain landslide susceptibility and hazard, and it can be of help for any further hazard and risk analysis. Although  many landslides inventories have already been created worldwide, often these archives of historical landslide events  lack precise information on the date of landslide occurrence. Many of these inventories also lack completeness especially in case of smaller landslides which is also caused by  landslides erosion processes, human impact, and vegetation  regrowth. Precise determination of landslide occurrence time is a big challenge in  landslide research. Optical and Synthetic Aperture Radar (SAR) images with multi-spectral and textural features, multi-temporal revisit rates, and large area coverage provide opportunities for landslide detection and mapping. Landslide-prone regions are frequently obscured by cloud cover, limiting the utility of optical imagery. The capacity of SAR sensors to penetrate clouds allows the use of SAR satellite data to provide a more precise temporal characterization of the occurrence of landslides on a regional scale. The archived Copernicus Sentinel-1 satellite, which has a 6 to 12-day revisit period and covers the majority of the world's landmass, allows for more precise identification of landslide failure timings. The time-series of SAR amplitude, interferometric coherence, and polarimetric features (alpha and entropy) have strong responses to landslide failures in vegetated regions. This is characterized by a sudden increase or decrease in their values. Consequently, the abrupt shifts in the time-series of SAR-derived parameters, triggered by the failure, can be recognized and regarded as the failure occurrence time. The aim of this study is to determine the time period of failure occurrences by automatically detecting abrupt changes in the time series of SAR-derived parameters. We present a strategy for anomaly detection in time-series based on deep-learning to identify the failure time using four parameters derived from SAR time series. In this strategy, we introduce a gated relative position bias to an unsupervised Transformer model to detect anomalies in a multivariate time-series composed of four SAR-derived parameters. We conduct an experiment involving multiple landslides and compare the performance of our proposed strategy for detection of the failure time period with that of the LSTM model. Our strategy successfully identifies the time of landslide failure, which closely approximates the actual time of occurrence when compared to the LSTM model employed in this study.