Generalized Image Segmentation Model for Multi-annual Retrogressive Thaw Slumps Mapping

Abstract

The railways and highways along the Qinghai Tibet Engineering Corridor (QTEC) were established on the frozen ground. Intensive engineering construction and human activities have significantly disturbed the permafrost environment. The melting of ice-rich permafrost closely relates to the carbon release. Beyond that, elevated pore fluid pressures may initiate or enlarge retrogressive thaw slumps (RTSs), which may further damage the foundation of critical transportation lifelines. However, the precise locations and margins of hundreds of RTSs around QTEC have not been systematically identified and delineated due to their remote locations and divergent surface features. The development of deep learning makes it possible to automatically and accurately identify and delineate the margins of RTSs. However, inventorying multi-temporal and large-scale RTS is challenged by the low generalization of deep learning model. Here we will apply the DeepLabv3+ segmentation algorithm to decipher 3-meter resolution PlanetScope optical images for a RTSs detection model. Fine-tuning, CycleGAN, and domain adversarial training will be used to improve the model's generalization ability. The time-dependent metric changes of RTSs will be investigated based on 2018-2022 multi-temporal RTS inventories. We will further extract the ground displacements over the mapped RTS using European Space Agency’s Copernicus Sentinel-1 satellite images and time-series Interferometric Synthetic Aperture Radar (InSAR) analysis. Our study leverages remote sensing big data and deep learning methods for hazard mitigation over the frozen ground in high environmental vulnerability.