Multi‐Annual Inventorying of Retrogressive Thaw Slumps Using Domain Adaptation
Journal Article
Overview
abstract
AbstractRetrogressive Thaw Slumps (RTSs), a form of thermokarst hazards, pose risks to hydrological and ecological environments and the safety of the Qinghai‐Tibet Engineering Corridor. We still lack the knowledge about the geographic locations of RTSs and their dynamically changing spatial margins. However, visual interpretation is labor‐intensive while the present‐day deep learning methods become ineffective when the model trained in one year is directly transferred to another. To enhance the model's generalization ability, here we implemented and compared three domain adaptation methods, that is, the classic supervised fine‐tuning method and two proposed unsupervised methods: Image StyleTransfer Domain Adaptation (ISTDA) and the Tversky Adversarial Domain Adaptation (TADA) network. In our proposed ISTDA, we uniformed the contextual information of multi‐temporal images by Cycle Generative Adversarial Network (CycleGAN). We introduced the Tversky loss and the automatic adjustment of weights for multiple loss functions to suppress false positives and to improve the generalization of TADA. We tested three methods' performance in Beiluhe region over the Qinghai‐Tibet Plateau using PlanetScope optical images during 2019–2022. The three domain adaptation methods are successful in generating regional, multi‐annual RTS inventories. Remarkably, TADA sustains good performance in complex transfer scenarios without additional label cost, achieving an F1 increase of 14.32%–24.17% compared to classic methods. Our work is the first to apply an unsupervised domain adaptation to automatically map the RTSs on a multi‐annual timescale, demonstrating a strong potential of its applicability for monitoring large‐scale, multi‐temporal evolution of geomorphological features.
