Domain adaptive semantic segmentation of multi-annual retrogressive thaw slumps.

Under review

In this study, we applied the DeepLab v3+ semantic segmentation model to 3-m-resolution PlanetScope optical images to construct a single-temporal RTS detection model in the Beiluhe region, and applied three domain adaption methods including fine-tuning based domain adaptive segmentation, Cycle-Consistent Generative Adversarial Networks (CycleGAN) based domain adaptive segmentation, and Domain Adversarial Neural Networks (DANN) based domain adaptive segmentation respectively to improve its temporal generalization ability.