KDA: Knowledge Distillation Adversarial Framework With Vision Foundation Models for Landslide Segmentation

Abstract

Landslides pose severe threats to infrastructure and safety, and their segmentation in remote sensing imagery remains challenging due to irregular boundaries, scale variation, and complex terrain. Traditional lightweight models often struggle to capture rich semantic features under these conditions. To address this, we leverage vision foundation models (VFMs) as teachers and propose a knowledge distillation adversarial (KDA) framework to transfer high-capacity knowledge into compact student models. Additionally, we introduce a dynamic cross-layer fusion (DCF) decoder to enhance global–local feature interaction. The experimental results demonstrate that, compared to the previous best-performing model SegNeXt [89.92% precision and 84.78% mean intersection over union (mIoU)], our method achieves a precision of 91.93% and mIoU of 86.53%, yielding improvements of 2.01% and 1.75%, respectively. Source code is available at https://github.com/PreWisdom/KDA