Multi-source heterogeneous feature fusion framework for identifying retrogressive thaw slumps on the Qinghai-Tibet plateau

Abstract

Accurate detection of retrogressive thaw slumps (RTSs) remains a significant challenge due to their complex morphological features and subtle spectral contrasts with surrounding landscapes. This study presents a robust deep learning framework to improve RTS identification by integrating multi-source remote sensing data. Focusing on the permafrost-dominated central Qinghai-Tibet Plateau, we conducted three pioneering investigations: (i) a detailed RTS inventory integrated with topographic, environmental, spectral, and thermal analyses to uncover spatio-temporal distribution patterns; (ii) a comprehensive evaluation of twelve leading-edge semantic segmentation models for RTS identification; and (iii) the formulation of an innovative FusionSA-SegFormer model, which employs dual-level (pixel-level and feature-level) heterogeneous feature fusion of optical, spectral, thermal, and topographic remote sensing datasets. Our results highlight prominent RTS clustering within the 4700–4800 m elevation range, on moderate slopes (3–7°), in mid - slope to valley settings, and in proximity to water bodies. Temporally, the active evolution of RTSs from 2019 to 2024 was characterized by sustained degradation patterns, manifested as a continuous decline in vegetation indices and a concurrent rise in land surface temperatures. Comparative model assessments identified SA-SegFormer as the most effective baseline architecture. Building upon this, the proposed FusionSA-SegFormer demonstrated significant improvements, showing 8.8% IoU and 10.9% recall enhancements on the validation set, and achieving a superior F1-score of 0.843 (Precision: 0.838, Recall: 0.899) on the test set. Crucially, independent spatial and temporal transferability evaluations confirmed the framework's robust generalization capacity across unseen regions and varying years, maintaining consistent identification performance and high overall accuracies (>0.90). Furthermore, feature importance analysis emphasized the critical influence of spectral bands, particularly the blue band, alongside notable contributions from thermal indices. This work establishes a new benchmark for mapping permafrost disturbances and provides a valuable tool for monitoring thermokarst dynamics in warming climates.