Abstract
<title>Abstract</title> <p> Global warming has accelerated thaw settlement on the Qinghai-Tibet Plateau, significantly threatening infrastructure and environment, especially in the permafrost regions of northern Xizang. Traditional susceptibility mapping is often affected by random sampling bias and its performance is overestimated due to the neglect of spatial autocorrelation. This study develops an integrated framework coupling a two-step negative sampling strategy with XGBoost model supported by spatial block cross-validation. By combining the Information Value model and spy technology, we determined an optimal probability threshold (α <sub>min</sub> = 0.018) to filter potential risks and purify the training pool. Subsequently, we implemented spatial cross-validation to enhance the spatial generalization ability of the model and address the issue of performance overestimation. The results show that the two-step strategy significantly improves prediction reliability, achieving an AUC of 0.88. Solar Radiation, Equivalent Latitude, Altitude, NDVI, and △ALT were identified as the primary drivers, contributing over 60% to the model. The resulting map reveals that Very High risk zones (16.81% of the area) capture over 70% of historical events, demonstrating superior discriminative power. This scientifical approach provides essential decision-support for infrastructure protection and hazard mitigation in northern Tibet. </p>