Real-Time, Depth-Resolved Permafrost Thermal Monitoring Across the Qinghai-Tibet Plateau Enabled by Bayesian-Optimized Deep Learning

Climate warming accelerates permafrost degradation across the Qinghai-Tibet Plateau (QTP), which stores Earth’s largest mid-latitude frozen carbon reservoir. This process threatens global climate stability through greenhouse gas feedbacks and compromises infrastructure security. Existing ERA5 reanalysis products suffer critical limitations including a 2–3 months temporal latency that prevents real-time monitoring of freeze-thaw transitions, a reliable detection depth confined to 7 cm obscuring most active-layer dynamics, and systematic biases in heterogeneous terrain. To address these challenges, we developed a Bayesian-optimized ensemble deep learning model (BCL) that integrates convolutional neural networks (CNN) for spatial feature extraction with bidirectional long short-term memory (LSTM) architectures to reconstruct real-time, depth-resolved soil temperatures across the QTP. Trained on high-resolution in situ soil measurements spanning 0–40 cm depth and ERA5 meteorological drivers, the BCL model achieves superior accuracy and temporal scalability (hourly to monthly). Critically, the BCL framework simultaneously resolves ERA5’s latency limitation. Besides, spatial upscaling demonstrates >90% coherence with ERA5 in the 0–7 cm layer while simultaneously overcoming depth constraints, capturing subsurface thermal inertia. This study establishes a transformative tool for real-time permafrost thermal assessment, directly supporting climate feedback quantification and geohazard mitigation across cold regions.