Physics-guided machine learning approach for reconstructing air temperature in warm permafrost on the Qinghai‒Xizang Plateau

作者：C.H. Peng, Dong-Liang Luo, Yu Sheng, Ji-Chun Wu, Jia Liu, Shi-Zhen Li, Ya-Juan ZAO · 发表于：Advances in Climate Change Research · 年份：2026 · DOI：10.1016/j.accre.2026.04.013 · 研究领域：Climate change and permafrost、Soil Moisture and Remote Sensing、Smart Materials for Construction

High-resolution air temperature data are essential for quantifying eco-hydrological processes in climate-sensitive regions like the Qinghai‒Xizang Plateau. However, in-situ observations are frequently interrupted by extended data gaps. To address this, we developed a physics-guided machine learning (PGML) framework to reconstruct a 9-mon gap in a 30-min resolution temperature series. The framework employs a LightGBM model that adaptively integrates bias-corrected regional climate forcing, a reconstructed NDVI phenology proxy, and intermittent local in-situ observations. A key innovation is the design of a custom objective function that embeds domain-specific physical constraints, specifically permafrost thermal stability and thermodynamic consistency, directly into the learning process. Comparative experiments revealed that, while standard physically-agnostic models achieve high overall accuracy, they fail to capture the high thermal inertia of frozen ground, resulting in unphysical warm biases under extreme cold conditions (RMSE = 1.79 °C). In contrast, the PGML model effectively mitigates these artifacts, reducing errors in the critical cold regime by 36% (RMSE = 1.14 °C). Across the full record, residuals remain centered near zero (mean = −0.10 °C), and the predicted 90% confidence interval achieves an empirical coverage of 90.1%. SHAP analysis further reveals that these physical constraints regularize the model’s inferential logic, enabling an adaptive shift from transien...