Mountainous land types are characterized by a scarcity of observational data, particularly in remote areas such as the Kunlun Mountains, where conventional Automatic Weather Stations (AWSs) typically do not record land surface temperature (LST) data. This study aims to develop and evaluate models for converting air temperature (TA) to LST using newly established meteorological station data from the Kunlun Mountain Gradient Observation System, thereby providing time-continuous LST data for AWSs. We constructed a conceptual model to explore the relationship between 1.5 m TA and LST and instantiated it using three machine learning algorithms: Support Vector Machine (SVR), Convolutional Neural Network (CNN), and CatBoost. The results demonstrated that the CatBoost algorithm outperformed the others under complex terrain and climatic conditions, achieving a coefficient of determination (R2) of 0.997 and the lowest root mean square error (RMSE) of 0.627 °C, indicating superior robustness and accuracy. Consequently, CatBoost was selected as the optimal model. Additionally, this study analyzed the spatiotemporal distribution characteristics of cloud cover in the Kunlun Mountain region using the MOD11A1 product and assessed the uncertainties introduced by the 8-day average compositing method of the MOD11A2 product. The results revealed significant discrepancies between the monthly average LST derived from polar-orbiting satellites and the hourly composite monthly LST measured on-site or under ideal cloud-free conditions. These differences were particularly pronounced in high-altitude regions (4000 m and above), with the greatest differences occurring in winter, reaching up to 10.2 °C. These findings emphasize the importance of hourly LST calculations based on AWSs for accurately assessing the spatiotemporal characteristics of LST in the Kunlun Mountains, thus providing more precise spatiotemporal support for remote sensing applications in high-altitude regions.
Read full abstract