Reconstructing all-weather remotely sensed air temperature via a kernel-based temporal filling and bias correction (KTF-BC) framework

Abstract

Thermal infrared (TIR) remote sensing provides an effective means of mapping near-surface air temperature (T_a) at large scales. However, cloud coverage introduces substantial data gaps, posing a considerable challenge for producing all-weather T_a datasets. This study proposed a two-step framework, termed Kernel-based Temporal Filling and Bias Correction (KTF-BC), to reconstruct all-weather remotely sensed T_a. In the first stage, a kernel-based temporal filling method was developed to estimate the theoretical clear-sky T_a for cloud-covered pixels. In the second stage, a bias correction model was constructed to adjust these theoretical estimates toward the actual T_a under cloudy conditions. The proposed framework was applied across China from 2019 to 2023 to generate spatially complete daily mean T_a at 1-km resolution. Validation against meteorological stations under cloudy conditions demonstrated consistently high accuracy, with R² values up to 0.99, mean absolute errors (MAEs) ranging from 0.91 to 0.95 °C, root mean square errors (RMSEs) ranging from 1.21 to 1.26 °C, and biases close to 0 °C. The method effectively captured fine-scale thermal heterogeneity and demonstrated robust performance across varying cloud conditions and surface environments. This study provides a practical and reliable solution for reconstructing all-weather T_a from satellite observations.