Mapping the air temperature in China from time-normalized MODIS land surface temperature data via zone-based stacking ensemble models

High-resolution near-surface air temperature (<i>T</i>_<i>a</i>) data are important for studies on human life and plant. Remote sensing offers an effective means to acquire spatially continuous <i>T</i>_<i>a</i> over large areas. However, several critical issues have been overlooked in current remote sensing-based <i>T</i>_<i>a</i> estimation, including temporal inconsistencies of land surface temperature (LST), spatial variability over large-scale area, and limitations of single-algorithm approaches. This study developed a three-step ensemble framework to address these issues and map high-accuracy daily average <i>T</i>_<i>a</i> from remote sensing data. This framework consists of temporal normalization, zone-based modeling, and ensemble integration. First, Terra/MODIS LST was temporally normalized using ERA5 reanalysis data to eliminate the uncertainty caused by differences in observation times. Then, the whole study area was divided into subzones, and nine base models were developed in each zone using machine learning (ML) methods to estimate <i>T</i>_<i>a</i>. Finally, the <i>T</i>_<i>a</i> estimation results from the base models were integrated using five ensemble methods to develop an optimal integration strategy for mapping <i>T</i>_<i>a</i>. The results showed that temporal normalization effectively reduced the time difference in LST. Zone-based modeling exhibited enhanced performance compared to holistic modeling strategy. The stacking-based ensemble model outperformed each of the base models. Among them, the generalized additive model (GAM)-based ensemble model achieved highest accuracy of <i>T</i>_<i>a</i> estimation, with <i>R</i>² of 0.99 across all years, mean absolute errors (MAEs) ranging from 0.73℃ to 0.79℃, and root mean square errors (RMSEs) ranging from 0.99℃ to 1.07℃ from 2019 to 2023. This work provides a valuable reference for accurately mapping <i>T</i>_<i>a</i> from remote sensing data, especially at large scales.