Abstract
The near-surface air temperature (NSAT) is crucial for understanding thermal and urban environments. Traditional estimation methods using general remote sensing images often focus on the types of spatial data or machine learning models used, neglecting the importance of seasonal and temporal variations, limiting their accuracy. This study introduces a novel ensemble model that incorporates both seasonal and temporal information integrated with satellite-derived land surface temperature (LST) data to enhance NSAT estimation, along with a rigorous feature importance analysis to identify the most impactful parameters. Data from 2022, collected from 147 South Korean weather stations, were used to develop and evaluate the models. Thirteen initial variables, including the LST and other auxiliary data, were considered. Random forest regression was employed to build separate models for each season. This novel approach of separating data by season allowed optimized feature selection tailored to each season, improving the model efficiency and capturing finer seasonal and daily temperature variations. These seasonal models were then combined to form an ensemble model. The seasonal models demonstrated varying accuracy, with the R(2) values indicating a strong correlation between the predicted and actual NSAT, particularly high in spring and fall and lower in summer and winter. The ensemble model showed improved performance, achieving an MAE of 0.534, an RMSE of 0.391, an R(2) of 0.996, and a cross-validated R(2) of 0.968. These findings highlight the effectiveness of incorporating seasonal and temporal information into NSAT estimation models, offering significant improvements over traditional approaches. The developed models support precise temperature monitoring and forecasting, aiding environmental and urban management.