An improved machine learning-based model for prediction of diurnal and spatially continuous near surface air temperature

Abstract

Abstract Near-surface air temperature (Tair) is crucial for assessing urban thermal conditions and their impact on human health. Traditional Tair estimation methods, reliant on sparse weather stations, often miss spatial variability. This study proposes a novel framework using a federated learning artificial neural network (FLANN) for fine-scale Tair prediction. Leveraging spatially complete thermal data from Landsat 8/9, Sentinel 3, and Himawari 8/9 (105 acquisition days, 2013–2023), and data from automatic weather stations, 23 predictor variables were extracted. After rigorous selection processes, nine variables significantly correlated with Tair were identified. Comparative analysis against established machine learning and linear models, using cross-validation data, showed FLANN’s superior performance with a Pearson correlation coefficient (r) of 0.98 and a root mean square error (RMSE) of 0.97 K, compared to r and RMSE of 0.85 and 1.09, respectively, for the linear model. FLANN showed greater improvements for urban stations with r and RMSE differences of 0.19 and − 2.03 K. Application of FLANN to predict Tair in Hong Kong in July 2023 enabled detailed urban heat island (UHI) analysis, revealing dynamic spatial and temporal UHI patterns. This study highlights FLANN’s potential for accurate Tair prediction and UHI analysis, enhancing urban thermal environment management.

Keywords

• Federated learning
• Neural network
• Near-surface temperature
• Urban heat island
• Remote sensing
• Machine learning