A multiscale and seasonal model for urban surface temperature prediction based on landscape, land use and spectral indices

Abstract

Land surface temperature(LST) data are crucial for agricultural production, climate change, and urban thermal environment studies. However, because the factors influencing LST distributions are unclear, it remains difficult to obtain high-precision, large-scale and spatially continuous LST data. Herein, support vector machine (SVM), random forest (RF), and extreme gradient boosting (XGBoost) models were established on the basis of Landsat 8 OLI_TIRS data from various seasons to predict the LST. Moreover, the performance levels of the different models were comparatively analysed across different spatial scales and seasons using Baohe District (Hefei city, Anhui Province), Xihu District (Hangzhou city, Zhejiang Province), and Yangzhong city (Zhenjiang city, Jiangsu Province), China, as the study areas. Among the three machine learning models, the XGBoost model provided the highest accuracy, and the accuracy stabilized with increasing spatial scales. When the spatial scale increased to 150–250 m, the value of evaluation indices such as the coefficient of determination (R²) remained stable, and the R² value exceeded 0.91. The prediction performance remained constant under repeated validation and spatial cross-validation. The model provided the best performance in summer (R²=0.95), followed by spring (R²= 0.94) and autumn (R²= 0.93), whereas the worst performance was obtained in winter (R²=0.90). At the various spatial and temporal scales, the remote sensing spectral indices contributed the most to the predictions. Most explanatory variables demonstrated stable performance across spatiotemporal scales, except for a few landscape elements sensitive to seasonal variations. The use of the five explanatory variables with the highest SHapley Additive exPlanations (SHAP) values, combined with variance inflation factor (VIF) analysis as input variables can guarantee high model prediction accuracy and computational efficiency. Because of its wide applicability and easily accessible input variables, the prediction model developed in this study could represent a novel technique for obtaining LST data with high spatial and temporal resolutions.