LCZ-Informed Analysis of Surface Urban Heat Island Intensity and Daily Thermal Dynamics Using CNN-Based Mapping and ECOSTRESS Data

To evaluate the application potential of convolutional neural network (CNN)-based Local Climate Zone (LCZ) mapping in urban thermal environment studies, this study employed a lightweight convolutional neural network model (Light-model) to classify LCZs within the area enclosed by the Fourth Ring Road of Xuzhou City. ECOSTRESS data obtained from summer (June to September) at different times were integrated to analyze the temporal and spatial changes of surface temperature (LST) and surface urban heat island intensity (SUHII). The classification results demonstrate that the Light-model achieved an overall accuracy of 84.46%, which is markedly higher than that of the random forest model (72.07%). It also outperformed random forest in built-up area identification (built-up overall accuracy: 69.41% vs. 46.91%) and non-built-up area identification (natural overall accuracy: 91.85% vs. 84.43%), as well as in Kappa coefficient and mean F1-score. Time-series analysis based on ECOSTRESS observations revealed a typical diurnal LST pattern characterized by the lowest temperatures before dawn, a peak in the afternoon, and a decline at night. High-density built-up zones (LCZ1–LCZ3) and large impervious areas (LCZ8) exhibited the highest daytime temperatures and the slowest nocturnal cooling, whereas bare soil areas (LCZF) showed the largest diurnal temperature range and the greatest fluctuations. Vegetation-covered and bare land zones (LCZA and LCZD) generally maintained lower temperatures, while water bodies (LCZG) functioned as persistent cooling sources throughout the day due to their high specific heat capacity. Overall, the findings suggest that CNN-based LCZ classification, when integrated with high-temporal-resolution LST observations, provides a reliable technical framework for urban thermal environment monitoring and regulation at the regional scale.