Beyond the riverbank: deep learning and individual conditional expectation for geospatial assessment of urban river cooling

This study investigates how urban rivers mitigate the urban heat island effect in a tropical coastal city. We examine a 3000 m buffer surrounding the rivers in Da Nang’s urban core, Vietnam, during the dry–hot period from May to August 2025. Using remote sensing data, geospatial analysis, and a deep fully connected neural network (DFCNN), we quantify the river’s cooling influence in terms of effective cooling range and average cooling intensity within the riparian zone. Explanatory variables—including distance to the river boundary, local river width, topography, land use/land cover, urban morphology, and proximity-based measures—were used to model spatial variation in land surface temperature (LST). The DFCNN effectively captured nonlinear relationships between LST and its conditioning factors. This method achieved a mean absolute percentage error of 2.56% and explaining 87% of LST variation. In addition, the individual conditional expectation (ICE) analysis, combined with the DFCNN model, estimated an effective cooling range of approximately 1470 m and an average cooling intensity of 1.38 °C. Constrained ICE was then applied to examine how built-up density and green space density influence the cooling performance. The results indicate that increased green space enhances cooling intensity, whereas higher built-up density has the opposite effect. The research findings provide data-driven insights for nature-based urban heat mitigation in Da Nang and similar tropical coastal cities.