Method to predict urban heat island hot spots at microscale level based on convolutional neural network land cover

The ongoing shifts in climate change suggest an increase in urban temperatures during hot periods, emphasising the need for supported public policy to be more efficient in reducing Urban Heat Island effects. This paper presents a CNN-based method to classify land cover and simulate micro-scale temperature variations for identifying UHI hot spots. The five primary urban heat sources - buildings, grass, trees, roads, and water - are analysed for their impact on local temperatures. Using a dataset of 775,000 tiles (60 × 60 pixels), dedicated CNN models are trained for each class. Model validation achieved over 97% accuracy for all classes except roads, which reached 89.1%. Simplified microscale models simulate temperature variations, producing a map that visualises UHI hot spots across urban areas. The results highlight the tool's effectiveness and practicality in identifying high-temperature zones, offering valuable support for UHI mitigation and targeted policy efforts.