Predicting Heat Stress Level with Convolutional Neural Network Using Local Climate Zone Pattern and High Density Air Temperature Observation Network: A Case Study of Chiayi City

With the rapid advancement of urbanization, the urban heat island (UHI) effect has significantly intensified. Traditional studies primarily rely on linear correlation analyses between land use ratios and temperature, often failing to explore the detailed impacts of different urban morphology distribution patterns on microclimates. This study integrates techniques and methodologies from convolutional neural networks (CNN) in machine learning to assess the intensity of the UHI. The study applies Landsat 8 satellite imagery to collect comprehensive data on the built environment, overcoming past limitations of non-real-time data availability. By employing the Local Climate Zones (LCZ) system, which systematically classifies the three-dimensional features of urban morphology, including building and landscape height and density, this method provides a nuanced understanding of urban thermodynamics that traditional approaches based on area ratios cannot achieve. CNN is utilized to analyze LCZ classifications and their correlation with temperature variations. This machine learning technique enhances prediction accuracy by training models using temperature data obtained from a high-density air temperature observation network established across Chiayi City in Taiwan. The results demonstrate an overall prediction accuracy of 82% in forecasting temperature variations under different urban patterns. This study not only validates the feasibility of integrating LCZ and CNN for accurate heat risk assessment but also identifies potential hotspot areas crucial for formulating urban cooling policies. The findings provide urban planners and designers without a background in meteorology with a more accessible way to understand the impact of planning on microclimates.