Fine-grained local climate zone classification using graph networks: A building-centric approach

Abstract

Local Climate Zone (LCZ) classification provides a refined framework for urban climate studies, particularly in assessing the urban heat island effect. Traditional LCZ mapping approaches primarily rely on remote sensing data and machine learning methods. Recent advancements have explored multi-source data with deep learning models, such as convolutional neural networks, to enhance LCZ classification accuracy. However, these methods are often limited by inappropriate resolutions of spatial units and overlook the impact of spatial proximity on classification results. To overcome these challenges, this study proposes a building-centric LCZ classification approach that leverages street view imagery, points of interest, and transport stops/stations with graph neural networks, specifically using the GraphSAGE model, to classify built-type LCZ classes. This innovative approach captures local-level urban environmental features while accounting for the influence of spatial relationships between buildings, significantly improving built-type LCZ classification accuracy and refining the spatial scale to the building level. Applied to three cities – Singapore, Berlin, and the urban core area of Sydney – our approach achieves an overall accuracy of 96%, 94%, and 82%, respectively. The models exhibit better performance than traditional remote sensing-based LCZ classification. Importantly, the inclusion of spatial distance shows a significant influence on building-centric LCZ classification accuracy. The building-centric graph network approach offers a more precise tool for LCZ mapping at the building level, prompting urban climate assessment and aiding in the design of climate-resilient cities in the face of rapid urbanization and climate change.