Urban local climate zone classification through deep learning using spatio-temporal thermal imagery

Abstract

Rapid urbanization challenges urban micro-climates, strains resources and affects public health. Understanding micro-climate dynamics is key to effective mitigation and sustainable development. Local Climate Zone (LCZ) classification supports climate-resilient planning but is complicated by the diversity and complexity of diverse urban landscapes and the coexistence of varying land uses and materials within small areas. While LCZ classification typically uses multispectral imagery, LiDAR, and land-use data, these sources often miss temporal thermal dynamic patterns. Thermal satellite imagery improves LCZ classification by distinguishing zones with similar structures but differing thermal behavior. This research proposes using deep learning-based multitemporal semantic segmentation to classify urban LCZs based solely on temporal thermal patterns from ECOSTRESS satellite imagery. The methodology is applied in a in a case study around the near coastal cities of Rotterdam and The Hague in The Netherlands and demonstrates how spatial and temporal factors (both diurnal and seasonal) influence the performance of the semantic segmentation model on different LCZ classes. The study shows that a U-Net architecture applied on spatio-temporal thermal imagery effectively classifies urban LCZs, achieving a test accuracy of 0.75. Temporal factors significantly impact model performance, with higher accuracies observed for daytime (0.8) and Spring/Summer imagery (0.78), as these conditions provide clearer thermal separability for distinguishing LCZs. The model achieved its highest test accuracy (0.83) when trained and tested on thermal images with the highest LST values. This suggests that focusing on high-value LST images with sufficient variability enhances classification performance compared to a generalized approach using the full dataset.