Explainable AI for Socio-Spatial Drivers of Urban Heat Exposure Across Contrasting Climatic Conditions in Western Sydney, Australia

Urban heat exposure is an increasing environmental risk in rapidly urbanising regions, with impacts shaped by both land surface characteristics and underlying socioeconomic conditions. In Western Sydney, Cumberland Local Government Area (LGA) experiences elevated summer land surface temperatures (LST), driven by low vegetation cover and high-density built environments. Vulnerable populations, including culturally and linguistically diverse (CALD) communities, may experience disproportionate exposure due to underlying socioeconomic disadvantage. This study develops an explainable artificial intelligence (XAI) framework to examine the relative and temporally stable contributions of biophysical and socioeconomic drivers of urban heat.

The study integrates Landsat 9 Collection 2 Level-2 LST observations from three austral summers (2022–23 La Niña, 2023–24 El Niño, and 2024–25 neutral) with Sentinel-2 derived vegetation and built-up indices (NDVI, NDBI). Socioeconomic variables from the 2021 Australian Bureau of Statistics census at Statistical Area Level 1 (SA1) include median income, rental tenure, elderly population proportion, and English language proficiency. XGBoost regression models are trained separately for each summer to predict mean SA1-level LST using combined predictor sets. Spatial block cross-validation is applied to reduce spatial autocorrelation bias. Model interpretability is assessed using SHapley Additive exPlanations (SHAP), enabling comparison of feature contributions across climatic conditions.

Analysis indicates that urban heat exposure is intramural to both biophysical and socioeconomic characteristics, with vegetation cover and built-up intensity emerging as the primary predictors across all summer periods. SHAP-based interpretation reveals that key socio-spatial drivers maintain broadly consistent importance rankings across contrasting climatic conditions, supporting the temporal stability of these determinants.

This framework provides an interpretable and transferable approach for analysing socio-spatial determinants of urban heat using remote sensing and machine learning. The findings are expected to support equity-focused urban heat adaptation strategies in socioeconomically and culturally diverse urban regions.