Machine learning modeling of vegetation and limited two dimensional urban morphology effects on land surface temperature in Osaka using open data.

Urban heat islands (UHI) significantly elevate land surface temperatures (LST) in high-density subtropical cities like Osaka, Japan, exacerbating energy demand, health risks, and climate vulnerability. This study investigates LST drivers using freely accessible Landsat 9 data (August 27, 2024) and OpenStreetMap (OSM) building footprints at 100 m resolution. Due to the absence of reliable 3D height data, we focus on vegetation indicators (mean NDVI and vegetation fraction) and basic 2D morphology (building coverage ratio [BCR] and building area density ratio [BADR], assuming uniform 10 m height) as a pragmatic open-data baseline. Vegetation metrics show weak positive associations with LST (Pearson r = 0.173 for NDVI_mean, 0.114 for VegFrac), while 2D morphology exhibits negligible links (r ≈ 0.008). These results highlight the limited explanatory power of planimetric indicators alone in humid subtropical settings. Machine learning models (Random Forest [RF], XGBoost [XGB], Artificial Neural Network [ANN]) substantially outperformed multiple linear regression (R² = 0.045), with XGB achieving the highest performance (R² = 0.233, RMSE = 29.6 °C). NDVI_mean dominated feature importance (55.3%). Spatial predictions identified LST hotspots in central districts (high BCR, low vegetation), where partial dependence analysis suggests an indicative marginal LST reduction of ≈ 1.0-1.5 °C associated with a 10% point increase in VegFrac (model-derived statistical association, not causal; subject to considerable uncertainty due to the modest R² and excluded confounders). Results emphasize the need for multi-variable frameworks incorporating 3D morphology (e.g., sky view factor, height variation), landscape patterns, and meteorological factors to enhance predictive accuracy and inform targeted greening, ventilation corridors, and cool materials in Osaka's urban planning. As a replicable, low-cost open-data baseline, this study offers practical insights for resource-constrained subtropical cities, contributing to Sustainable Development Goals (SDGs) 11 and 13.