Explainable multi-sensor machine learning for global mapping of urban morphology and its implications for urban climate

Accurate global characterization of urban morphology is essential for urban climate modeling but remains fundamentally constrained by spectral ambiguity and the opaque, "black-box" nature of machine learning. This study presents an interpretable, multi-sensor data fusion framework for mapping urban morphological archetypes by coupling Sentinel-1 SAR and Sentinel-2 multispectral imagery across 20 cities spanning six continents. Utilizing a dataset of 60,000 samples, we operationalize a functional aggregation of the Local Climate Zone taxonomy into three distinct archetypes: Dense Organic, Dense Formal, and Open Sprawl. To overcome pixel-level classification limitations, we developed an optimized Extreme Gradient Boosting (XGBoost) model integrated with spatial smoothing to capture neighborhood-scale structural context, achieving a high overall accuracy of 81.35% and a Cohen's Kappa of 0.697. The framework was rigorously validated through hierarchical feature clustering and unsupervised K-means alignment to mitigate a priori classification bias. Quantitative thermal analysis leveraging Landsat 8/9 data confirmed that the mapped archetypes correspond to statistically unique surface thermal regimes (p < 0.001), reinforcing their climatic relevance. Explainable AI (SHAP) analytics successfully decoded the model's internal decision logic, identifying short-wave infrared (SWIR) reflectance and smoothed cross-polarized radar backscatter (VH_smooth) as globally consistent structural fingerprints. Furthermore, regional diagnostics revealed a substantially greater reliance on structural radar metrics within the Global South (Africa: 51.8%; Asia: 40.3%) relative to Oceania (11.5%), with Europe exhibiting the highest overall reliance (56.1%), underscoring the limitations of optical-only approaches. This research provides a transparent, scalable methodology for urban canopy parameterization, bridging the divide between remote sensing analytics and physical urban climate theory to support evidence-based mitigation worldwide.