Beyond Land Surface Temperature: Explainable Spatial Machine Learning Reveals Urban Morphology Effects on Human-Centric Heat Stress

Abstract: Heat exposure connects the built environment and public health, directly shaping the livability and sustainability of urban areas. Understanding the spatial heterogeneity of heat exposure and its drivers is vital for climate-adaptive urban planning. However, most planning-oriented studies rely on land surface temperature (LST), and whether LST adequately represents human heat exposure and how it differs from physiologically relevant heat stress remains insufficiently examined. Here, adopting Landsat-retrieved 30-m LST and GPU-accelerated 1-m universal thermal climate index (UTCI) in Singapore, this study establishes a comprehensive "Modeling-Comparing-Assessing" framework to systematically evaluate the spatial and mechanistic discrepancies between the two metrics. We further investigate pronounced non-stationary and threshold-based quantitative relationships of the two metrics with urban factors by employing a novel geographically weighted XGBoost (GW-XGBoost) and generalized additive model (GAM) workflow. Our results demonstrate notable discrepancies in spatial patterns of LST and UTCI, along with substantial spatial heterogeneity in how 2D and 3D urban factors impact these two thermal metrics, as revealed by explainable GW-XGBoost models (global out-of-bag R2 = 0.855 for LST and 0.905 for UTCI, respectively). Crucially, spatially explicit SHAP interprets that sky view factor plays a central role in explaining UTCI variability but exhibits a comparatively marginal independent contribution to LST, indicating that LST inadequately captures shading-driven and radiative processes governing actual human heat stress. Notably, SHAP-GAM analysis indicates that higher albedo is associated with increased UTCI. These novel findings provide evidence for integrating physiologically relevant thermal indices to inform targeted heat risk management and climate-adaptive urban planning.

• The paper demonstrates that LST is an inadequate proxy for human heat exposure by revealing significant discrepancies with UTCI using explainable spatial ML.
• It employs a GW-XGBoost framework with SHAP to expose non-linear, threshold-dependent impacts of factors like SVF, albedo, and canopy density.
• The study provides actionable insights for urban heat mitigation by advocating dual-index frameworks and spatially tailored intervention strategies.

Explainable Spatial Machine Learning for Urban Heat Stress: Critique of LST Proxies and Urban Morphology Effects

Introduction

The persistent usage of satellite-derived Land Surface Temperature (LST) as a proxy for urban heat exposure fails to address the multidimensional drivers of human-centric thermal stress. The paper "Beyond Land Surface Temperature: Explainable Spatial Machine Learning Reveals Urban Morphology Effects on Human-Centric Heat Stress" (2604.22433) systematically interrogates the spatial and mechanistic discrepancies between radiometric surface temperature (LST) and the Universal Thermal Climate Index (UTCI) in the spatially heterogeneous context of Singapore, integrating interpretable geographically weighted machine learning and high-resolution urban morphology datasets.

Methodological Innovation

The authors implement a "Modeling-Comparing-Assessing" analytical framework that simultaneously leverages satellite LST (30 m), GPU-accelerated UTCI (1 m, via SOLWEIG), and an extensive suite of 2D/3D urban form and environmental covariates. The methodological core is a geographically weighted XGBoost (GW-XGBoost) approach, designed to address limitations of conventional statistical and ML models—namely, their inability to capture non-linearity and spatial non-stationarity in urban climate processes.

Explainability is addressed through the integration of SHAP (Shapley Additive exPlanations), which enables spatially explicit attribution of variable influence across the study area. Model performance is robustly validated via nested cross-validation for global XGBoost and leave-one-out spatial CV and pseudo-out-of-bag validation for GW-XGBoost, with mean global $R^2$ reaching 0.905 for UTCI and 0.855 for LST.

Empirical Results: Divergence of Surface and Human-Centric Heat Metrics

Spatial and Statistical Discrepancies

The study demonstrates that spatial patterns of LST and UTCI are notably divergent. LST exhibits higher maximum values (>50°C) and greater spatial range, while UTCI is bounded, reflecting the air temperature, radiative flux, wind, and humidity modulations experienced by pedestrians. High LST regions do not always correspond to high-UTCI zones: LST overestimates human heat exposure in dense, shaded environments and underestimates pedestrian risk in unvegetated, exposed spaces.

The relationship between LST and UTCI is non-linear with domain-specific saturations and even reversals (notably over water bodies, where LST is low but UTCI is high due to humidity and unshaded solar load).

Variable Importance and Mechanisms

SHAP-based global rankings reveal a clear separation in the environmental drivers of LST versus UTCI. LST is primarily governed by surface indicators—wetness index (WET), normalized difference built-up index (NDBI), NDVI, and building density—demonstrating little sensitivity to volumetric 3D morphology. In contrast, UTCI is dominantly modulated by 3D radiative and morphological metrics, most notably the Sky View Factor (SVF), elevation (DEM), albedo, and canopy structure (height and density).

A strong, spatially explicit claim is supported: SVF is the central determinant of UTCI variability, but exerts little independent influence on LST. This result directly contradicts the common use of LST as a planning metric for urban thermal risk, especially in high-density settings dominated by complex shading and radiative geometry.

Non-linear SHAP-GAM analyses expose further threshold effects:

• SVF exhibits a clear cooling-to-warming transition at ~0.51, below which increasing enclosure/shade reduces UTCI, with monotonic warming above this threshold.
• Albedo, contrary to expectations, is associated with increased UTCI once SVF exceeds 0.15–0.17, especially in open zones. This warming is attributed to additional multiple reflections in unshaded environments, intensifying the mean radiant temperature at pedestrian height.
• Tree canopy effects are only significant above a density threshold near 0.81, with dense, contiguous canopy generating substantial UTCI reductions, whereas sparse greening provides minimal mitigation.

Implications for Urban Heat Assessment and Planning

The findings challenge the prevailing methodological paradigm in urban climate adaptation planning, which often defaults to LST due to data convenience. The evidence that LST is an inadequate and potentially misleading proxy for human heat stress in heterogeneous, high-density cities is quantitative and mechanistically explicit. Optimizing urban heat mitigation requires prioritization of morphological metrics (SVF, canopy continuity), spatial configuration (patch density), and cautious deployment of reflective materials, as their efficacy is strongly context-dependent.

Practically, this evidence compels a shift toward:

• Dual-index frameworks where UTCI or similar biophysical indices guide climate adaptation in conjunction with LST, rather than as secondary diagnostics.
• Spatially differentiated interventions: For example, maximizing tree canopy continuity is critical but only above certain density thresholds; shading strategies must be prioritized for high-SVF, open spaces; cooling via reflective pavements is effective only in enclosed/shaded scenarios.

Theoretically, the work advances urban climate modeling by establishing a robust explainable GeoAI workflow that reconciles spatial autocorrelation, heterogeneity, and non-stationarity, providing a blueprint for reproducible, mechanism-oriented urban environmental analytics.

Prospects for Future Research

Given context sensitivity, further studies are necessary to evaluate the transferability of the observed thresholds (SVF, canopy, albedo) across climates with different radiative regimes. The computational cost of physically based UTCI simulation at hyperlocal scales highlights a strong opportunity for hybrid physical-ML models—such as Physics-Informed Neural Networks—to deliver scalable, path-dependent heat stress mapping and scenario modeling [shaeri2025multimodal, ren2025physics].

Additionally, integrating multi-source high-frequency urban observations—including dense sensor networks, mobile campaigns, and advanced meteorological downscaling—remains a key research avenue to better constrain heat exposure at the scale of vulnerable populations.

Conclusion

The paper robustly establishes that LST-derived indices systematically misestimate human-centered heat exposure in tropical cities with complex urban morphologies. Explainable geographically weighted machine learning—applied to high-resolution UTCI, morphology, and environmental data—quantifies non-linear, threshold-dependent, and spatially heterogeneous drivers of urban heat stress. The inadequacy of LST as a standalone proxy is mechanistically substantiated, requiring planners and climate resilience practitioners to adopt biophysically grounded, spatially explicit, and explainable modeling frameworks for effective adaptation. The framework presented offers a technically rigorous foundation for the next generation of spatial urban climate analytics (2604.22433).