Urban morphology plays a pivotal role in shaping urban heat islands, by influencing land surface temperature (LST) and heat stress. Numerous studies have extensively explored the relationships between urban morphology and LST, as LST provides a spatially continuous, long-term, and openly available dataset for assessing urban thermal patterns. However, LST is a poor proxy for near-surface air temperature or physiologically relevant human thermal comfort, and relying on it alone may misinform urban planning decisions. Furthermore, although the target temperature variables inherently exhibit spatial autocorrelation, most models fail to simultaneously capture the complex nonlinear interactions and spatial heterogeneity resulting from diverse urban morphologies and environmental conditions across regions. This study aims to reveal the mismatch between LST and heat stress and to investigate the non-stationary relationships between thermal indices and their influencing urban factors by applying Geographically Weighted machine learning (GW-XGBoost) models. Using high-resolution LST and 1-m universal thermal climate index (UTCI) data, combined with 3D urban morphology metrics, landscape indices, biophysical variables, socio-economic factors, we systematically analyse the spatial variations in these associations. SHapley Additive exPlanations (SHAP) enhance interpretability of explainable GeoAI models, revealing localized impacts of key influencing factors. Our findings demonstrate notable discrepancy in spatial patterns of LST and UTCI, along with substantial spatial variations in how 2D and 3D urban morphological factors on these two target temperature metrics across subzones, as revealed by the explainable GW-XGBoost models (out-of-bag R$^2$ = 0.855 (MAE = 0.600°C) for LST and 0.905 (0.172°C) for UTCI). Crucially, important 3D characteristics, especially sky view factor, are substantially underrepresented in LST, underscoring its limited capacity to represent the shading effect that govern actual human thermal exposure. Moreover, SHAP analysis reveals that higher albedo can increase heat stress. These insights provide a robust foundation for targeted heat stress mitigation strategies and adaptive urban planning.