- English
Urban heat in rapidly growing megacities is spatially uneven, yet the relative influence of built form, socioeconomic activity and blue-green spatial configuration remains insufficiently resolved in cities of the Global
South. This study examines neighbourhood-scale controls on land surface temperature (LST) in Delhi, India,
using a single clear-sky Landsat 8 thermal observation acquired during the severe pre-monsoon heatwave of 4
June 2024. Landsat-derived LST, Sentinel-2 land-cover information, socioeconomic proxies, built-environment
indicators, terrain data and blue-green landscape metrics describing the amount, fragmentation, shape, cohesion
and contiguity of vegetation and water patches were integrated within a 500 m grid. This scale was selected
to represent neighbourhood-level thermal variation while limiting redundancy from highly detailed urban surfaces.
An Extreme Gradient Boosting (XGBoost) model evaluated through 10-fold cross-validation achieved a
mean R2 of 0.723 and a mean RMSE of 1.79 ◦C. Shapley Additive Explanations (SHAP) and partial dependence
analysis were used to identify dominant predictors, nonlinear response patterns and conditional attribution
across predictor pairs. Global SHAP rankings showed that mean contiguity and patch cohesion had slightly
greater average attribution than impervious surface percentage and substantially greater attribution than
building density. Partial dependence analysis indicated a pronounced but nonlinear decline in fitted LST with
increasing mean contiguity. These results show that the spatial arrangement of blue-green elements contained
explanatory information beyond land-cover fractions alone. The numerical relationships remain specific to the
single-date observation and 500 m analytical scale. Our approach relies on available geospatial datasets and
provides an adaptable diagnostic framework for heat assessment in data-scarce megacities.
- English
