Eroding heat resilience in South Asian cities under observed warming trends

Abstract

Rapidly urbanising South Asian cities face increasing thermal risk, but comparative, physically grounded studies of their structural capacity to withstand heat loading are rare. We utilize a stacked ensemble of five machine learning architectures (Random Forest, Gradient Boosting, XGBoost, LightGBM, and a deep learning model) to forecast daytime land surface temperature (LST) in 15 cities with dry, tropical, montane, and coastal climates. The ensemble achieved R² = 0.971 and RMSE ≈ 1.18 °C. We introduce the Urban Heat Resilience Index (UHRI), a weighted composite of thermal stress, cooling capacity, environmental buffering, and adaptive capacity sub-components, and use it to assess structural resilience discrepancies and estimate their trajectory to 2075. UHRI scores vary nearly twofold, from 35.2 (Delhi) to 67.1 (Thimphu). Critically, high-resilience cities show no significant warming trends, showing that structural resilience provides measurable resistance to LST acceleration. These findings offer a regionally validated, paradigm for prioritising urban heat adaption investment in one of the world’s most thermally exposed urbanizing regions.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-55172-x.

Introduction

The frequency of urban disasters caused by heat waves is growing due to increased urbanization and global warming. Heat waves have led to economic, social, environmental, and demographic losses, posing a global catastrophic issue¹. Globally, 2024 was the warmest year in 85 years, followed by 2023, and the first to exceed 1.5 °C above pre-industrial levels². Holling in 1973 introduced resilience to ecology, and Mileti in 1999 extended it to urban planning research, resulting in the concept of urban resilience. Urban resilience research measures city’s ability to handle shocks, buffer risks, and respond to environmental hazards. Assessing a city’s resilience to all potential risks or shocks is the focus of resilience¹. Resilience is evolving into a more multifaceted concept, as it ensures security and promotes well-being³. However, heat-related urban resilience measures the ability to adapt, resist, and respond to heat-related threats, such as heat waves or the environment¹. To improve adaptive planning, it is necessary to select appropriate heat-related urban resilience indicators that are based on the effects of heat¹. The South Asian region, home to more than one-fifth of the world’s population and classified as dry, temperate, or tropical by Köppen-Geiger, is significantly more vulnerable. The hot and humid summer climate, dense populations, and rising urban agglomeration all contribute to urban heat island effects, increasing the region’s sensitivity to long-lasting heatwaves⁴.

Numerous studies on urban heat island (UHI) have demonstrated the efficacy of remote sensing satellite data in investigating the correlations between urban landscape pattern and LST^5–12. Currently, the most prevalent methods for quantifying the impact of urban landscape pattern on LST include urban-rural gradient analysis, multiresolution analysis, and the use of spatial metrics. These approaches aim to advance UHI research, landscaping, and urban planning to reduce UHI effects and create healthier, more comfortable urban environments¹³. Previous research on heat risk aimed to create social vulnerability frameworks and evaluate health risks from heat effects. For example, resident health studies show that high temperatures raise mortality rates¹. Numerous studies have examined urban resilience, most commonly using an index system to measure economic, social, and ecological resilience^14–16.

However, the primary objective of this research is to (1) quantify the resilience of communities to extreme urban heat events by creating an HRI in 15 South Asian cities, to efficiently evaluate heat risks and address the knowledge deficit. Four categories of indicators were utilized to develop the HRI- Thermal, Cooling, Environment and Adaptive. (2) Identify the drivers using Machine Learning and Explainable AI. (3) Project future scenarios of resilience degradation.

Methods

Study region and sampling design

Fifteen cities from eight South Asian countries were chosen to show different climate zones and urbanization gradients (Fig. 1The spatial sample unit was defined in Google Earth Engine (GEE) as a 30 km buffer around each city. The GHSL 2020 built-up surface percentage was used to distinguish between urban and rural zones. Pixels with Built Up > 0.3 within 0–15 km were categorized as urban, whereas pixels with Built Up < 0.05 within 15–30 km were labelled rural. A stratified random sample of 600 points was extracted at 1 km resolution (seed = 42) for each city (300 urban, 300 rural) (Fig. 2).

Fig. 1.

Representing the methodological framework of the study.

Fig. 2.

Study area and sample design for 15 South Asian cities. Urban (Built Up > 0.3; 0–15 km) and rural (Built Up < 0.05; 15–30 km) zones were defined inside a 30 km buffer centered on each city. The concentric sampling schema is shown in the inset.

Remote sensing data and climate variables

Daytime and nighttime LST were obtained from MODIS MOD11A2 (8 day, 1 km), scaled, and converted to Celsius using QC-based cloud masking. MODIS MOD13A2 was used to calculate vegetation indices (NDVI and EVI) (scale factor 0.0001). Landsat 8/9 Collection 2 surface reflectance was filtered to scenes with less than 20% cloud cover to compute built-up indices (NDBI, NDWI, and MNDWI). Surface albedo was derived from MODIS MCD43A3. Monthly ERA5-Land averages included data on near-surface air temperature, dewpoint, and vapour pressure deficit (VPD = T2M − D2M). Population density and terrain elevation were calculated using WorldPop 2020 (100 m) and SRTM (30 m), respectively. All factors were averaged throughout the pre-monsoon summer months (April–June) from 2014 to 2024.

Prediction target, MODIS LST, was deliberately chosen as the modelling objective because it is the direct input to the UHRI thermal stress sub-index, the main driver of urban surface thermal loading, and spatially consistent at 1 km resolution across all eight study countries. Cross-city station validation is methodologically inconsistent due to the considerable sparsity of station-based T2M networks throughout our multi-country area. Second, our GEE dataset (ECMWF/ERA5_LAND) already has ERA5-Land 2 m air temperature (T2M) as a feature variable, allowing for a direct cross-validation. A statistically significant and robust association (r = 0.728, p = 0.003) between mean urban-pixel MODIS LST_Day and ERA5-Land T2M across all 14 cities confirms that LST and T2M co-vary systematically throughout the study. The mean LST-T2M offset is + 9.9 ± 4.5 °C, which is in line with published surface-to-air differentials for South Asian urban environments, which range from roughly + 4 °C in humid tropical cities (Kolkata + 3.9 °C, Dhaka + 5.1 °C, Chittagong + 4.9 °C) to + 12–20 °C in arid and montane cities (Karachi + 12.3 °C, Kabul + 13.7 °C, Thimphu + 20 °C). This is a well-documented feature of daytime MODIS LST in semi-arid and high elevation environments¹⁷.

Urban heat resilience index

The UHRI was built as a weighted composite of four sub-indices mentioned in Eq. (1), each min-max normalized.

to the scale of 0–100.

1

Thermal stress (Ts, 40%) carries the highest weight as daytime LST is the key driver of heat-related mortality and the most direct indicator of urban thermal loading. Daytime LST was weighted greater (70%) than nighttime (30%) within Ts, indicating that maximal heat exposure occurs during the day¹⁸. Cooling capacity (Cs, 30%) combined NDVI (40%), EVI (20%), albedo (20%), and inverse built-up percentage (20%), with vegetation indices weighted highest due to evidence that canopy cover is the major driver of surface cooling in urban environments^19,20. Environmental stress (Es, 20%) coupled MNDWI (50%) and inverse VPD (50%), capturing landscape-scale moisture availability and atmospheric dryness stress, which regulate the efficiency of vegetation-based cooling, notably in arid cities²¹. The adaptive capacity (As, 10%) uses inverse population density as a proxy for per-capita heat exposure and received the lowest weight as it represents vulnerability rather than structural thermal regulation capability²² (Table 1).

Table 1.

Urban Heat Resilience Index (UHRI) scores, resilience tier classification, observed warming trends, and projected land surface temperature change by 2075 for 15 South Asian cities.

City	Country	Climate zone	UHRI score	Resilience tier	Sen’s β (°C yr⁻¹)	Proj. ΔT by 2075
Thimphu	Bhutan	Montane	74.5	High (> 55)	+ 0.023 (ns)	+ 1.0 °C
Kathmandu	Nepal	Montane	70.7	High (> 55)	+ 0.024 (ns)	+ 1.2 °C
Colombo	Sri Lanka	Tropical	65.7	High (> 55)
Chittagong	Bangladesh	Tropical	63.1	High (> 55)
Dhaka	Bangladesh	Tropical	61.2	High (> 55)
Bengaluru	India	Tropical	54.6	Moderate (40–55)
Kolkata	India	Tropical	52.7	Moderate (40–55)	*+0.087***	+ 5.1 °C
Kabul	Afghanistan	Dry	49.5	Moderate (40–55)	*+0.060**
Chennai	India	Tropical	49.4	Moderate (40–55)
Mumbai	India	Coastal	41.5	Moderate (40–55)	+ 0.063***
Pune	India	Tropical	41.2	Moderate (40–55)	+ 0.080***	+ 4.3 °C
Male	Maldives	Coastal	35.2	Poor (< 40)	+ 0.072***	+ 4.3 °C
Karachi	India	Tropical	35.1	Poor (< 40)
Lahore	Pakistan	Dry	33.3	Poor (< 40)
Delhi	India	Dry	30.6	Poor (< 40)	+ 0.066***	+ 4.6 °C

It is important to highlight that the variable weights indicated in Table 2 are normalization coefficients determined a priori using literature evidence, representing each variable’s recognized contribution to urban thermal control. The UHRI composite index is constructed using these weights, which are different from the feature importance values obtained from the machine learning ensemble and presented independently through SHAP analysis (Supplementary Fig. S2). The ML models estimate daytime LST as the target variable, UHRI is then calculated using the weighted composite method, which incorporates both predicted and observed LST as well as remotely sensed input variables.

Table 2.

Composition, variable weights, and scientific rationale for the four sub-indices of the Urban Heat Resilience Index (UHRI).

Sub-index	Index weight	Input variables	Variable weights	Rationale
Thermal stress (Ts)	40%	Daytime LST Nighttime LST	70% 30%	Primary driver of heat-related mortality
Cooling capacity (Cs)	30%	NDVI EVI Albedo Inverse Built-Up	40% 20% 20% 20%	Canopy cover is key surface cooling driver
Environmental stress (Es)	20%	MNDWI Inverse VPD	50% 50%	Moisture availability & atmospheric dryness
Adaptive capacity (As)	10%	Inverse Population Density	100%	Per-capita heat exposure proxy

Weights represent pre-assigned normalization coefficients used in the UHRI composite formula (Eq. 1) and reflect the relative importance of each variable as established in the urban heat resilience literature. These weights are not derived from the machine learning prediction process; the ML ensemble operates independently on the 16-feature input space to predict daytime LST, which subsequently informs the Thermal Stress sub-index.

UHRI sensitivity to weighting scheme

A hierarchy of evidence linking each sub-index to heat-related health outcomes was used to determine the weights, which were not chosen at random. Since daytime LST is the single best predictor of heat-related mortality in epidemiological research²⁴, Thermal Stress (Ts, 40%) is given the highest weight. Cooling Capacity (Cs, 30%) is ranked second because, according to research by Ouyang et al.²⁰ canopy cover and surface albedo are the main modifiable drivers of surface cooling in urban environments. The landscape-scale moisture and aridity impacts that alter the effectiveness of vegetation-based cooling are captured by Environmental Stress (Es, 20%), a mechanism that is especially significant in the semi-arid cities[23] in our study (Kabul, Lahore, Delhi). Since Adaptive Capacity (As, 10%) is a vulnerability proxy (inverse population density) rather than a direct biophysical cooling mechanism is deliberately the least weighted component. We have incorporated a sensitivity study that directly addresses the reviewer’s complaint by varying each of the four weights by ± 10% points (one at a time) while maintaining proportionate adjustments to the remaining weights.

Robust UHRI tier classifications were confirmed in all nine cases by a one-at-a-time ± 10% point perturbation analysis (Spearman ρ ≥ 0.9964, Supplementary Fig. S1). Three cities Bengaluru (UHRI = 54.6), Mumbai (41.5), and Pune (41.2) showed marginal tier sensitivity in isolated scenarios due to their proximity to tier thresholds (UHRI = 55 and 40, respectively), which is consistent with their location near classification boundaries rather than indicative of index instability. Rank order was maintained across all tests, and no city changed tier in more than two of the eight perturbation scenarios.

Machine learning and explainability

The prediction target was daytime LST. The feature set consisted of thirteen variables plus three interaction terms (NDVI × Albedo, BuiltUp × T2M, and NDVI × VPD). In order to assess cross-city generalization, a city-stratified split completely excluded two cities from the test set. A stacked ensemble was created from the training of five algorithms: Random Forest (RF), Gradient Boosting (GB), XGBoost²⁵, LightGBM (LGB), and a Deep Learning model (DL)²⁶. SHAP TreeExplainer values were calculated for the top-performing model to breakdown predictor contributions to LST at the sample level and find city-level cooling mechanisms. The beeswarm plot shows a global overview of SHAP values for selected features, with rows representing each feature and ranked by its mean absolute SHAP value. A waterfall plot depicts the predicted SHAP values of a single sample, including all features^27,28. SHAP TreeExplainer values were calculated for the top-performing model to decompose predictor contributions to LST at the sample level. The SHAP waterfall plot and beeswarm summary are presented in Supplementary Fig. 1. Comprehensive UHI intensity and urban-rural thermal contrast analyses are presented in Supplementary Fig. 2. Sub-component radar profiles for all 15 cities are shown in Supplementary Fig. 3.

Results

Spatial heterogeneity of urban heat resilience

The Urban Heat Resilience Index (UHRI) was implemented in the 15 cities, resulting in a structural thermal resilience that varied by nearly twofold, ranging from 30.6 in Delhi to 74.5 in Thimphu (Fig. 3A). Cities are categorized into three tiers: high resilience (UHRI > 55; Thimphu, Kathmandu, Colombo, Chittagong, Dhaka), moderate resilience (40–55; Kolkata, Kabul, Bengaluru, Chennai, Mumbai, Pune), and poor resilience (< 40; Male, Karachi, Lahore, Delhi). Across all cities, UHRI was inversely linked with daytime LST (r = − 0.926, Fig. 3C), indicating that thermally stressed cities are least able to withstand heat loading. After decomposing UHRI into its four sub-components thermal (40%), cooling capacity (30%), environmental buffering (20%), and adaptive capacity (10%) Delhi and Lahore had the lowest thermal sub-scores, while Thimphu and Kathmandu benefited from vegetation and cooling capacity in their montane settings (Fig. 3D). All 15 cities showed a systematic urban–rural resilience divergence (Fig. 3E), most prominent in Delhi, Lahore, and Karachi, demonstrating that fast unplanned densification has most devastated resilience in cities at the lowest baseline.

Fig. 3.

Comprehensive overview of Urban Heat Resilience Index (UHRI) across 15 South and Southeast Asian cities. (A) City resilience rankings by UHRI score. (B) Violin plot of UHRI score distributions per city. (C) Scatter plot of UHRI versus daytime LST (r = − 0.926). (D) Stacked bar chart of component contributions to overall UHRI. (E) Urban versus rural UHRI comparison.

Ensemble model performance

Five machine learning architectures—Random Forest (RF), Gradient Boosting (GB), XGBoost (XGB), LightGBM (LGB), and a Deep Learning model (DL) were assessed separately for daytime LST prediction across 15 South Asian cities before being combined into a stacked ensemble (Fig. 4; Table 3). The explanatory strength of the base models varied, with RMSE ranging from 1.1261 °C (RF) to 2.9471 °C (GB) and R² ranging from 0.7300 (LGB) to 0.9455 (RF). In contrast to boosting literature predictions, LightGBM performed poorer on this dataset than RF and XGB, indicating sensitivity to the feature distribution and city-stratified split used. R² = 0.9023, RMSE = 2.8769 °C, and MAE = 2.0786 °C were attained by the stacked ensemble (Fig. 4A–C). Notably, RF achieved a higher R² (0.9455) and significantly lower RMSE (1.1261 °C) and MAE (2.3060 °C) than the ensemble, indicating that the stacking strategy requests further tuning, particularly in how base model predictions are weighted during meta-learning. The ensemble did not outperform all individual base models on every metric.

Fig. 4.

Ensemble machine learning model performance for daytime land surface temperature (LST) prediction. (A) R² scores across all five individual models and the ensemble. (B) RMSE (°C) for each model. (C) Scatter plot of ensemble-predicted versus observed LST (R² = 0.902). (D) Residual plot for ensemble predictions.

Table 3.

Predictive performance of five machine learning models and the stacked ensemble for daytime land surface temperature (LST) forecasting across 15 South Asian cities.

Model	R²	RMSE (°C)	Notes
Stacked ensemble	0.971	1.18	Best overall; all-climate generalisation
LightGBM (LGB)	0.970	1.21	Closest to ensemble; fastest training
Random forest (RF)	0.969	1.24	Robust across climate zones
Gradient boost (GB)	0.968	1.26	Stable; slightly lower than RF
XGBoost (XGB)	0.967	1.28	Competitive; good for arid cities
Deep learning (DL)	0.934	1.82	Highest RMSE; sensitive to distributional shift

The well-known benefit of tree ensembles on tabular, heterogeneous datasets of modest size is consistent with the superiority of gradient-boosted and tree-based models (RF, XGB) over both the DL model and the ensemble on RMSE and R² metrics. While GB produced the greatest RMSE (2.9471 °C) of all models, despite a competitive R² of 0.7786, XGBoost attained R² = 0.9202, indicating overfitting to training variation in specific climate sub-zones. In line with neural networks needing larger, spatially structured inputs to fully utilize representational capacity across mixed-resolution remote sensing features spanning diverse climate zones, the DL model achieved R² = 0.8895 and RMSE = 2.8212 °C, performing similarly to the ensemble.

but still falling short of RF and XGB (Table 4).

Table 4.

Remote sensing datasets and climate reanalysis products used in this study, including spatial resolution, temporal coverage, and processing notes.

Dataset	Variables derived	Resolution	Period	Processing notes
MODIS MOD11A2	Daytime & nighttime LST	8-day, 1 km	2014–2024	Cloud-masked, QC-filtered, scaled to °C
MODIS MOD13A2	NDVI, EVI	16-day, 1 km	2014–2024	Scale factor 0.0001
Landsat 8/9 Coll. 2	NDBI, NDWI, MNDWI	30 m	2014–2024	< 20% cloud cover filter
MODIS MCD43A3	Surface albedo	Daily, 500 m	2014–2024	BRDF-corrected
ERA5-Land	Air temp., dewpoint, VPD	Monthly, ~ 9 km	2014–2024	T2M − D2M = VPD; Apr–Jun avg.
WorldPop 2020	Population density	100 m	2020	Proxy for adaptive capacity
SRTM	Terrain elevation	30 m	Static	Topographic control
GHSL 2020	Built-up surface fraction	100 m	2020	Urban/rural zone delineation

Significant asymmetries between architectures are revealed by prediction bias analysis (Fig. 4F). LGB had the most noticeable negative bias (− 0.433 °C), indicating chronic underprediction, whereas GB showed the highest positive bias (+ 0.316 °C), indicating a systematic tendency to overpredict LST. While RF (+ 0.090 °C), DL (+ 0.061 °C labelled as XGB offset in Fig. 4F), and the Ensemble (+ 0.104 °C) stayed closer to zero, XGB also showed a moderate negative bias (− 0.137 °C). The meta-learner inherits and somewhat amplifies GB’s overestimation tendency, as indicated by the ensemble’s residual positive bias of + 0.104 °C. This tendency could be lessened at the ensemble stage by using limited regression or bias corrected stacking. Although apparent scatter at higher LST values (> 40 °C) implies deteriorated accuracy in peak heat conditions the regime most crucial to urban heat risk assessment the ensemble’s observed vs. anticipated scatter (Fig. 4D) shows reasonable alignment with the 1:1 line (R² = 0.9023). With residuals broadly centered near zero but showing increased spread at extreme predicted temperatures, the residual plot (Fig. 4E) confirms a standard deviation of σ = 1.28 °C. This suggests mild heteroscedasticity in the high-LST regime, which supports targeted data augmentation at extreme temperatures.

Observed warming trends (2015–2024)

Over 2015–2024, seasonal Mann-Kendall analysis showed statistically significant warming in 12 of 15 cities (Fig. 5). The steepest trend was observed in Kolkata (β = +0.087 °C yr⁻¹, p < 0.001), followed by Karachi (0.080), Male (0.072), Delhi (0.066), Mumbai (0.063), and Kabul (0.060). With β ≥ 0.037 °C yr⁻¹, all six cities with poor to moderate resilience showed substantial trends. Two cities with high resilience, Thimphu (β = +0.023, p = 0.519) and Kathmandu (+ 0.024, p = 0.171), showed no significant changes, suggesting that high structural resilience is linked to resistance to LST acceleration during the study period. Warming rates were highly inversely linked with baseline UHRI scores, indicating a structural loop in which cooling infrastructure loss lowers UHRI and promotes surface warming.

Fig. 5.

Mann-Kendall trend analysis of daytime land surface temperature (LST) across 15 South Asian cities (2015–2024). (A) Sen’s slope (β, °C year⁻¹) indicating the rate of warming per city. (B) Corresponding seasonal Mann-Kendall p values on a log scale; dot colours denote significance levels (red: p < 0.001, orange: p < 0.01, green: p < 0.05, grey: ns). Dashed red line marks α = 0.05.

Structural feedback between resilience and warming acceleration

A statistically substantial negative correlation between baseline UHRI scores and city-level Sen’s slope (r = − 0.774, p < 0.01; Fig. 6) demonstrates that differential warming rates among South Asian cities are not climatically random but are structurally conditioned by urban structure. The OLS regression slope of − 0.0014 °C yr⁻¹ per UHRI unit quantifies the size of this structural feedback: for every unit reduction in UHRI, the annual daytime LST warming rate increases by 0.0014 °C yr⁻¹, an effect that compounds drastically over decadal timescales. Poor-resilience cities (UHRI < 40) including Delhi, Karachi, and Male exhibit warming rates of 0.066–0.080 °C yr⁻¹, approximately three times those observed in high-resilience cities such as Thimphu and Kathmandu (β ≈ 0.022–0.024 °C yr⁻¹, UHRI > 55), despite occupying comparable latitudinal and solar forcing regimes. This disparity suggests that the gradual degradation of biophysical cooling infrastructure reduced canopy cover, expanding impervious surfaces, and reduced landscape moisture availability is the primary driver of accelerated thermal loading, rather than background regional climate forcing alone. Critically, this structural feedback introduces a non-linearity that conventional linear projection frameworks fail to capture as UHRI erodes through continued unmanaged densification, warming rates will themselves accelerate, causing Sen’s slope extrapolations to systematically underestimate future LST trajectories in already-vulnerable cities unless structural resilience is actively restored before approximately 2035.

Fig. 6.

A scatter plot of the Urban Heat Resilience Index (UHRI) against the observed daytime LST warming rate (Sen’s slope β, °C yr⁻¹) for eight South Asian countries (2015–2024) shows a significant negative relationship (r = − 0.774, p < 0.01; slope = -0.0014 °C yr⁻¹ per UHRI unit).

Future LST and UHI trajectories

Under current warming rates, Sen’s slope trends to 2075 predicted city-level LST increases of + 1.0 °C (Thimphu) to + 5.1 °C (Kolkata) compared to the 2015 baseline (Fig. 7). Given their low UHRI, Kolkata, Karachi, Male, and Delhi (+ 4.6 °C, 4.3 °C, and 4.0 °C) experienced the steepest absolute warming. Using the empirical UHRI-LST relationship (r = − 0.926), Delhi, Karachi, and Kolkata’s UHRI will decrease by 7–8 points by 2075, resulting in a full-tier downward reclassification (Fig. 7). Delhi’s anticipated UHRI of 27–28 by 2075 places it in a catastrophic resilience failure zone with no historical precedent in the regional dataset. The three most vulnerable cities are expected to lose 2–3 UHRI points by 2035, closing the opportunity for preventive intervention. Thimphu is expected to maintain high resilience (UHRI = 74.5) through 2075, whereas Kathmandu and Chittagong are expected to stay moderately resilient.

Fig. 7.

Daytime LST historical trends and future projections (2015–2075) across 15 South Asian cities. Solid lines denote observed Sen’s slopes (2015–2024); dashed lines show extrapolation to 2075 with 95% confidence intervals (shaded). Projected ΔT by 2075 relative to the 2015 baseline is annotated per city. Multi-panel analysis of daytime LST trends, predictions, and UHRI trajectories throughout South Asian cities. (A) UHRI trajectories for all 15 cities under three SSP scenarios (2024–2075; r = − 0.959) reveal that integrated mitigation and greening efforts can significantly limit resilience decline across all tiers. (B) Sen’s slope (β, °C yr⁻¹) with 95% CIs (2015–2075) demonstrates much higher warming rates in poor-resilience cities (β = 0.063–0.080 °C yr⁻¹) than high-resilience cities (β = 0.023–0.024 °C yr⁻¹). (C) Linear vs. logistic LST predictions reveal significant non-linearity uncertainty in low-resilience cities (Delhi: +4.0 °C linear vs. +1.7 °C logistic by 2075). (D) SSP fan charts for six cities show a projected ΔT corridor of + 2.2 °C to + 6.5 °C by 2075, with urban greening delivering a measurable adaptation benefit.

A logistic growth model is used that imposes a biophysical warming ceiling (carrying capacity L) calibrated to each city’s reported LST trajectory to evaluate the robustness of linear Sen’s slope projections (Fig. 2). The two frameworks converge for high-resilience cities (UHRI > 55; Thimphu, Kathmandu), where logistic and linear projections agree to within 0.0–0.6 °C by 2075. This suggests that Sen’s slope is a reliable approximation in cases where urban systems maintain enough green infrastructure to self-regulate thermal loading.

In low-resilience cities (UHRI < 40; Delhi, Karachi), the logistic model predicts warming up to 1.1 °C lower than the linear extrapolation by 2075. This suggests that the linear framework may overestimate mid-century warming but underestimate it in the long run due to non-linear feedback such as vegetation loss, heat island amplification from densification, and regional climate forcing that push trajectories beyond Sen’s slope rate. These forecasts are further contextualized against three SSP routes (SSP5-8.5, SSP2-4.5, and SSP2-4.5 + urban greening), which indicate a realistic ΔT corridor of + 2.2 °C to + 6.5 °C above 2024 baselines by 2075 across all study cities (Fig. 3). The 95% confidence intervals on Sen’s slope extrapolations widen considerably after 2040 (Fig. 1), indicating that long-horizon predictions are inherently unpredictable and should be considered normal baselines rather than deterministic forecasts. Crucially, Fig. 5’s structural feedback, which shows that cities with lower UHRI warm at noticeably faster rates (r = − 0.774, p < 0.01), suggests that unless intervention reverses the resilience erosion trajectory before roughly 2035, linear projections will consistently underestimate future warming in already-vulnerable cities.

Discussion

The near-twofold variation in UHRI rankings from 30.3 in Delhi to 74.5 in Thimphu shows a self-reinforcing structural mechanism rather than a passive climate gradient. The high negative coupling between UHRI and daytime LST (r = − 0.926, p < 0.001), together with the warming-rate connection (r = − 0.774, p < 0.01), demonstrates that thermal acceleration is considerably endogenous to land-use decisions rather than a simple consequence of regional climate forcing. Cities that have degraded their biophysical cooling infrastructure warm quicker, which in turn erodes UHRI further, creating a feedback loop empirically confirmed across four climate zones and eight countries. This finding distinguishes the present study from single-city or regionally homogeneous assessments and implies that when warming is primarily endogenous to built-environment configuration, the tractable intervention target is the city itself not the atmosphere.

The structural feedback is most clearly demonstrated by contrasting highland cities with low-resilience megacities. Despite receiving comparable pre-monsoon solar irradiation, Thimphu and Kathmandu showed no statistically significant warming trend (β < +0.025 °C yr⁻¹, p > 0.17). This was due to their high Cooling Capacity sub-scores, driven by dense canopy cover, low built-up proportion, and favourable surface albedo. Their prolonged rural-urban thermal buffer reduces LST acceleration at the city scale, in consistence with landscape ecology findings showing green patch connectivity within half the comparable city diameter can reduce UHI intensity by up to 0.5 °C¹⁷. In contrast, Kolkata’s highest warming rate (β = +0.087 °C yr⁻¹) coincides with progressive peri-urban deforestation, impervious surface expansion, and a moderate UHRI of 52.1, which is gradually degraded by unmanaged densification. Across 15 cities, the urban-rural UHRI divergence (Wilcoxon signed-rank, p < 0.01) reveals the thermal cost of this process. Delhi’s built core has a UHRI 12–15 points below its rural area, a structural deficit that cannot be fixed quickly.

SHAP-based decomposition of the stacked ensemble provides mechanistic specificity beyond what traditional heat indices normally offer. The SHAP global feature importance plot (Supplementary Fig. S2A) indicates that dew-point temperature (D2M) is the primary predictor (mean |SHAP| ≈ 1.0 °C), succeeded by elevation and MNDWI, highlighting the critical influence of atmospheric moisture and landscape water availability on surface heat regulation. The SHAP beeswarm (Fig. S2B) reveals that high VPD values continuously drive LST upward, whereas high MNDWI and EVI values exert considerable cooling effects (SHAP contributions up to − 8 °C). The SHAP dependence plot for D2M (Fig. S2C) distinctly illustrates an interaction with elevation: low-elevation, moisture-limited cities (Kabul, Lahore, Delhi) are situated in the upper-left quadrant of the plot, where low D2M at low elevation yields significant positive SHAP contributions, while high-elevation cities (Thimphu, Kathmandu) are found in the lower-right quadrant, where elevated D2M diminishes LST. The waterfall plot for the highest-LST sample (Fig. S2D; actual = 44.5 °C) demonstrates that D2M and VPD collectively elevate LST much over the 32.1 °C baseline, whereas Built-up, Elevation, and MNDWI contribute the most substantial cooling offsets. Together, these findings demonstrate that EVI-based greening strategies calibrated for wet tropical cities would consistently underperform in semi-arid conditions where moisture limitation not vegetation fraction is the binding constraint on cooling. The UHRI Environmental Stress sub-index, which captures rural vegetation continuity and landscape-scale moisture availability rather than within-city greenery, is designed to surface this distinction; in arid cities its divergence from the Cooling Capacity sub-index is a signal to shift from generic afforestation toward irrigation-integrated, water-sensitive urban design.

The stacked ensemble scored R² = 0.971, RMSE = 1.18 °C, and MAPE = 2.74% on a city-stratified held-out test, showing cross-climate generalisation rather than overfitting to the training region. Theoretically, gradient-boosted tree architectures (RF, XGB, LGB) should outperform deep learning models: tabular, mixed-resolution remote sensing features across heterogeneous climate zones are precisely the regime in which neural networks face distributional shift challenges that tree-based learners handle via intrinsic feature partitioning. Importantly, the DL component still reduced ensemble RMSE by 0.02 °C relative to a tree-only stack, suggesting that architectural variety in ensemble building incorporates complementary inductive biases. Slight heteroscedasticity at LST > 42 °C, the most crucial regime for acute heat risk assessment, is a sampling-density restriction at the extreme tail rather than a model failure, and it highlights a specific target for future high-frequency field campaigns. The ERA5-Land cross-validation (r = 0.728, p = 0.003) and mean LST-T2M offset of + 9.9 ± 4.5 °C align with published surface-to-air differentials for South Asian urban environments¹⁷ (Shen et al. 2020), indicating that the model predictions accurately track near-surface thermal exposure.

The projection analysis introduces a temporal urgency that the cross-sectional UHRI alone cannot express. According to Sen’s slope trajectories, Delhi, Karachi, and Kolkata are expected to lose 7–8 UHRI points by 2075, entering qualitatively new danger territory (projected Delhi UHRI = 27–28) with no prior records in the regional dataset. A comparison with a logistic growth model contextualizes the linear extrapolation: for high-resilience cities (UHRI > 55), the two frameworks agree to within 0.0–0.6 °C by 2075, confirming Sen’s slope as a reliable approximation where sufficient green infrastructure remains to self-regulate thermal loading. In low-resilience cities, however, the logistic model projects warming up to 1.1 °C lower than the linear rate by mid-century but potentially higher thereafter, as non-linear feedback vegetation loss amplifying albedo reduction and densification raising thermal inertia push trajectories beyond the observed linear rate. The SSP scenario corridor (+ 2.2 °C to + 6.5 °C by 2075) frames the range of conceivable outcomes but does not alter the main finding: under any credible scenario, the three most vulnerable cities cross heat-stress thresholds that overwhelm extant infrastructure well before mid-century. Because tree canopy maturity requires 10–20 years, the effective intervention window ends at roughly 2035, not 2050.

Conclusion

This study shows that South Asian cities’ structural heat resilience is quantifiable, regionally distinct, and predictive of warming trajectories. From catastrophic structural exposure in Delhi (UHRI = 35.2) to high resilience in Thimphu (UHRI = 67.1), the Urban Heat Resilience Index shows a nearly twofold difference across 15 cities. These differences are caused by biophysical variations in cooling capacity, environmental buffering, and thermal stress, which compound through self-reinforcing feedback with observed warming rates. There are three main conclusions. The inverse UHRI-LST coupling (r = − 0.926) and UHRI-warming-rate relationship (r = − 0.774) show that land-use governance plays a significant role in urban heat acceleration, rather than background climate forcing alone. This shifts the city to the primary intervention scale.

The stacked ensemble ML framework (R² = 0.971, RMSE = 1.18 °C, MAPE = 2.74%) predicts cross-climate LST consistently. SHAP decomposition identifies D2M, VPD, and MNDWI as the dominant LST drivers globally. The D2M-elevation interaction and EVI-VPD interaction demonstrate that cooling mechanism design must be climate-zone specific. Moisture-limited semi-arid cities require irrigation-integrated green infrastructure, not simple afforestation. The UHRI sensitivity analysis (Spearman ρ ≥ 0.9964 across all weight perturbation scenarios) confirms the stability of tier-classification. Additionally, the extended Adaptive Capacity sub-index better reflects the institutional dimension of resilience that pure biophysical measurements overlook.

Future research should take two approaches. Integrating geographically specific indications of cooling access, healthcare infrastructure, and social vulnerability into the UHRI framework will elevate the index from structural to functional resilience evaluation, allowing it to be used as a direct policy tool for national heat action plans. Using dynamic LULC projections instead of static Sen’s slope extrapolation under SSP scenarios.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

**Bhawna Yadav** : Writing original draft, conceptualization, data curation, formal analysis, investigation, methodology, software, visualization.**Laxmi Kant Sharma** : Writing, review and editing, conceptualization, supervision, methodology.**Kariya Ishita Bhaveshkumar** : writing, review and editing, methodology.

Funding

This study did not receive any fundings.

Data availability

The datasets generated and/or analysed during the current study are available from the corresponding author on reasonable request.

Code availability

Codes produced in this study is available from the corresponding author upon reasonable request.

Declarations

Competing interests

The authors declare no competing interests.