Seasonal Nonlinearity and Spatial Non-Stationarity of Urban Thermal Comfort Drivers: Implications for Sustainable Urban Planning in Tianjin, China

Abstract

Rapid urbanization and intensifying thermal stress pose challenges to urban livability, climate resilience, and sustainable development. However, studies overlook the seasonal nonlinearity and spatial non-stationarity of urban thermal comfort and its driving factors, constraining seasonally differentiated, climate-adaptive planning strategies. Taking central Tianjin, China, as a case study, this research integrates a geographically weighted XGBoost model with shapley additive explanations to disentangle the seasonal and spatial differentiation, nonlinear relationships, and conditional interaction patterns of the Universal Thermal Climate Index (UTCI). Results indicate: (1) UTCI exhibits seasonal differences in spatial thermal gradients, with relatively high values concentrated in central and northern-central areas; (2) driver contributions shift across seasons, with evapotranspiration leading in summer, digital elevation model in autumn, and building density in winter, while landscape composition contributes most strongly in winter; (3) candidate nonlinear breakpoints include winter sky view factor (0.683; 95% interval 0.662–0.752) and water patch density (15.6 patches/100 ha; 13–17.8); (4) conditional interactions vary seasonally, involving environmental–terrain combinations in summer and building-morphology–landscape combinations in winter. These findings reveal season-specific thermal comfort associations, providing quantitative evidence for translating nonlinear relationships and spatial heterogeneity into climate-adaptive urban design. This supports targeted thermal-environment interventions for urban resilience and sustainability.

1. Introduction

Global warming and rapid urbanization have intensified the Urban Heat Island (UHI) effect in high-density cities [1,2], exacerbating outdoor thermal comfort issues and making the precise analysis of thermal comfort a critical bridge for mitigating public health risks in sustainable urban development [3]. While traditional field measurement studies provide localized data [4], faced with the holistic complexity of the urban environment, research is increasingly shifting towards large-scale spatial analysis. To accurately evaluate the urban thermal environment, indicator systems have evolved from single physical parameters, such as Land Surface Temperature (LST), to comprehensive physiological indices like the Universal Thermal Climate Index (UTCI).

Existing outdoor thermal comfort studies cover a variety of assessment methods from local to macro-urban scales, yet each method possesses specific applicability limitations regarding data acquisition and indicator representation. At the local scale, existing studies typically obtain high-precision assessments at specific time points through in situ measurements, rendering cross-temporal and seasonal assessments challenging. At the urban scale, most studies primarily utilize satellite inversion or weather station data, relying on LST or air temperature (AT) as proxy indicators for the outdoor thermal environment [5,6]. Because thermal stress is governed by more than temperature alone, reliance on single parameters such as LST may lead to incomplete or biased assessments of outdoor thermal conditions [7,8]. To overcome this, the Universal Thermal Climate Index (UTCI) is widely adopted [9,10], as it comprehensively integrates air temperature, humidity, wind speed, and radiation into a single biometeorological model to measure dynamic human physiological responses across diverse climate zones [11,12].

The seasonal and spatial differentiation of urban outdoor thermal comfort is associated with multidimensional influencing factors, including topographic features, landscape composition and pattern, environmental conditions, building morphology, and socio-economic characteristics [13,14]. Topographic features like elevation and Aspect alter local airflow and radiation budgets [15,16], while landscape elements and surface attributes, such as water bodies, vegetation, and high-albedo materials, regulate temperatures through latent heat fluxes and reduced heat storage [17,18]. Furthermore, three-dimensional building morphology alters street canyon radiation dissipation and wind attenuation [19,20], while socio-economic characteristics may be associated with differences in human activity intensity and thermal exposure [21,22]. Collectively, these multidimensional factors are associated with complex urban microclimatic responses [23].

Although UTCI establishes theoretical global thermal-stress categories, research indicates that the actual neutral range of thermal comfort can be influenced by seasonal climate memory and psychological expectations [24]. Empirical studies in the Mediterranean and Central Europe have reported seasonal asymmetry and hysteresis in thermal comfort responses, indicating that residents may provide different thermal evaluations across seasons even under comparable UTCI conditions [25,26].

Existing studies are concentrated at either macro-urban/regional scales or micro-scale settings, leaving a clear gap at intermediate spatial scales. At macro-urban and regional scales, studies mainly reveal spatial patterns of thermal stress under broad climate backgrounds and their statistical associations with land cover [27,28]. UTCI is primarily applied to assess climate change trends, extreme heatwave risks, and biometeorological impacts of the UHI effect. Relevant works rely on meteorological station data or mesoscale weather models to map the geographic distribution of thermal stress, aiming to establish quantitative relationships between thermal comfort and macro-climate zones or land cover types [29,30]. While providing effective macro-warnings, the coarse grid resolution of such studies fails to capture specific disturbances of buildings on the wind-thermal environment, making them unable to offer direct guidance for refined urban design. Conversely, at the micro-human-centric scale, research focuses on verifying the improvement effects of specific microclimate adaptive design strategies. Here, UTCI serves as a core dependent variable to test micro-renewal strategies, shrinking the scope to single points or lines. The extensive literature discusses the instantaneous effects of street tree shading, permeable pavements, or street canyon Aspect ratios [31,32]. While highly precise, these studies often have limited ability to capture the spatial continuity and interactions of climatic processes across groups of buildings. Consequently, a gap remains in characterizing within-city spatial variation while integrating finer-resolution explanatory information. Existing macro models have limited spatial detail, while micro-scale measurements have limited areal coverage. The present study addresses this gap through intra-urban gridded analysis on the native nominal 1 km UTCI grid, with finer-resolution explanatory variables aggregated to the native UTCI cells before modeling.

Methodologically, characterizing these within-city spatial associations requires handling complex spatial non-stationarity and nonlinear threshold responses [33,34]. To better address these spatial and methodological gaps [35], this study adopts the Geographically Weighted XGBoost (GW-XGBoost) model to capture spatial heterogeneity and nonlinear relationships among driving factors [36,37]. Furthermore, to overcome the inherent “black-box” opacity of complex machine learning models, GW-XGBoost is paired with the SHAP (SHapley Additive exPlanations) framework. This combined approach provides feature attribution [38], supports the quantitative identification of candidate nonlinear breakpoints [39], and characterizes conditional interaction patterns among built-environment factors [40].

Existing studies rarely systematically explore the seasonal spatial differentiation of UTCI within cities, particularly lacking quantitative analysis of spatial non-stationarity and nonlinear threshold effects. To bridge this gap, taking the central urban area of Tianjin as a case study, this research utilizes the GW-XGBoost model and SHAP interpretability analysis to systematically address the following scientific questions: Under different seasonal climatic backgrounds, what seasonally varying spatial patterns do multidimensional influencing factors exhibit in their associations with thermal comfort? How can the limitations of single macro-surface temperature representation be overcome to characterize the spatial heterogeneity of UTCI at its native nominal 1 km grid while integrating finer-resolution explanatory information? How can candidate nonlinear breakpoints, marginal relationships, and interaction patterns be quantitatively characterized? This study aims to provide a quantitative basis for seasonally differentiated and spatially targeted urban thermal-environment interventions by quantifying candidate nonlinear breakpoints, spatially varying contributions of driving factors, and conditional interaction patterns among key factors. By translating these findings into climate-adaptive planning guidance, the study seeks to support more resilient, livable, and sustainable urban development under increasing climate and urbanization pressures. Such translation requires rigorous spatial validation [41] and careful consideration of the transferability of planning implications [42].

2. Data and Methods

2.1. Study Area

This study focuses on the central urban area of Tianjin, geographically defined by the Outer Ring Road. As a major coastal city in northern China, the study area covers approximately 374 km² (Figure 1). Located in the northeastern North China Plain, the region is characterized by flat terrain and high urbanization, exhibiting a monocentric, high-density spatial structure expanding concentrically along the Hai River axis. As a typical temperate semi-humid continental monsoon climate zone (Köppen classification: Dwa), Tianjin is controlled by the East Asian monsoon circulation and exhibits pronounced seasonal climatic contrasts. In summer, under the influence of the subtropical high, Tianjin is typically hot, rainy, and humid, with frequent heatwave conditions. In winter, the Mongolian-Siberian high produces cold and dry conditions with prevailing northwesterly winds. The pronounced contrast between hot-humid summers and cold-dry winters provides a suitable climatic setting for examining seasonal differences in urban thermal comfort.

Figure 1. Location of study area. (a) location of Tianjin in China; (b) location of the study area in Tianjin; (c) land cover of the study area.

2.2. Data and Preprocessing

The primary target variable for this study is thermal comfort, derived from the GloUTCI-M dataset (Table 1). To mitigate the noise from single-month meteorological fluctuations and characterize seasonal thermal regimes, we performed seasonal aggregation on monthly gridded estimates for March 2021–February 2022. By calculating seasonal averages, we generated UTCI raster images that reflect seasonal thermal-environment conditions. Spring, summer and autumn comprise March–May, June–August and September–November 2021, respectively; winter comprises December 2021–February 2022. The product represents its daytime sampling convention (approximately 11:00–14:00 local time), rather than all-day exposure. UTCI retains its original geographic grid (nominally 1 km).

Table 1. Data sources.

Data Type	Data Feature	Data Usage	Data Source
GloUTCI-M	Raster, 1 km, seasonal, 2021–2022	Thermal comfort assessment	Zenodo (https://doi.org/10.5281/zenodo.8310513, accessed on 20 September 2025)
Landsat-8 OLI/TIRS reflectance	Raster, 30 m, seasonal, 2021	Calculate NDVI, albedo, and water proportion	USGS EarthExplorer (https://earthexplorer.usgs.gov/, accessed on 20 September 2025)
EcoVision Land Cover	Raster, 5 m, 2022	Land cover classification and landscape metrics extraction	Zenodo (https://zenodo.org/records/14921585, accessed on 20 September 2025)
ASTER GDEM V2	Raster, 30 m, 2021	Topographic feature derivation (Aspect, TWI, TPI)	NASA Earthdata (https://earthdata.nasa.gov/, accessed on 20 September 2025)
Building data	Vector, building geometry and height, 2021	Calculate Building Height and Floor Area Ratio	Gaode Map (https://www.amap.com/, accessed on 20 September 2025)
WorldPop	Raster, stratified by age group, 100 m, 2021	Calculate urban Population Density	WorldPop (https://www.worldpop.org/, accessed on 20 September 2025)

To construct the set of explanatory variables, we integrated multi-source remote sensing and geographic data. First, regarding surface biophysical attributes, Landsat-8 OLI/TIRS images were acquired via the USGS EarthExplorer platform (United States Geological Survey, Reston, VA, USA). Following radiometric calibration and atmospheric correction, cloud-free composite images were generated to retrieve NDVI, surface Albedo, and water proportion. Second, to capture fine-scale landscape patterns, we used the supplied EcoVision-derived land cover map. These data were aggregated to extract the Landscape Shape Index (LSI), Aggregation Index (AI), and Patch Density (PD) for various land use types. Specifically, to clarify the EcoVision classifications within the high-intensity urban core, variables such as Shrub Patch Density (PD_Shrub) and the Landscape Shape Index of unbuilt bare surfaces (LSI_Soil) were extracted, retaining descriptors derived from the supplied polygons for subsequent area-weighted aggregation. Finally, to account for three-dimensional urban morphology and topography, vector building data were rasterized to compute Building Height (BH), BD, and Floor Area Ratio (FAR), while topographic features, including Slope, Aspect, and Topographic Wetness Index (TWI), were derived from ASTER GDEM V2 data.

To ensure spatial consistency, all explanatory variables were reprojected to WGS 1984 UTM Zone 50N and summarized to the native UTCI grid. UTCI was not downscaled and retained its native nominal 1 km spatial support. Landsat- and terrain-derived variables were processed at their original 30 m scale and subsequently aggregated to the native UTCI cells using area-weighted procedures. NDVI, Albedo, and DEM were aggregated by intersection area to native UTCI cells; other predictors were joined to the supplied polygons by New_ID and aggregated by polygon–pixel intersection area. Aspect used a circular mean. Each native UTCI pixel contributes one response observation. The supplied polygons served only as intermediate source geometries and were not treated as independent thermal analysis units. Within each training fold, missing values were treated first; highly correlated predictors (|r| > 0.8) were then removed, followed by iterative VIF screening until all remaining VIF values were below 10. These procedures were used to construct the Multi-source Integrated Feature dataset for UTCI modeling (MIF-UTCI).

During data-quality screening, 80 native UTCI cells showed an invariant value of exactly 0 °C across all 12 monthly layers and were therefore excluded as potentially non-physical observations. Legitimate negative values and isolated near-zero observations were retained. Modeling additionally required at least 50% coverage by the supplied polygons and complete optical/DEM support (coverage fraction at least 0.99995), yielding 422 samples in summer, autumn, and spring and 418 in winter. The subset with at least 75% coverage by the supplied polygons is reported as a sensitivity check using the same out-of-fold predictions. These procedures improve spatial-support consistency, while residual uncertainty associated with the Modifiable Areal Unit Problem (MAUP) is explicitly acknowledged.

2.3. Variables

To comprehensively characterize the multidimensional associations with UTCI, this study used the MIF-UTCI dataset spanning six major dimensions (Table 2). First, terrain-related factors characterize the physical background of the urban thermal environment. DEM and Slope describe elevation-related surface variation, while DEM-derived Aspect represents local surface orientation at the DEM support scale and may influence solar exposure and airflow conditions. Given the generally flat and highly urbanized context of central Tianjin, Aspect may partly reflect composite surface morphology, but it should not be interpreted as a direct measure of street or building orientation. Hydrological indicators such as TWI characterize relative topographic moisture conditions. Building on this foundation, landscape composition and environment factors jointly characterize the thermophysical attributes of land cover. Pland, PD, and NDVI were used to characterize land cover composition and vegetation conditions, while Albedo and ET characterize surface radiative and moisture-related conditions. Under the operational grouping adopted in this study, class-specific PD, Pland, AI, and LSI were grouped as landscape composition descriptors, whereas landscape-level Cohesion, LPI, SHDI, and Split were grouped as landscape pattern metrics. These variables jointly characterize land cover composition, aggregation, shape complexity, connectivity, diversity, and fragmentation. Finally, environmental, building-morphology, and socio-economic variables were included to characterize conditions in high-density built-up areas. SVF, classified here as an environment variable, characterizes sky exposure, whereas FAR, BD, and BH characterize building morphology and urban compactness. Meanwhile, Population Density (POPD) characterizes demographic distribution, human settlement intensity, and potential thermal exposure vulnerability.

Table 2. Explanatory variables used in this study.

Category	Variable	Formula	Description
Terrain	Digital elevation model (DEM)	$\mathrm{DEM}={\mathrm{Z}}_{\mathrm{xy}}$	Digital representation of ground surface elevation.
	Aspect	$\mathrm{Aspect}=\arctan ({\mathrm{f}}_{\mathrm{y}}/{\mathrm{f}}_{\mathrm{x}})$	DEM-derived local surface orientation associated with differences in solar exposure and airflow conditions.
	Topographic Wetness Index (TWI)	$\mathrm{TWI}=\ln (\mathsf{\alpha }/\tan (\mathsf{\beta }))$	Characterizes relative topographic wetness based on upslope contributing area and local slope.
	Slope	$\mathrm{Slope}=\arctan (\sqrt{{\mathrm{p}}^2+{\mathrm{q}}^2})\times \left(180/\mathsf{\pi }\right)$	Rate of change in elevation, affecting runoff and land use.
	Topographic position index (TPI)	$\mathrm{TPI}={\mathrm{Z}}_0-\overline{\mathrm{Z}}$	Measures topographic position relative to surroundings.
	Stream power index (SPI)	$\mathrm{SPI}={\mathrm{A}}_{\mathrm{s}}\times \tan (\mathsf{\beta })$	Estimates erosive power of overland flow.
	Canopy height (CH)	$\mathrm{CH}=\mathrm{DSM}-\mathrm{DTM}$	Vegetation canopy height; for ecology/biomass studies.
Landscape composition	Patch Density (PD)	$\mathrm{PD}={\displaystyle \frac{\mathrm{N}}{\mathrm{A}}}$	Patches per unit area; reflects subdivision.
	Landscape Shape Index (LSI)	$\mathrm{LSI}={\displaystyle \frac{0.25\mathrm{E}}{\sqrt{\mathrm{A}}}}$	Patch shape complexity; indicates fragmentation.
	Aggregation Index (AI)	$\mathrm{AI}=\left[{\displaystyle \frac{{\mathrm{g}}_{\mathrm{ii}}}{\max \left({\mathrm{g}}_{\mathrm{ii}}\right)}}\right]\times 100$	Clustering degree of same-type land cover patches.
	Percentage of landscape (Pland)	$\mathrm{Pland}=\left({\displaystyle \frac{\sum {\mathrm{a}}_{\mathrm{ij}}}{\mathrm{A}}}\right)\times 100$	Proportion of landscape occupied by a specific class.
Landscape pattern	Cohesion	$\mathrm{Cohesion}=\left[1-{\displaystyle \frac{\sum {\mathrm{p}}_{\mathrm{ij}}}{\sum ({\mathrm{p}}_{\mathrm{ij}}\sqrt{{\mathrm{a}}_{\mathrm{ij}}})}}\right]\cdot {\left[1-{\displaystyle \frac{1}{\sqrt{\mathrm{Z}}}}\right]}^{-1}\times 100$	Physical connectedness between same-type patches.
	Largest Patch Index (LPI)	$\mathrm{LPI}={\displaystyle \frac{\max ({\mathrm{a}}_{\mathrm{ij}})}{\mathrm{A}}}\times 100$	Percentage of landscape covered by the largest patch.
	Shannon’s Diversity Index (SHDI)	$\mathrm{SHDI}=-\sum ({\mathrm{P}}_{\mathrm{i}}\ln ({\mathrm{P}}_{\mathrm{i}}))$	Variety and evenness of land cover types.
	Split	$\mathrm{Split}={\displaystyle \frac{{\mathrm{A}}^2}{\sum {\mathrm{a}}_{\mathrm{ij}}^2}}$	Fragmentation degree; higher values mean more dispersion.
Environment	Sky View Factor (SVF)	$\mathrm{SVF}={\displaystyle \frac{1}{2\mathsf{\pi }}}{{\displaystyle \int }}_0^{2\mathsf{\pi }}{{\displaystyle \int }}_0^{\mathsf{\pi }/2}\cos \mathsf{\theta }\sin \mathsf{\theta }\mathrm{d}\mathsf{\theta }\mathrm{d}\mathsf{\varphi }$	Visible sky fraction; characterizes sky exposure and radiative openness.
	Normalized difference vegetation index (NDVI)	$\mathrm{NDVI}={\displaystyle \frac{\mathrm{NIR}-\mathrm{Red}}{\mathrm{NIR}+\mathrm{Red}}}$	Spectral index indicating vegetation greenness and density.
	Albedo	$\mathrm{Albedo}={\displaystyle \frac{{\mathrm{SW}}_{\mathrm{out}}}{{\mathrm{SW}}_{\mathrm{in}}}}$	Surface solar reflectance; influences energy balance.
	Evapotranspiration (ET)	$\mathrm{ET}={\displaystyle \frac{{\mathrm{R}}_{\mathrm{n}}-\mathrm{G}}{\mathsf{\lambda }}}$	Combined water loss from soil and plants.
Building morphology	Floor Area Ratio (FAR)	$\mathrm{FAR}={\displaystyle \frac{\sum {\mathrm{A}}_{\mathrm{floor}}}{{\mathrm{A}}_{\mathrm{site}}}}$	Ratio of total floor area to land area.
	Building Density (BD)	$\mathrm{BD}={\displaystyle \frac{{\mathrm{A}}_{\mathrm{footprint}}}{{\mathrm{A}}_{\mathrm{site}}}}\times 100$	Percentage of land covered by buildings; reflects compactness.
	Building Height (BH)	$\mathrm{BH}={\displaystyle \frac{\sum {\mathrm{H}}_{\mathrm{bldg}}}{\mathrm{N}}}$	Average vertical dimension of buildings; reflects vertical urban form.
Socio- economic	Population Density (POPD)	$\mathrm{POPD}={\displaystyle \frac{{\mathrm{P}}_{\mathrm{total}}}{{\mathrm{A}}_{\mathrm{area}}}}$	Number of inhabitants per unit area, indicating human settlement intensity.

The candidate predictor set comprised NDVI, Albedo, ET, DEM, Aspect, Slope, SPI, TPI, TWI, CH, SVF, BH, BD, FAR, AI, Cohesion, LPI, LSI, SHDI, Split, class-specific AI/LSI/PD/Pland for shrub, tree, bare land, and water, and POPD_Elder/POPD_Child. LSI_Soil represents the shape complexity of the bare-land class, while PD_Shrub denotes Shrub Patch Density in the supplied processed land cover classification. PD is patches per 100 ha (equivalent numerically to patches/km²), whereas Pland is percent. The native-grid landscape predictors are area-weighted means of descriptors derived from the supplied polygons, not metrics recomputed for a new native-sized patch mosaic. Undefined polygon-derived values are excluded from the area-weighted numerator and denominator; missing aggregate values are imputed within training folds. TWI and ET were retained in the numerical scales provided in the source dataset.

2.4. Methods

This study established a four-stage technical framework comprising data preparation, indicator construction, model training, and interpretability analysis (Figure 2). Step 1 involved multi-source data acquisition and preprocessing. Step 2 established a multi-dimensional indicator system covering terrain, landscape, and socio-economic elements. Step 3 focused on constructing the GW-XGBoost model using geographically separated validation and comparing its predictive performance with the global XGBoost baseline. Step 4 introduced the SHAP framework to analyze feature contributions and relationships from multiple dimensions, including spatial distribution, nonlinear responses, and conditional interaction patterns. All computational analyses were performed in Python 3.12 using pandas 3.0.5, Rasterio 1.5.1, scikit-learn 1.9.1, XGBoost 3.4.1, and SHAP 0.52.0.

Figure 2. Research framework. The sub-figures shown in Step 4 are merely illustrative examples of the analytical outputs, which are fully detailed in subsequent result sections (Figures 3–7).

2.4.1. Regression Model

Although tree-ensemble algorithms like XGBoost are inherently resilient to multicollinearity, feature screening was conducted within each training fold to reduce redundant predictors and improve the stability and interpretability of SHAP attribution. The geographically weighted component supplies distance-based weights to local tree ensembles; it is not a separately estimated linear GWR model. Feature screening and imputation were performed independently within each training fold following the procedures described in Section 2.2, without using held-out outcomes or predictor statistics. The full 38-candidate list and retained variables for each season and fold are provided in Table A1.

Model performance was assessed using five-fold geographically separated validation [41]. Native-cell centroids were divided into five contiguous easting-quantile bands. Each band was held out in turn, and training centroids within 1000 m of any test centroid were excluded. Consequently, no subdivision of a native thermal pixel can enter another fold. Training sample counts after buffering ranged from 296 to 323, and test counts from 83 to 85. Coordinates in UTM zone 50N were used for distances.

To capture the spatial non-stationarity of the urban thermal environment, this study constructed a Geographically Weighted Extreme Gradient Boosting (GW-XGBoost) model. The model combines geographically weighted local modeling with the capacity of XGBoost to represent high-dimensional nonlinear relationships [34]. Specifically, GW-XGBoost captures spatial non-stationarity by introducing a geographic weight based on a spatial kernel function into the loss function. For a local model at spatial location $i\left(u_i,v_i\right)$, its optimized objective function $Ob{j}^{\left(t\right)}$ can be expressed as

$$Ob{j}^{\left(t\right)}={{\displaystyle \sum }}_{j=1}^n{w}_{ij}\left[l\left(y_j,{\widehat{y}}_j^{\left(t-1\right)}+f_t\left(x_j\right)\right)\right]+\Omega \left(f_t\right)$$

(1)

In this equation, $w_{ij}$ is the spatial weight from observation point $j$ to regression point $i$ (calculated in this study using a bi-square kernel with an adaptive bandwidth of 120 nearest training samples); $l$ is a differentiable convex loss function; and $\Omega \left(f_t\right)$ is the regularization term controlling model complexity.

For each held-out location, the 120 nearest eligible training samples define h, the distance to the 120th neighbor; weights are w = [1 − (d/h)²]² within h and zero outside. Thus, 120 is a sample count, not meters; 80 and 160 samples provide bandwidth sensitivity checks. XGBoost uses 120 trees, maximum depth 2, learning rate 0.05, minimum child weight 3, L2 regularization 5, and full row/column sampling; the random-forest comparator uses 200 trees and minimum leaf size 3. These hyperparameters were fixed a priori for all validation folds and were not tuned using held-out predictions. All responses and reported SHAP values remain in °C, with random seed 20260913.

To evaluate spatial predictive performance, the GW-XGBoost model was evaluated against the global Extreme Gradient Boosting (XGBoost) baseline using the same held-out native pixels. Comparative metrics including R², RMSE, and MAE were calculated to assess spatial predictive performance. Performance values are reported in Section 3.1, rather than in the Methods.

2.4.2. Model Interpretation

To interpret the machine-learning model and characterize nonlinear relationships among key factors, this study constructed a joint analytical approach integrating GW-XGBoost, SHAP, and Locally Weighted Scatterplot Smoothing (LOWESS). This methodology enables analysis across three dimensions: global importance, nonlinear response characteristics, and local spatial patterns. First, at the global level, SHAP importance plots were utilized to display the ranking of feature importance, quantifying the relative contribution magnitude of the candidate factors to UTCI predictions. Regarding nonlinear responses and candidate breakpoint identification, we analyzed the relationships between key urban planning factors (e.g., SVF and PD_Water) and UTCI by plotting SHAP Dependence Plots [37]. On this basis, Locally Weighted Scatterplot Smoothing (LOWESS) was introduced to fit the trend of SHAP value scatterplots, with a separate segmented-regression procedure used to quantify candidate breakpoints and uncertainty, and SHAP interaction analysis used to characterize conditional interaction patterns between selected factor pairs. Furthermore, local SHAP values were mapped back to geographic space to characterize the spatial heterogeneity of local feature contributions [39]. Together, these analyses support spatially differentiated interpretation of the modeled UTCI relationships and their potential planning relevance.

Each test sample retains its native_id, centroid coordinates, and original native-pixel polygon. Exact path-dependent TreeSHAP is calculated from that sample’s own geographically weighted model, then joined one-to-one by native_id to map the attribution on the original pixel polygon. The sum of feature contributions plus that local model’s expected value was checked against its prediction (maximum tolerance 10⁻⁴ °C). Local expected values differ across locations, and fold-excluded features have zero contribution. The SHAP interaction matrix was checked to sum to the same contributions. The spatial distribution maps display absolute contributions for local importance; The dependence analyses utilize signed contributions. The interaction analysis reports a single off-diagonal matrix entry per pair (not twice that value), together with LOWESS curves for partner values below and above the median.

Quantitative breakpoint identification uses a continuous one-break segmented model: SHAP(x) = β₀ + β₁x + β₂ max(0, x − τ), following the segmented-regression principle. We search 31 quantiles from the 20th to 80th percentile and choose τ minimizing residual sum of squares, requiring at least 15 observations and 10% of the sample on each side. ΔBIC = BIC(linear) − BIC(segmented) = n log(RSSlinear/RSSsegmented) − 2 log(n), counting four parameters for the segmented fit, including τ, versus two for the line. Positive ΔBIC favors the segmented model, whereas negative ΔBIC favors the linear model. ΔBIC is used here for model comparison and should not be interpreted as a formal significance test. Two hundred bootstrap replicates resample occupied 4 km × 4 km spatial tiles with replacement and give 95% percentile intervals. Fold-specific fits and equivalent global XGBoost/random-forest SHAP fits assess stability. The bootstrap intervals are conditional on the fitted SHAP attribution values and do not represent full model-refitting uncertainty. LOWESS was used only for visualization, while breakpoint estimation was based on segmented regression.

3. Results

3.1. Seasonal and Spatial Heterogeneity Patterns of UTCI

Prior to analyzing the seasonal spatial patterns, model performance was validated on geographically separated test sets. The results demonstrate that the GW-XGBoost model achieved higher mean spatial predictive accuracy across the four seasons (average R² = 0.408, RMSE = 0.745 °C, MAE = 0.547 °C) compared to the global XGBoost baseline (average R² = 0.276, RMSE = 0.888 °C, MAE = 0.698 °C). Following this validation, Figure 3 illustrates the seasonal spatial distribution of UTCI within the study area. Spatially, the region exhibits a generally warmer core relative to parts of the periphery across seasons. The magnitude and extent of this spatial gradient differ across seasons.

In terms of spatial distribution, relatively high UTCI values are concentrated in parts of the central and northern-central built-up area, while lower values occur mainly toward the southern and peripheral areas. In summer (Figure 3a), relatively high UTCI values are concentrated in parts of the central built-up area, with a maximum of 32.73 °C. The minimum among the eligible native cells is 29.22 °C, indicating clear spatial variation in summer UTCI. In winter (Figure 3c), a pronounced north–south spatial contrast is observed. Relatively high UTCI values occur across much of the central and northern area, reaching a maximum of 2.26 °C, whereas the southern fringe contains the lowest values, with a minimum of −2.19 °C.

Across seasons, UTCI exhibits marked differences in spatial gradients. In autumn (Figure 3b), relatively high UTCI values occur in parts of the central and northern area, with a maximum of 18.35 °C, whereas lower values occur toward the southern fringe, with a minimum of 14.52 °C. In spring (Figure 3d), relatively high UTCI values occur in parts of the central and northern area, reaching a maximum of 23.71 °C, while lower values occur mainly toward the southern fringe, with a minimum of 18.08 °C. Overall, relatively high UTCI values are repeatedly observed in the central and northern-central portions of the study area, while the magnitude and spatial extent of the contrast with the southern and peripheral areas vary among seasons.

Figure 3. Seasonal spatial distribution of UTCI across eligible native cells within the study area. (a) Summer; (b) autumn; (c) winter; (d) spring.

Figure 3. Seasonal spatial distribution of UTCI across eligible native cells within the study area. (a) Summer; (b) autumn; (c) winter; (d) spring.

3.2. Relative Contributions and Seasonal Shifts in UTCI Driving Factors

Having revealed the seasonal and spatial differentiation of UTCI, this study utilized the GW-XGBoost model to further quantify the relative contributions of the selected predictors to UTCI. The model demonstrated season-dependent spatial predictive performance across the four seasons, with the highest spatial predictive R² in winter (R² = 0.678), followed by spring (R² = 0.411) and autumn (R² = 0.300), and lower performance in summer (R² = 0.244). These results indicate clear seasonal differences in the spatial predictive performance of the model.

SHAP-based feature importance (Figure 4) reveals substantial seasonal variation in the relative contributions of predictor categories. The relative contribution of landscape composition increases from summer to winter. In summer (Figure 4a) and autumn (Figure 4b), terrain accounts for 22.5% and 29.4% of the total contribution, respectively, while landscape composition accounts for 30.1% and 42.4%. These results indicate substantial relative contributions from both terrain and landscape composition during summer and autumn. In winter (Figure 4c), the relative contribution of landscape composition increases to 52.7%, exceeding that of terrain (18.1%) and representing the largest category contribution. This indicates a greater relative contribution of landscape composition to the modeled spatial variation in winter UTCI. In spring (Figure 4d), landscape composition remains the largest category (42.4%), followed by terrain (27.8%).

At the individual-feature level, the rankings also vary across seasons. In summer, ET and Elderly Population Density (POPD_Elder) rank first and second, followed by NDVI and elevation (DEM). The Socioeconomic category reaches its highest contribution among the four seasons at 9.9% in summer, indicating a relatively stronger modeled contribution of population-related variables during this season. Autumn is led by DEM and NDVI, followed by Slope. In spring (Figure 4d), ET ranks first, followed by Pland_Water and Slope, indicating notable contributions from environmental, landscape composition, and terrain variables. In winter, the top five features are Building Density (BD), AI, Pland_Water, the Bare Land Landscape Shape Index (LSI_Soil), and Water Aggregation (AI_Water), indicating relatively large contributions from building-morphology and landscape composition variables.

In summary, the relative contributions of UTCI predictors vary substantially across seasons, with landscape composition contributing more strongly in winter than in summer. Terrain, environmental, socioeconomic, building-morphology, and landscape variables also show season-dependent changes in their relative importance, indicating a clear seasonal reorganization of the modeled feature contributions.

Figure 5 illustrates the spatial distribution of local feature importance, revealing substantial spatial heterogeneity in the contributions of individual predictors to UTCI. Unlike the global feature-importance rankings, the local SHAP maps identify areas of relatively high predictor importance by quantifying feature contributions at each native UTCI cell. The maps therefore represent local importance intensity at the native UTCI-cell support. Figure 5 presents the three leading factors for each season using absolute SHAP magnitudes, while the directions of their contributions are examined using signed SHAP values in the dependence plots.

Socioeconomic, environmental, and terrain variables exhibit distinct spatial patterns of local importance across seasons. In summer, high absolute SHAP magnitudes for Elderly Population Density (POPD_Elder; Figure 5b) are concentrated in parts of the central and eastern built-up area, with a peak local contribution of 0.459 °C. ET (Figure 5a) shows relatively high local importance in the northern part of the study area, whereas NDVI (Figure 5c) exhibits a more concentrated central pattern. In autumn, terrain variables show pronounced spatial variation in local importance. High absolute SHAP magnitudes for Elevation (DEM; Figure 5d) are concentrated toward the southern and southwestern fringe, reaching a peak contribution of 1.097 °C, while Slope (Figure 5f) also exhibits elevated local importance along parts of the western and southern fringe. These patterns demonstrate distinct spatial footprints of predictor importance across seasons.

Figure 4. Global feature-importance ranking of explanatory factors based on mean absolute SHAP values across four seasons. (a) Summer; (b) autumn; (c) winter; (d) spring.

Figure 4. Global feature-importance ranking of explanatory factors based on mean absolute SHAP values across four seasons. (a) Summer; (b) autumn; (c) winter; (d) spring.

Figure 5. Spatial distribution of local feature importance for the three leading drivers in each season. (a) ET; (b) POPD_Elder; (c) NDVI; (d) DEM; (e) NDVI; (f) Slope; (g) BD; (h) AI; (i) Pland_Water; (j) ET; (k) Pland_Water; (l) Slope.

Figure 5. Spatial distribution of local feature importance for the three leading drivers in each season. (a) ET; (b) POPD_Elder; (c) NDVI; (d) DEM; (e) NDVI; (f) Slope; (g) BD; (h) AI; (i) Pland_Water; (j) ET; (k) Pland_Water; (l) Slope.

Landscape composition and building-morphology variables also exhibit spatially heterogeneous importance patterns. Water cover (Pland_Water; Figure 5i,k) shows relatively high absolute SHAP magnitudes in western portions of the study area in winter and a more southerly concentration in spring, with peak contributions of 0.631 °C and 0.744 °C, respectively. This indicates season-dependent spatial variation in the local importance of water-related landscape composition. In winter, high absolute SHAP magnitudes for Building Density (BD; Figure 5g) are concentrated in the western part of the study area, reaching a maximum of 1.050 °C, whereas AI, a landscape composition metric, exhibits relatively high local importance in the eastern area (Figure 5h). These contrasting spatial patterns further demonstrate the seasonal and spatial heterogeneity of local predictor importance.

3.3. SHAP Dependence Relationships and Candidate Breakpoints of Key Driving Factors

Following the spatial analysis of predictor importance, Figure 6 further examines the functional form of SHAP relationships between key predictors and UTCI, including both approximately linear and nonlinear patterns. Figure 6 presents signed SHAP dependence relationships together with LOWESS trends and segmented fits, allowing comparison between approximately linear responses and candidate nonlinear breakpoints. Each seasonal row displays three factors, together with quantitative segmented-fit estimates, ΔBIC values, and 95% spatial-bootstrap intervals. Positive ΔBIC favors the segmented description, whereas negative ΔBIC favors the linear description.

Socioeconomic variables show a gradual rather than sharply bounded relationship in summer. Elderly Population Density (POPD_Elder; Figure 6c) exhibits a generally increasing relationship with SHAP values, while the model comparison favors a linear description (ΔBIC = −9.9), indicating a gradual density-related association rather than a well-supported breakpoint. Terrain variables show clearer evidence of nonlinear responses. For DEM, the estimated breakpoints are 4.44 m in summer (Figure 6a) and 6.13 m in autumn (Figure 6d), with 95% spatial-bootstrap intervals of 4–6.84 m and 5.64–6.6 m, respectively. The corresponding ΔBIC values are 35.7 and 46.1, supporting segmented descriptions. Below these estimates, fitted SHAP values generally increase with elevation; beyond them, the summer curve flattens while the autumn segmented curve turns downward.

Environmental predictors exhibit different functional forms across seasons. In summer, SHAP values for ET (Figure 6b) generally decline as the supplied ET index increases, while the model comparison favors a linear description (ΔBIC = −11.3). Similarly, autumn NDVI (Figure 6e) shows a generally negative association, with the linear description favored over the one-break model (ΔBIC = −5.0). In contrast, Water Patch Density (PD_Water; Figure 6f) in autumn shows an increasing response followed by flattening, with an estimated breakpoint of 9.42 patches/100 ha (95% interval 7.65–13.6; ΔBIC = 69.1). These results demonstrate that response shape and statistical support for a specific breakpoint should be distinguished.

In winter and spring, building-morphology and water-related predictors also exhibit contrasting functional forms. The winter Sky View Factor (SVF; Figure 6i) declines toward zero before flattening near 0.683 (95% interval 0.662–0.752; ΔBIC = 3.8), providing modest support for a change in slope. Winter PD_Water (Figure 6h) shows stronger segmented support, increasing below approximately 15.6 patches/100 ha and becoming nearly flat above this level (95% interval 13–17.8; ΔBIC = 35.8). In contrast, Building Density (Figure 6g) is better described by a linear relationship (ΔBIC = −8.7). In spring, AI_Water (Figure 6l) shows a downturn near 86.4% (95% interval 84.3–90.9; ΔBIC = 59.1), while ET (Figure 6j) and Slope (Figure 6k) show estimated changes near 3.52 and 2.2°, respectively. Where ΔBIC is positive, these segmented-fit estimates are interpreted as candidate statistical breakpoints and locally testable nonlinear reference points rather than fixed planning thresholds. Where ΔBIC is negative, the linear description is preferred, and the segmented-fit estimate is not interpreted as a supported breakpoint.

Figure 6. (a–l) SHAP dependence relationships and candidate breakpoints for selected driving factors across four seasons. (a) DEM; (b) ET; (c) POPD_Elder; (d) DEM; (e) NDVI; (f) PD_Water; (g) BD; (h) PD_Water; (i) SVF; (j) ET; (k) Slope; (l) AI_Water. The red solid lines represent the Locally Weighted Scatterplot Smoothing (LOWESS) trends, and the purple dashed lines indicate the segmented-regression fits.

Figure 6. (a–l) SHAP dependence relationships and candidate breakpoints for selected driving factors across four seasons. (a) DEM; (b) ET; (c) POPD_Elder; (d) DEM; (e) NDVI; (f) PD_Water; (g) BD; (h) PD_Water; (i) SVF; (j) ET; (k) Slope; (l) AI_Water. The red solid lines represent the Locally Weighted Scatterplot Smoothing (LOWESS) trends, and the purple dashed lines indicate the segmented-regression fits.

3.4. Conditional Interaction Patterns Among Key Driving Factors

Figure 7 further characterizes conditional interaction patterns between selected pairs of key predictors across seasons. In summer, the ET–DEM interaction (Figure 7a) has a mean absolute interaction contribution of 0.0032 °C. The fitted curves for lower and higher DEM groups remain close to zero over much of the ET range, although individual observations exhibit both positive and negative interaction contributions. The ET–POPD_Elder interaction (Figure 7b) shows clearer differentiation between Population Density contexts mainly at lower ET values, with the fitted curves converging as ET increases. The ET–CH interaction (Figure 7c) likewise shows variation in the ET interaction pattern across canopy-height contexts.

Figure 7. (a–l) SHAP interaction dependence plots for selected driver pairs across four seasons. (a) ET × DEM; (b) ET × POPD_Elder; (c) ET × CH; (d) DEM × PD_Water; (e) DEM × AI_Soil; (f) NDVI × PD_Water; (g) AI × LPI; (h) BD × Pland_Water; (i) AI × PD_Water; (j) DEM × Pland_Water; (k) ET × SPI; (l) Slope × Pland_Water.

Figure 7. (a–l) SHAP interaction dependence plots for selected driver pairs across four seasons. (a) ET × DEM; (b) ET × POPD_Elder; (c) ET × CH; (d) DEM × PD_Water; (e) DEM × AI_Soil; (f) NDVI × PD_Water; (g) AI × LPI; (h) BD × Pland_Water; (i) AI × PD_Water; (j) DEM × Pland_Water; (k) ET × SPI; (l) Slope × Pland_Water.

Autumn and spring exhibit different combinations of terrain, vegetation, and water-related interactions. In autumn, the DEM–PD_Water interaction (Figure 7d) has a mean absolute interaction contribution of 0.0043 °C. At lower elevations, the fitted curve for the lower PD_Water group is more positive than that for the higher PD_Water group, while both approach zero as elevation increases. The DEM–AI_Soil and NDVI–PD_Water pairs (Figure 7e,f) show additional conditional differences. In spring (Figure 7j–l), the selected interactions involve DEM–Pland_Water, ET–SPI, and Slope–Pland_Water. Among these, DEM–Pland_Water has the largest mean absolute interaction contribution at 0.0057 °C. These seasonal differences indicate changes in the combinations of predictors exhibiting the strongest interaction contributions.

Winter interactions are concentrated among landscape composition, landscape pattern, and building-morphology variables. The AI–Largest Patch Index (LPI) interaction (Figure 7g) has the largest mean absolute interaction contribution at 0.0060 °C. Although individual observations show both positive and negative interaction contributions, the fitted curves remain close to zero across much of the AI range. The Building Density–water cover (BD–Pland_Water) interaction (Figure 7h) has a mean absolute contribution of 0.0057 °C, followed by AI–PD_Water at 0.0052 °C (Figure 7i). Overall, the winter results indicate that building morphology and landscape characteristics exhibit stronger conditional interaction contributions than those observed for the selected summer pairs.

4. Discussion

4.1. Seasonal Heterogeneity of Urban Thermal-Environment Regulation Mechanisms

This study indicates that the associations between UTCI and its predictors vary substantially across seasons [43,44,45]. Such seasonal variation suggests that thermal-comfort priorities differ between hot and cold periods: summer interventions primarily focus on limiting heat stress, whereas winter strategies need to consider the maintenance of outdoor thermal comfort under cold conditions.

Urban-form and radiative-exposure conditions also show season-dependent associations with UTCI. In summer, lower SVF may increase shading and reduce direct solar exposure, thereby contributing to lower daytime thermal loads [46,47]. In winter, the SVF relationship shows a candidate breakpoint near 0.683 (Figure 6i). Below this estimate, the fitted SHAP contribution declines as SVF increases, whereas above it the fitted relationship remains close to zero. A possible physical interpretation, consistent with previous studies, is that more enclosed urban spaces can provide shelter from cold airflow and reduce convective heat exchange [45]. Given the modest ΔBIC support and bootstrap uncertainty, 0.683 should therefore be interpreted as a candidate morphological reference for local testing rather than as a fixed physical or design boundary.

During cooler periods, differences in local surface orientation and exposure may also influence thermal comfort through variations in received solar radiation. In highly urbanized areas, these surface orientation effects may coexist with the influence of built morphology on shading and airflow. These processes provide a possible physical context for the observed statistical associations.

Water-related landscape variables likewise exhibit season-dependent relationships with UTCI. In summer, previous studies have shown that water bodies can reduce surrounding thermal loads through evaporative cooling [22,48]. In winter, the SHAP dependence relationship for Water Patch Density (PD_Water; Figure 6h) increases below approximately 15.6 patches/100 ha and becomes nearly flat above this level. This pattern is consistent with the potential role of water-related landscapes as local microclimate stabilizers under cold conditions [49,50]. One possible interpretation is that water heat-storage processes may contribute to relatively moderated thermal conditions in surrounding areas. These findings suggest potential value in maintaining functional blue-green networks, while their specific thermal performance should be evaluated under local conditions.

4.2. Nonlinear Breakpoints and Conditional Interactions Among Driving Factors

Beyond seasonal variation in predictor importance, the results identify candidate nonlinear breakpoints for several UTCI relationships [51,52] together with conditional interaction patterns among selected predictors. Their planning interpretation should therefore consider both estimated numerical ranges and their associated uncertainty.

Socioeconomic variables exhibit comparatively gradual associations with UTCI. In summer, Elderly Population Density (Figure 6c) shows a generally increasing SHAP relationship, while model comparison favors a linear description rather than a segmented breakpoint. Population distribution therefore remains relevant to thermal exposure and vulnerability [53,54]. In contrast, several environmental and landscape variables show stronger support for candidate nonlinear relationships [55]. Winter PD_Water exhibits a change in slope near 15.6 patches/100 ha (Figure 6h), whereas SVF shows a weaker candidate change near 0.683 (Figure 6i) [56,57]. The corresponding spatial-bootstrap intervals quantify the uncertainty around these estimates and support their interpretation as locally testable statistical reference points rather than exact planning standards.

Conditional interaction patterns also vary across seasons [58,59]. In summer, the ET–DEM interaction (Figure 7a) remains close to zero across much of the ET range, while the ET–population interaction (Figure 7b) shows greater differentiation between population contexts at lower ET values. In autumn, DEM–PD_Water (Figure 7d) exhibits conditional differences at lower elevations, whereas spring interactions involve combinations of water cover, terrain, and environmental variables (Figure 7j–l). In winter, the selected interactions are concentrated among building-morphology and landscape variables, with AI–LPI and BD–Pland_Water showing relatively large mean absolute interaction contributions (Figure 7g,h). These seasonal differences highlight the importance of considering predictor combinations and their conditional relationships rather than interpreting individual factors in isolation.

4.3. Spatial Planning Strategies for Climate-Adaptive Cities

The results suggest that urban thermal-environment planning should account for seasonally varying spatial associations, candidate nonlinear breakpoints, and local validation rather than relying on uniform interventions [60,61,62]. Planning responses should also distinguish between existing urban regeneration areas and new development areas because their capacities for modifying urban form differ substantially.

For existing urban regeneration areas, where large-scale modification of buildings, roads, and property boundaries is often constrained, thermal adaptation may prioritize incremental and surface-oriented interventions. Population Density shows a gradual association with summer thermal conditions (Figure 6c), making thermal exposure and local microclimate conditions relevant considerations for regeneration planning. In high-density built-up areas, interventions can therefore emphasize quality-oriented improvements, including vertical greening, high-reflectance coatings, permeable surface treatments, and enhancement of existing blue-green connections [63,64], rather than relying primarily on large-scale reductions in development intensity.

For new development areas and urban fringes, greater spatial flexibility allows climate adaptation to be incorporated before urban form and development intensity are fixed. Ventilation corridors may be reserved where local wind and thermal assessments support their effectiveness [65], while street-canyon geometry, solar exposure, and blue-green infrastructure can be coordinated during the initial planning stage. The winter SVF breakpoint near 0.683 (Figure 6i; 95% interval 0.662–0.752) may serve as a candidate quantitative reference for local testing of street-canyon conditions rather than as a mandatory design limit. Compact urban forms may enhance winter shelter while also constraining summer ventilation [66,67]; consequently, open-space and street-canyon design should jointly consider solar exposure, wind conditions, and seasonal thermal objectives [68].

Landscape planning should likewise consider landscape composition and configuration rather than coverage alone. The winter PD_Water estimate near 15.6 patches/100 ha (Figure 6h) represents an area-weighted patch-density descriptor rather than a water-cover percentage. Water cover, aggregation, patch density, and surrounding morphology can therefore be assessed jointly when evaluating cold-season thermal conditions [69,70]. In naturally moisture-retaining areas and existing drainage or blue-green corridors, planning may prioritize the preservation of permeable and vegetated surfaces and avoid unnecessary sealing or concreting. Vegetation placement can additionally account for terrain, moisture availability, and existing blue-green networks [71,72], supporting evapotranspirative cooling while allowing possible cold-season buffering effects to be evaluated locally.

4.4. Limitations

Although this study has advanced the quantitative understanding of season-dependent associations and candidate nonlinear breakpoints of urban thermal comfort (UTCI), several limitations remain. Constrained by the spatiotemporal resolution of multi-source data, uncertainties persist in characterizing the heterogeneity of micro-scale thermal environments. In addition, although explanatory variables were aggregated to the native UTCI grid, residual uncertainty associated with the Modifiable Areal Unit Problem (MAUP) remains. While candidate nonlinear breakpoints were estimated, current raster data struggle to precisely capture the refined impact of instantaneous wind fields and complex shadows within street canyons on human thermal experience, particularly regarding radiant heat load. Some explanatory variables focus on passive physical regulation, while active anthropogenic heat sources such as traffic and air-conditioning exhaust are not directly represented, which limits model explanatory power in high-intensity activity zones. Furthermore, although SHAP interaction analysis identifies nonlinear associations, it is fundamentally based on statistical correlation. It currently lacks physical validation of fluid-dynamic processes within ventilation corridors and requires further process-based validation of energy-transfer mechanisms. Future work could expand on multi-source data fusion and mechanism coupling, incorporating airborne LiDAR to build sub-meter 3D urban canopy models to enhance data precision, and combining crowd movement data to explore the transition from static evaluation to dynamic thermal comfort assessment. The analysis covers a single seasonal cycle and integrates multi-source datasets with differing spatial and temporal resolutions. In addition, the reported bootstrap intervals are conditional on the fitted SHAP attribution values. Future studies incorporating multi-year observations and higher-resolution data could further evaluate the robustness and transferability of the identified seasonal relationships.

5. Conclusions

This study developed an interpretable spatial machine learning framework integrating GW-XGBoost and SHAP to investigate the seasonal differentiation of urban thermal comfort in the central urban area of Tianjin, China. By jointly considering nonlinear responses and spatial non-stationarity, the study systematically revealed the seasonal and spatial differentiation of UTCI across its native grid and the season-dependent associations of its explanatory factors. The findings provide quantitative evidence for climate-adaptive thermal-environment management and sustainable urban planning. The main conclusions are as follows.

First, UTCI exhibits pronounced spatial structure and seasonal differences in thermal gradients. Across all four seasons, the study area generally maintains higher UTCI values in the core relative to parts of the periphery. In summer, relatively high UTCI values are concentrated in parts of the central built-up area, reaching a maximum of 32.73 °C. In winter, relatively high UTCI values occur across much of the central and northern area, reaching a maximum of 2.26 °C, whereas the minimum among eligible native cells is −2.19 °C. Across seasons, autumn values range from 14.52 °C to 18.35 °C and spring values from 18.08 °C to 23.71 °C, demonstrating pronounced seasonal differences in spatial thermal gradients. These findings indicate substantial seasonal variation in the spatial distribution of urban thermal comfort, which should therefore be considered in season-specific climate adaptation.

Second, the relative contributions of UTCI predictors exhibit substantial seasonal shifts. In summer and autumn, environmental, terrain, and socioeconomic variables all contribute to the modeled UTCI variation, with ET leading in summer and DEM in autumn. In winter, the leading individual factor is Building Density and the largest category is landscape composition, indicating increased relative contributions from building-morphology and landscape composition variables. Overall, predictor importance changes markedly across seasons, highlighting the need to move beyond uniform year-round thermal-environment interventions toward seasonally adaptive urban design.

Third, candidate nonlinear breakpoints were quantified for selected predictors using segmented fits and spatial-bootstrap intervals. The gradual association of Elderly Population Density with summer thermal conditions highlights the relevance of complementary surface and microclimate strategies in density-related planning. The SVF breakpoint of approximately 0.683 has a 95% interval of 0.662–0.752. In addition, winter PD_Water exhibits a change in slope near 15.6 patches/100 ha (95% interval 13–17.8), while spring AI_Water turns downward near 86.4% (95% interval 84.3–90.9). These estimates provide candidate quantitative reference points for local testing of street-canyon design and blue-green infrastructure planning rather than fixed planning thresholds.

Fourth, season-dependent conditional interaction patterns were identified among selected predictors. In summer, leading interaction patterns involve ET together with DEM, Population Density, and canopy height. In winter, AI–LPI and Building Density–Water Cover are among the interaction pairs with relatively large mean absolute contributions. These results indicate that thermal-comfort interventions should consider not only individual factors but also conditional predictor combinations and their seasonal context.

In summary, this study advances the understanding of urban thermal comfort from the identification of general predictor importance toward the quantification of season-specific nonlinear relationships and candidate breakpoints, spatially heterogeneous contributions, and conditional interaction patterns. By translating these findings into uncertainty-informed, locally testable, and spatially differentiated planning guidance, the study provides quantitative support for climate-adaptive urban design, including differentiated density management, seasonally responsive urban morphology, and targeted blue-green infrastructure interventions. More broadly, the proposed framework contributes to improving urban thermal resilience and outdoor livability and provides an evidence base for more sustainable urban development under intensifying climate and urbanization pressures.