Deciphering the Seasonal Thermal Environments in Kunming’s Central Urban Area Using LST and Interpretable Geo-Machine Learning

Highlights

What are the main findings?

• Seasonal LST revealed a seasonal thermal dichotomy: extreme summer UHI in urban cores versus anomalous widespread spring heat in natural and vegetated landscapes.
• Geo-machine learning uncovered a seasonal rotation of thermal drivers, showing that built environments dominate summer and autumn warming, while natural and topographic factors govern spring and winter.

What are the implications of the main findings?

• The spring thermal anomaly highlights that drought-induced vegetation stress and peri-urban greenhouse agriculture act as major, localized heat sources.
• The strong spatial non-stationarity of thermal drivers proves that identical urban features behave differently depending on location, requiring spatially targeted climate-adaptive planning in plateau-basin cities.

Abstract

Rapid urbanization and complex topography complicate Urban Heat Island (UHI) spatio-temporal dynamics. Traditional models and coarse-resolution imagery often fail to capture fine-scale, spatially non-stationary seasonal driving mechanisms. This study investigates the multi-dimensional drivers of surface thermal dynamics in Kunming, a typical low-latitude plateau city, using seasonal median LST composite (2018–2025). Integrating eXtreme Gradient Boosting (XGBoost) with eXplainable Artificial Intelligence (XAI) models decoupled the nonlinear impacts of these drivers. Results reveal a seasonal thermal dichotomy: Summer exhibits the most intense UHI effect with extreme peak temperatures, while Spring presents an anomaly where natural and vegetated Local Climate Zones (LCZs) show pronounced warming. SHapley Additive exPlanations (SHAP) analysis identified a seasonal rotation: anthropogenic and structural factors dominate Summer and Autumn warming, whereas natural and topographic regulators govern Spring and Winter. GeoShapley deconstruction demonstrated strong spatial non-stationarity. Building-density warming is amplified in poorly ventilated urban cores, and fragmented vegetation’s cooling is offset by anthropogenic heat during peak summer. This study provides new insights into the seasonal drivers of urban thermal environments in plateau cities.

1. Introduction

Rapid urbanization has profoundly transformed the surface energy balance of cities worldwide, giving rise to intensified Urban Heat Island (UHI) effects [1,2,3,4,5]. Early theoretical work established the physical foundation of UHI formation, highlighting the role of altered surface roughness and enhanced heat storage capacity in urban environments [6]. Subsequent studies have demonstrated that the large-scale replacement of natural vegetation and permeable soils with impervious materials such as concrete and asphalt modifies surface radiative properties, reduces evapotranspiration, and increases sensible heat storage [7,8]. Urban areas often exhibit significantly higher land surface temperatures (LST) than their surrounding rural environments. Under the dual pressures of global climate warming and accelerating urban expansion, elevated urban temperatures exacerbate cooling energy demand [9], intensify air pollution [10], and increase heat-related morbidity and mortality [11,12]. Consequently, understanding the spatial distribution, temporal evolution, and driving mechanisms of intra-urban thermal environments has become a critical research frontier for sustainable urban development.

Over the past decades, substantial progress has been made in characterizing urban thermal environments using satellite-derived Land Surface Temperature (LST) products. Moderate-resolution sensors such as MODIS have enabled large-scale intercity comparisons and global assessments of urban heat islands [13], while higher-resolution observations from the Landsat program have facilitated detailed intra-urban investigations [14,15]. Numerous studies have revealed that urban heat patterns are governed by a multidimensional system integrating two-dimensional (2D) surface biophysical characteristics (e.g., vegetation coverage and water bodies), three-dimensional (3D) urban morphology (e.g., building density, height, and sky view factor), topographic variability, and socioeconomic activities [16,17,18]. To better characterize the relationship between urban form and thermal behavior, the Local Climate Zone (LCZ) system was proposed by Stewart and Oke, providing a standardized classification scheme that links urban morphology with thermal responses under different climatic contexts [19]. Subsequent empirical studies demonstrated that heating and cooling contributions vary significantly across LCZ types and macroclimatic backgrounds [20,21,22,23], thereby improving the comparability of urban climate studies across different cities.

At the same time, growing evidence suggests that urban thermal processes are inherently dynamic and seasonally modulated. For instance, vegetation-induced cooling effects are closely associated with evapotranspiration intensity and vegetation phenology [8,13], while water bodies may exhibit seasonal thermal reversals depending on background climate conditions and heat capacity differences [24]. The dominant drivers of LST may shift between summer and winter due to variations in solar radiation and atmospheric humidity [25]. In monsoon-influenced regions, strong inter-annual climate variability may further reshape seasonal thermal patterns [5]. These findings indicate that urban heat cannot be fully understood through single-date or single-season analyses.

While high-resolution satellite imagery enables detailed observation of urban heat patterns, another major research frontier lies in identifying and quantifying the driving mechanisms underlying urban thermal patterns. Machine learning models have also been widely employed to analyze the relationships between LST and multidimensional environmental drivers such as vegetation, urban morphology, topography, and socioeconomic activities [26]. However, despite their high predictive accuracy, many machine learning models inherently operate as black-box models, making it difficult to interpret the relative importance and spatial variability of different driving factors. To overcome this limitation, explainable artificial intelligence (XAI) techniques have recently been introduced into environmental modeling. In particular, the SHapley Additive exPlanations (SHAP) framework provides a theoretically grounded approach based on cooperative game theory to quantify feature contributions to model predictions. GeoShapley extends classical SHAP theory into spatial analytics by explicitly incorporating geographic context into the attribution process [18,27]. By decomposing model predictions into intrinsic feature contributions and spatial interaction components, GeoShapley enables geographically explicit interpretation of nonlinear and spatially non-stationary mechanisms governing urban thermal environments.

In this study, we investigated the seasonal thermal environments of the central urban area of Kunming over the 2018–2025 period. First, Landsat 30 m seasonal median LST composites are calculated to characterize the spatiotemporal evolution of urban heat patterns, effectively overcoming persistent cloud contamination challenges during the monsoon season. Second, an eXtreme Gradient Boosting (XGBoost) attribution model coupled with XAI techniques is applied to interpret the driving mechanisms of LST variations. SHAP is employed to quantify the nonlinear contributions of morphological, topographic, and socioeconomic drivers, while GeoShapley further decomposes these effects to reveal their spatial heterogeneity across seasons. By coupling high-resolution thermal mapping with spatially explicit interpretability analysis, this study aims to advance the understanding of seasonal urban heat mechanisms in plateau cities and provide scientific evidence for climate-adaptive urban planning.

2. Materials and Methods

2.1. Study Area

The study area focuses on the central urban area of Kunming, the capital city of Yunnan Province in southwestern China (Figure 1). As a typical low-latitude plateau city situated at an elevation of approximately 1900 m, Kunming provides a representative case for investigating complex urban thermal dynamics [28,29]. Previous studies have documented macroclimate-dependent thermal differences among cities in Yunnan Province [30], highlighting the unique influence of plateau environments on urban heat compared to typical lowland megacities.

Kunming experiences a subtropical highland monsoon climate characterized by a pronounced dry–wet seasonal dichotomy [28,30]. The dry season, spanning from November to April, is marked by abundant sunshine, strong solar radiation, high diurnal temperature variations, and significant surface radiative cooling at night. Conversely, the wet season, from May to October, is dominated by concentrated precipitation, substantial contrasts in atmospheric humidity, and persistent cloud cover. Notably, this persistent cloud contamination during the monsoon season poses significant challenges for acquiring reliable single-date thermal imagery.

The regional topography further complicates these thermal patterns. The terrain of Kunming is characterized by alternating mountain ridges and semi-enclosed basin landscapes surrounding Dianchi Lake, creating strong microclimatic heterogeneity across the urban region. Furthermore, the city’s southern boundary is delineated by Dianchi Lake, China’s sixth-largest freshwater lake, which encompasses an area of approximately 330 km². The massive thermal capacity of Dianchi Lake drives localized lake-land breeze circulations, significantly modulating the seasonal microclimate of the adjacent urban neighborhoods [31]. Against this complex geographical backdrop, Rapid urban expansion in recent decades has further intensified land-use fragmentation and reshaped surface thermal patterns [32,33].

2.2. Data

To capture the seasonal thermal dynamics of Kunming’s central urban area, this study leveraged multi-source remote sensing datasets processed on the Google Earth Engine (GEE) cloud computing platform. Landsat-8/9 LST and Sentinel-2 MSI surface reflectance data were first masked for clouds and cloud shadows, and relevant biophysical indices were extracted. Pixel-wise seasonal median composites were generated for each meteorological season (Spring: March–May; Summer: June–August; Autumn: September–November; Winter: December–February) across the 2018–2025 period. This approach mitigates biases caused by extreme weather or persistent cloud cover, providing a reliable basis for the subsequent seasonal attribution analysis. Details of the data used in this study are shown in

Table 1. Summary of multi-source datasets used in this study.

Data Category	Dataset	Variables/Applications	Spatial Resolution	Time Range	Data Source
Thermal Remote Sensing	Landsat-8/9 OLI/TIRS (Collection 2 Level-2)	Initial LST	30 m	2018–2025	USGS
Multispectral Remote Sensing	Sentinel-2 MSI (Level-2A)	Surface reflectance and biophysical indices (NDVI, NDBI, NDWI, NDMI, Albedo, EVI, MNDWI)	10 m	2018–2025	ESA
	MOD16A2 V6.1	Evapotranspiration (ET)	500 m	2018–2025	LP DAAC
Topographical Data	ALOS PALSAR DEM	Elevation, Topographic Position Index (TPI)	12.5 m	-	JAXA
Urban Morphology	Global map of Local Climate Zones	3D structural fabric, building density, and aerodynamic roughness proxy	100 m	2023	WUDAPT protocol
Socioeconomic Data	WorldPop Database	Gridded population density (POP)	100 m	2021	WorldPop
	NPP-VIIRS	Nighttime Light (NTL) as a proxy for urbanization intensity and energy expenditure	500 m	2025	NOAA

.

2.2.1. Remote Sensing Dataset

Initial LST observations were obtained from the Landsat-8 and Landsat-9 OLI/TIRS Collection 2 Level-2 Science Products (M9.1), provided by the U.S. Geological Survey (USGS), Reston, VA, USA, and the National Aeronautics and Space Administration (NASA), Washington, DC, USA, offering atmospherically corrected surface temperature measurements resampled to 30 m resolution [34]. Synchronously, Sentinel-2 MSI Level-2A surface reflectance products, sourced from the European Space Agency (ESA), Paris, France, were used to extract spectral biophysical indices [35].

To capture the multidimensional surface characteristics, dynamic indices were derived to represent instantaneous surface states: the Normalized Difference Vegetation Index (NDVI) for vegetation greenness, the Normalized Difference Built-up Index (NDBI) for urban impervious surfaces, the Normalized Difference Water Index (NDWI) for surface water, and broadband surface Albedo to capture overall surface reflectivity [36,37,38,39,40].

To provide a stable characterization for multi-year seasonal attribution, these dynamic indices were aggregated into seasonal means across the 2018–2025 period. Enhanced indices, including the Enhanced Vegetation Index (EVI) and the Modified Normalized Difference Water Index (MNDWI), were also incorporated to better delineate vegetation vigor and water bodies under complex plateau-basin conditions [41,42]. The resulting Enhanced Vegetation Index (EVI) and Normalized Difference Built-up Index (NDBI) represent multi-year seasonal averages of their respective instantaneous indices, providing a stable characterization of the urban morphological and biophysical background.

2.2.2. Urban Morphology and Topographical Data

For urban morphology, the high-resolution (100 m) Global map of Local Climate Zones was utilized [43]. Built upon the standard WUDAPT protocol and the foundational LCZ typology [19], this globally consistent dataset serves as a holistic categorical proxy for the city’s 3D structural fabric, building density, and aerodynamic roughness. Furthermore, given Kunming’s complex mountain-basin terrain, high-precision topographical controls were derived from the 12.5 m ALOS PALSAR Digital Elevation Model (DEM). Key variables extracted included absolute Elevation (DEM), Slope, Aspect, and the Topographic Position Index (TPI). TPI was specifically utilized to geographically differentiate between well-ventilated ridges and enclosed valleys prone to cold-air pooling.

2.2.3. Socioeconomic Data

To adequately quantify localized anthropogenic heat emissions and human disturbance during the spatial attribution modeling phase, two macro-level socioeconomic datasets were incorporated. High-resolution (100 m) gridded population density (POP) was acquired from the WorldPop database [44], reflecting the intensity of human presence and residential energy consumption. Additionally, Nighttime Light (NTL) data, acting as a robust spatial proxy for urbanization intensity and commercial energy expenditure, were sourced from the NPP-VIIRS (Visible Infrared Imaging Radiometer Suite) sensor, manufactured by Raytheon Company (Raytheon Space and Airborne Systems), Goleta, CA, USA, aboard the Suomi National Polar-orbiting Partnership (Suomi NPP) satellite platform developed by Ball Aerospace & Technologies Corp., Boulder, CO, USA, with data products distributed by the National Oceanic and Atmospheric Administration (NOAA), Silver Spring, MD, USA.

2.2.4. Evapotranspiration (ET) Data

To quantitatively investigate the hydrological controls on the regional thermal environment, Evapotranspiration (ET) data were acquired from the MOD16A2 Version 6.1 product. To ensure strict temporal parity and methodological consistency with the LST baseline, the ET time-series data were subjected to the identical temporal aggregation protocol. Specifically, ET observations from 2018 to 2025 were temporally grouped by meteorological season, and a pixel-wise median was calculated.

2.3. Method

Figure 2 illustrates the technical approach of this study. Seasonal thermal patterns in Kunming’s central urban area were analyzed by first calculating multi-year seasonal median LST from the Landsat observations to provide stable representations of thermal patterns across seasons. Finally, SHAP and GeoShapley, were applied to quantify and spatially decompose the contributions of morphological, topographic, and socioeconomic factors.

2.3.1. Calculation of Biophysical Indices

To capture the multidimensional surface characteristics, dynamic indices were derived to represent instantaneous surface states. We utilized the Normalized Difference Vegetation Index (NDVI) and the Enhanced Vegetation Index (EVI) to characterize vegetation greenness and vigor. The Normalized Difference Built-up Index (NDBI) was calculated to represent urban impervious surfaces, while the Modified Normalized Difference Water Index (MNDWI) was used to accurately delineate surface water bodies. Additionally, broadband surface Albedo was derived to capture overall surface reflectivity. To provide a stable characterization for multi-year seasonal attribution, these dynamic indices were aggregated into seasonal means across the 2018–2025 period.

The fundamental spectral indices were calculated using standard formulations [37,38,39,40]:

$$\mathit{NDVI}={\displaystyle \frac{{\rho }_{\mathit{NIR}}-{\rho }_{\mathit{Red}}}{{\rho }_{\mathit{NIR}}+{\rho }_{\mathit{Red}}}}$$

(1)

$$\mathit{NDBI}={\displaystyle \frac{{\rho }_{\mathit{SWIR}1}-{\rho }_{\mathit{NIR}}}{{\rho }_{\mathit{SWIR}1}+{\rho }_{\mathit{NIR}}}}$$

(2)

where ρ_Blue, ρ_Green, ρ_Red, ρ_NIR, ρ_SWIR1 represent the surface reflectance of the blue, green, red, near-infrared, and short-wave infrared bands of the Sentinel-2 imagery, respectively.

For the spatial attribution phase, to accurately delineate water bodies from complex urban shadows and quantify vegetation vigor more robustly, the MNDWI [42] and the EVI [41] were introduced:

$$\mathit{MNDWI} ={\displaystyle \frac{{\rho }_{\mathit{Green}}-{\rho }_{\mathit{SWIR}1}}{{\rho }_{\mathit{Green}}+{\rho }_{\mathit{SWIR}1}}}$$

(3)

$$\mathit{EVI}=2.5\times {\displaystyle \frac{{\rho }_{\mathit{NIR}}-{\rho }_{\mathit{Red}}}{{\rho }_{\mathit{NIR}}+6\times {\rho }_{\mathit{Red}}-7.5\times {\rho }_{\mathit{Blue}}+1}}$$

(4)

To transition to the multi-year seasonal analysis, the dynamic instantaneous indices were temporally aggregated. Specifically, the EVI and the NDBI utilized in the attribution models represent the multi-year seasonal averages of their respective instantaneous counterparts, ensuring a stable characterization of the urban morphological background. Finally, the broadband surface Albedo was derived utilizing a rigorous narrow-to-broadband conversion algorithm optimized specifically for the Sentinel-2 multispectral instrument [36]. The formulation calculates Albedo as a weighted linear combination of six spectral bands:

$$\mathit{Albedo}=0.2266{\rho }_{\mathit{Blue}}+0.1236{\rho }_{\mathit{Green}}+0.1573{\rho }_{\mathit{Red}}+0.3417{\rho }_{\mathit{NIR}}+0.1170{\rho }_{\mathit{SWIR}1}+0.0338{\rho }_{\mathit{SWIR}2}$$

(5)

where ρ_SWIR2 represents the surface reflectance of the second short-wave infrared band (Band 12) of Sentinel-2, and the coefficients reflect the proportional contribution of each spectral region to the total solar irradiance.

2.3.2. Seasonal LST Compositing and LCZ Analysis

To characterize seasonal thermal patterns, Landsat LST maps from 2018–2025 were grouped by meteorological season (spring, summer, autumn, winter), and a pixel-wise mean was calculated across all years to construct multi-year seasonal mean LST maps. These seasonal composites were then overlaid with the Local Climate Zone (LCZ) classification at 100 m resolution (LCZ 1–10 for built types; LCZ A–G for natural land cover) [19], and the mean LST was computed for each LCZ type. This procedure enables a morphology-oriented assessment of seasonal urban heat.

2.3.3. Correlation Analysis and Non-Linear XGBoost Modeling

To analyze the drivers of seasonal urban thermal patterns, all seasonal LST composites and their corresponding multi-dimensional factors (multi-year mean dynamic indices, static LCZ types, POP, NTL, and TPI) were resampled to a uniform 30 m grid using zonal statistics, aligning macro-socioeconomic data with micro-biophysical features. Prior to modeling, a correlation analysis was conducted, calculating Pearson’s r to identify and screen highly correlated predictors (|r| > 0.8), ensuring the interpretability of subsequent XAI analyses.

Four independent grid-based XGBoost models were then trained, one for each season. Model hyperparameters (e.g., learning_rate, max_depth, subsample) were automatically optimized using the Optuna framework, which iteratively explores the hyperparameter space to maximize predictive performance while preventing spatial overfitting [45].

2.3.4. Global and Spatial Attribution via SHAP and GeoShapley

While the grid-based XGBoost models accurately capture the synergistic forcings of multi-dimensional factors on LST, their inherent “black-box” nature obscures the geographical interpretation of these physical processes. To overcome this interpretability deficit, this study employed the SHAP framework, rooted in cooperative game theory [46]. Standard SHAP effectively decomposes global predictions into additive feature attributions, quantifying the overall importance of each morphological driver.

However, standard SHAP disregards spatial non-stationarity—the phenomenon where the same biophysical factor (e.g., vegetation) may exert vastly different cooling efficiencies depending on its geographic location (e.g., inside a valley vs. on a ridge). Therefore, this study advanced the analysis by introducing the GeoShapley framework [18,47]. Building upon the foundation of SHAP, GeoShapley explicitly integrates spatial coordinates and geographic context into the attribution decomposition:

$$f\left(x_i\right)={\varphi }_0+\sum _{j=1}^M{\varphi }_j\left(x_i\right)+{\varphi }_{\mathit{geo}}\left(x_i\right)$$

(6)

where ${\varphi }_0$ is the global baseline prediction, M is the number of driving factors, and ${\varphi }_j\left({\mathit{x}}_{\mathit{i}}\right)$ denotes the exact marginal attribution value (in °C) of the j-th feature at specific geographic location i. A positive ${\varphi }_j$ indicates a localized warming effect, whereas a negative value signifies cooling. By extracting and mapping these marginal attributions, GeoShapley transcends traditional non-spatial XAI, enabling the precise spatial diagnosis of where and to what extent a specific morphological or topographical factor dictates the localized urban microclimate across different seasons.

2.3.5. Seasonal Coupling Analysis of LST and ET

To empirically validate the thermodynamic coupling between surface moisture and thermal behavior across different seasons, a pixel-wise bivariate correlation analysis was conducted. The 8-year seasonal median LST and the corresponding 8-year seasonal median ET composites were spatially intersected. To ensure statistical robustness, valid pixels containing finite values in both datasets were extracted, and a comprehensive scatter analysis was performed for each season. Linear regression models were fitted to these extracted point clouds to calculate the Pearson correlation coefficient and the slope, thereby quantifying the seasonal shifts in the cooling efficiency of latent heat fluxes across the Kunming basin.

3. Results

3.1. Seasonal Distribution of LST and Thermodynamic Responses of LCZ

The seasonal LST maps of Kunming (Figure 3) visually reveal a complex seasonal thermal dichotomy. Rather than a simple linear warming trend, the maps display two distinct thermal paradigms. During the pre-monsoon Spring, the region presents a pervasive thermal anomaly. Instead of highly localized hotspots, the entire regional background, including vast peri-urban and rural areas, is elevated to a ubiquitous warm state. In Summer, the thermal intensity becomes highly agglomerated, with peak temperatures frequently exceeding 35 °C (deep red) clustered in specific northern and eastern built environments, creating a sharp contrast with the cooler “blue/yellow” zones of Dianchi Lake and the surrounding mountain ranges. In stark contrast, Spring presents a pervasive thermal anomaly. Crucially, the Summer-Spring difference map captures this divergent thermal trajectory: while natural mountainous regions exhibit minimal warming (white/light blue) due to the cooling effect of monsoon-driven evapotranspiration (ET), artificial and ET-suppressed surfaces experience sharp temperature escalations (deep red). Autumn reflects a gradual cooling phase, while Winter exhibits the lowest overall LST, with most of the region dropping below 15 °C.

The quantitative analysis of the seasonal mean LST further corroborates this thermal dichotomy, revealing significant and counterintuitive temporal shifts in the thermal behavior of various LCZs (Figure 4). Overall, the absolute highest temperatures occur in Summer; however, they are not driven by the compact built-up core. Instead, agricultural and low-plant covers—specifically LCZ D (greenhouses, 32.7 °C) and LCZ C (bush/scrub, 32.4 °C)—dominate the regional thermal peak. Concurrently, Spring exhibits a profound and anomalous thermal signature characterized by widespread heat stress across non-urban surfaces, setting the baseline for the Summer thermal divergence.

Regarding the thermal responses of different LCZ types, water bodies (LCZ G) consistently maintain the lowest temperatures across all four seasons (ranging from 16.7 °C in Winter to 28.4 °C in Summer), demonstrating the most stable and robust “cooling island” effect. Conversely, natural and agricultural cover types, particularly LCZ D (low plants), exhibit a severe reduction in their relative cooling capacity in Spring. The mean temperature of LCZ D in Spring reaches an elevated 30.6 °C, significantly exceeding the thermal output of high-density compact built-up areas such as LCZ 1 (27.8 °C). While absolute LSTs rise across the board in Summer, a drastic spatial inversion of the UHI effect emerges. The dense urban core (LCZ 1) exhibits a remarkably suppressed mean LST of 26.8 °C, highlighting the powerful “urban canyon shadow effect” at the 30 m resolution. In stark contrast, the contiguous, unshaded morphology of LCZ D, coupled with its severe ET suppression, sustains a persistent thermal buffer of 32.7 °C, making it the primary driver of regional daytime heat accumulation.

3.2. Linear Drivers and Multicollinearity in the Thermal Environment

The initial exploratory analysis utilized Pearson correlation matrices to quantify the linear relationships between diverse urban morphological, biophysical, and topographic indicators and LST across four seasons (Figure 5). The results reveal a consistent and strong linear dominance of urbanization metrics on thermal escalation. Specifically, the NDBI exhibits the strongest positive correlations with LST across all seasons, peaking during Summer (r = 0.77) and Spring (r = 0.76). Additionally, broadband surface Albedo also demonstrates a strong positive correlation with LST, particularly in Spring (r = 0.71) and Winter (r = 0.56). This underscores the fundamental role of artificial, low-albedo materials and dense urban fabrics in absorbing and retaining solar radiation. Furthermore, NTL, serving as a proxy for anthropogenic heat emissions and socioeconomic density, consistently demonstrates a moderate-to-strong positive association with LST, particularly in Summer (r = 0.68) and Autumn (r = 0.59).

Conversely, natural biophysical factors present more complex linear relationships. While MNDWI maintains a robust and stable negative correlation with LST year-round (reaching r = −0.64 in Spring), highlighting the consistent cooling capacity of urban water bodies, the NDVI displays surprisingly weak or even counterintuitive linear correlations. For instance, NDVI shows a slight positive correlation in Spring (r = 0.18) and a minimal negative correlation in Summer (r = −0.16). This suggests that in the specific high-altitude, low-latitude climatic context of Kunming, the cooling mechanism of vegetation is highly non-linear and likely overshadowed or confounded by background soil moisture and urban heat entrapment during dry seasons.

Crucially, the correlation matrices expose significant multicollinearity among the independent variables. For example, MNDWI and NDVI are highly inversely correlated (r = −0.81 in Summer), and Albedo exhibits strong inverse relationships with MNDWI (r = −0.68 in Summer). This inherent collinearity limits the explanatory power and reliability of traditional stepwise multiple linear regression (SMLR) models. For instance, increasing building density simultaneously reduces vegetation cover and alters local aerodynamics. This complex physical interplay necessitates the use of advanced, non-linear machine learning models to accurately decouple their individual contributions to the thermal environment.

3.3. Non-Linear Modeling Performance and SHAP Global Attribution

To address the limitations of linear frameworks, an XGBoost machine learning model was deployed, yielding exceptional predictive accuracy across all temporal scenarios. The model successfully captured the complex, non-linear dynamics of Kunming’s urban heat island, achieving R² scores of 0.88, 0.87, 0.83, and 0.79 for Spring, Summer, Autumn, and Winter, respectively (Figure 6). The corresponding RMSE remained remarkably low, ranging from 1.39 °C to 1.69 °C. The scatter plots of predicted versus observed LST demonstrate dense, symmetrical clustering along the 1:1 reference line, confirming that the XGBoost architecture effectively mitigates overfitting while capturing the synergistic effects of 2D landscape patterns, 3D building morphology, and topography. However, a closer inspection of the scatter plots reveals a minority of outlier points with larger prediction deviations, primarily distributed at the thermal extremes. The model occasionally underestimates extreme peak temperatures (where observed LST significantly exceeds predicted LST), which geographically correspond to highly localized micro-environments such as industrial facilities with low-thermal-inertia metal rooftops or specific peri-urban agricultural greenhouses that experience intense heat accumulation not fully captured by the input features. Conversely, prediction deviations in the moderate-to-lower temperature ranges are often situated at complex morphological boundaries, such as land-water interfaces or the edges of dense urban canyons affected by transient 3D shading. Furthermore, some anomalous points during the Summer season are likely artifacts of residual thin cloud contamination in the native Landsat imagery, which locally depresses the sensor-recorded LST below the model’s physically driven prediction.

The SHAP summary plots provide a quantifiable hierarchy of feature importance, revealing profound seasonal shifts in the primary drivers of the thermal environment (Figure 7). During the Summer and Autumn months, anthropogenic and structural urban features decisively dominate. NTL, NDBI, and LCZ emerge as the top contributors to warming. The SHAP color gradients clearly illustrate that higher values of these features (red dots) systematically push the model output towards higher temperatures. This indicates that during periods of high solar insolation, the physical volume and material composition of the urban core act as the primary thermal engines.

Interestingly, the driving mechanisms shift significantly during the Spring and Winter seasons. In Spring, natural factors (specifically NDVI, MNDWI, and Albedo) surpass urban structural metrics in global importance. The SHAP plots show that lower values of NDVI and MNDWI (blue dots) contribute to warming, while higher values (red dots) provide substantial cooling. Furthermore, the DEM acts as a persistent top-tier driver across all seasons. Higher elevations consistently yield negative SHAP values (cooling effect), affirming the profound local climate regulation provided by Kunming’s plateau-basin topography. The SHAP analysis definitively proves that UHI drivers in Kunming are not static; rather, they undergo a seasonal rotation between anthropogenic heat trapping in the wet and warm season and biophysical and topographic regulation in the dry and cool season.

3.4. GeoShapley Spatial Deconstruction and Heterogeneity

While the global SHAP summary plots (Figure 8) establish the overall hierarchical importance of various thermal drivers, the pixel-level GeoShapley analysis further reveals the profound non-stationarity of Kunming’s urban thermal environment. The resulting spatial SHAP maps delineate exactly how geographical location amplifies or attenuates specific warming or cooling contributions across the four seasons.

3.4.1. Spatial Amplification of the Built Environment

The spatial distribution maps of anthropogenic and morphological factors, specifically NTL, NDBI, and LCZ classifications, reveal a pronounced spatial agglomeration effect. Across all seasons, these features exhibit a massive, concentrated “red zone” (indicating a strong warming contribution) that is perfectly congruent with Kunming’s dense urban core.

This spatial pattern demonstrates a phenomenon: the warming impact of the built environment is highly location-dependent. Within the central business district (CBD) and high-density residential zones, the warming potential of these indicators is substantially amplified. Conversely, as these same NDBI or LCZ values extend toward the urban periphery and mountainous suburbs, their spatial SHAP values transition to pale red or even blue. This spatial divergence indicates that identical building densities or impervious surface fractions contribute significantly less to the local LST when situated adjacent to open rural or ecological landscapes.

3.4.2. Spatiotemporal Dynamics of Blue-Green Infrastructure

The GeoShapley analysis elucidates the highly complex, location-dependent cooling efficiency of natural land covers. Regarding water bodies (MNDWI), the spatial footprint presents a remarkably stable and visually distinct cooling pattern, characterized by a massive, persistent “blue cluster” in the southwestern quadrant across all seasons, perfectly delineating the Dianchi Lake basin. However, the spatial reach of this cooling effect exhibits seasonal variation: during Summer, the cooling gradient extends further inland, whereas in Winter, the negative SHAP values contract strictly to the immediate riparian zones.

In stark contrast to the relative stability of water, the cooling efficacy of vegetation (represented by NDVI and EVI) exhibits extreme spatial and seasonal volatility. During the Winter months, NDVI acts as a widespread cooling agent, evidenced by extensive blue distributions across non-urbanized peripheries. However, the GeoShapley map for Spring NDVI exhibits a highly fragmented thermal landscape, featuring a chaotic mix of blue and pale red pixels even in non-urbanized areas, which corroborates the anomalous positive linear correlation (r = 0.17) observed in the initial global analysis.

During the Summer season, the spatial map for NDVI reveals further heterogeneity within the urban core, where isolated green patches paradoxically register slight positive (red) SHAP values, indicating a localized warming contribution. Conversely, NDVI consistently registers robust, deep-blue negative SHAP values only when situated in large, contiguous suburban areas or along the mountainous fringes, underscoring that the thermal mitigation potential of green infrastructure is strictly dictated by its spatial context.

3.4.3. Topographic Effects in the Plateau-Basin

The GeoShapley maps for topographic factors reveal persistent, top-tier spatial drivers with stark binary manifestations. Across all four seasons, the DEM invariably presents a widespread warming effect (positive red SHAP values) concentrated within the central basin, while the surrounding high-elevation mountainous areas consistently present a strong cooling effect (deep blue). Furthermore, the Topographic Position Index (TPI) highlights micro-topographical influences, where localized depressions within the urban fabric consistently register higher positive SHAP values compared to localized ridges.

3.4.4. Seasonal Shifts in Dominant Drivers

Tracing single factors across the four seasons via GeoShapley reveals dynamic spatial footprints. For instance, the spatial warming footprint of NTL (anthropogenic intensity) in Spring is relatively modest and confined strictly to the commercial centers. However, during Summer, the NTL spatial warming effect expands significantly, manifesting as a massive, intense red cluster dominating the urban core. By Autumn, this spatial intensity recedes, and by Winter, the high positive SHAP values shrink back to the central business districts.

Similarly, surface Albedo exhibits a distinct and counter-intuitive spatial pattern. Rather than cooling the periphery, Albedo displays a concentrated cooling effect (deep blue) explicitly within the urban core across Summer, Autumn, and Winter. This suggests that the high-albedo artificial surfaces (e.g., concrete, reflective rooftops) in the dense city center provide a relative thermal buffering effect compared to heat-absorbing darker surfaces. In Spring, however, this spatial cooling footprint largely dissipates, transitioning to a weak, pervasive warming effect. In summary, the GeoShapley deconstruction conclusively demonstrates that Kunming’s thermal environment is characterized by a shifting seasonal hierarchy of drivers. The absolute geographic coordinate profoundly modulates whether a physical feature acts as a thermal liability (positive SHAP value) or an ecological asset (negative SHAP value), heavily dependent on its relative position within the basin core versus the mountainous periphery.

4. Discussion

4.1. Vegetation Phenology and Drought Effects

The anomalous phenomenon observed in Spring, where natural and vegetated LCZs exhibit pervasive heat stress and a severe loss of cooling capacity, is likely a profound reflection of the coupling effect between the macroclimate background and vegetation phenology in the southwest plateau monsoon climate zone. This mechanism is quantitatively corroborated by the seasonal LST-ET coupling analysis (Figure 9). During the typical extreme dry season in Spring, a robust negative correlation (R = −0.67) emerges between LST and ET. Natural vegetation is in its early growth stages with unclosed canopies, resulting in a severe lack of physical shading. Concurrently, extreme drought and water scarcity severely limit stomatal conductance and transpiration [48]. As evidenced by the lowest ET values in the Spring scatter plot, this paralysis of the latent heat flux causes abundant shortwave solar radiation to convert directly into surface sensible heat, dangerously narrowing the thermal gap between rural landscapes and the urban core. In contrast, the abundant rainfall during the Summer monsoon is expected to promote rapid plant elongation; the resulting peak in canopy density and Leaf Area Index (LAI) enhances physical shading and intense evapotranspiration, which effectively restores the thermal buffer of natural LCZs and exacerbates the relative Urban Heat Island intensity [48].

4.2. Thermal Effects of Greenhouse Agriculture

Importantly, the anomalous thermal signature of LCZ D in Spring, reaching elevated mean temperatures that astonishingly rival those of high-density compact built-up areas, is presumed to be closely associated with the unique agricultural land-use patterns in Kunming. In the peri-urban areas of Kunming, LCZ D primarily consists of vegetable and flower greenhouse growing regions, which are typically characterized by plantations covered by plastic foils [30]. The presence of these greenhouse covers is hypothesized to fundamentally alter the thermodynamic properties of the natural underlying surface. The plastic material of the greenhouses likely reduces vertical heat exchange and increases localized atmospheric back radiation, providing a highly effective thermal insulation effect [49]. Under intense solar radiation in Spring, heat rapidly accumulates within and on the surface of the greenhouses, potentially causing their satellite-retrieved LST to be significantly higher than that of conventional natural low vegetation. Indeed, due to this unique thermal signature, previous morphological studies have even defined agricultural greenhouses as a distinct category (LCZ H) in Kunming [30]. As Summer brings more rainfall, increased cloud cover, and possible artificial cooling interventions (e.g., mechanical ventilation and shading) within the greenhouses, their surface temperatures recede. Overall, the characteristic greenhouse agriculture in Kunming constitutes a non-negligible potential “heat source” in the peri-urban areas during the dry season, significantly elevating the regional background thermal baseline.

4.3. Spatial Non-Stationarity and Microclimate Mechanisms

The GeoShapley spatial deconstruction clearly indicates that Kunming’s thermal environment is governed by a shifting seasonal hierarchy, where geographic context determines whether a physical feature functions as a thermal liability or an ecological asset. The pronounced spatial amplification of warming effects in Kunming’s dense urban core (LCZs 1–3) is largely associated with the urban canyon effect. In these high-density central districts, overlapping anthropogenic heat emissions and severely restricted aerodynamic ventilation significantly enhance the baseline warming potential of impervious surfaces [6,50]. In contrast, similar building densities produce substantially weaker thermal impacts when located near the well-ventilated mountainous periphery, suggesting that urban geometry and ventilation conditions regulate heat retention more strongly than surface materials alone.

The cooling efficiency of blue–green infrastructure exhibits strong spatial and seasonal variability [51]. The persistent cooling footprint over the Dianchi Lake basin reflects the stable thermal buffering capacity of large water bodies, while its inland influence varies seasonally, extending further during Summer due to enhanced lake-breeze circulation and contracting to riparian zones in Winter. Vegetation cooling also demonstrates a clear spatial threshold. During the Spring dry season, sparse canopies and delayed phenology weaken evapotranspiration capacity, producing the mixed warming–cooling patterns observed in the GeoShapley maps. Even in Summer, small fragmented green spaces embedded within dense urban cores can be overwhelmed by surrounding impervious surfaces and anthropogenic heat, supporting previous findings that vegetation cooling effectiveness strongly depends on spatial configuration and connectivity [52].

Kunming’s plateau-basin topography further reinforces this spatial differentiation. Basin terrain tends to trap heat and restrict atmospheric circulation, amplifying urban heat accumulation in central low-elevation zones. Localized micro-topographic depressions can also function as heat sinks, contributing to the fragmented hotspot patterns observed in the TPI spatial SHAP maps.

4.4. Limitations and Future Perspectives

While this study provides a robust spatial deconstruction of Kunming’s seasonal thermal environment based on high-resolution observations, several limitations warrant attention. The summer accuracy degradation and occasional outliers reflect inherent challenges in complex terrains, where intense solar radiation, 3D shadows from dense buildings, rapid heat absorption by impervious surfaces, and localized cooling from vegetation and water bodies generate drastic micro-scale thermal gradients, compounded by mixed-pixel effects at LCZ boundaries, contributing to discrepancies between low- and high-resolution observations.

Additionally, reliance on optical and thermal infrared imagery (Landsat, Sentinel-2) introduces a clear-sky bias; although multi-year seasonal composites mitigate severe cloud contamination, residual thin clouds and aerosols, particularly during the plateau rainy season, mean thermal dynamics under overcast or rainy conditions are underrepresented [53]. This monsoon-induced data limitation is also reflected in the auxiliary Evapotranspiration (ET) products. As observed in the Summer LST-ET scatter plot, persistent cloud cover during the wet season severely disrupts the optical/thermal retrieval algorithms of ET, leading to significant data gaps, interpolation artifacts (vertical clustering), and the breakdown of observable LST-ET coupling. The analysis is also constrained to satellite overpass times (~10:30 AM), neglecting diurnal variations in urban heat island effects and the shifting importance of features such as the cooling efficiency of water bodies or thermal retention of buildings [25].

Finally, while biophysical, morphological, and topographical drivers were effectively decoupled, atmospheric interactions, especially in plateau-basin topographies prone to inversion layers, remain unaddressed. Integrating LST with Aerosol Optical Depth (AOD) and anthropogenic aerosol dynamics in future studies could elucidate heat-smog interactions and support comprehensive urban climate mitigation strategies [54].

5. Conclusions

In this study, we investigated the seasonal thermal dynamics of Kunming by calculating multi-year Landsat 30 m LST composites and integrating them with multidimensional environmental and anthropogenic predictors. Both SHAP and pixel-level GeoShapley frameworks were applied to quantify feature contributions and account for spatial heterogeneity, enabling a robust and interpretable analysis of urban heat dynamics.

The XBoost model achieved high predictive accuracy across all seasons, effectively capturing complex, non-linear thermal dynamics within heterogeneous urban landscapes. Multi-year seasonal analysis revealed a pronounced anomaly: while Summer exhibited the strongest UHI effect with extreme peak temperatures in built-up areas, Spring experienced pervasive heat stress in natural and vegetated LCZs, significantly elevating regional background temperatures. This Spring anomaly is likely associated with drought-induced vegetation stress, limited evapotranspiration, and the localized influence of greenhouse agriculture in peri-urban zones.

Seasonal driver analysis demonstrated a clear rotation of dominant influences. During the wet Summer and Autumn, anthropogenic and built-environment factors, including urban density, impervious surfaces, and night-time light intensity, were the primary drivers of thermal escalation. In contrast, during dry Spring and Winter, natural biophysical factors, such as vegetation cover, water bodies, and topography, exerted a stronger influence, highlighting the importance of climate context in modulating UHI drivers.

Pixel-level GeoShapley analysis further revealed strong spatial non-stationarity in the thermal effects of urban features. The warming impact of high-density built environments was amplified in poorly ventilated urban cores, whereas the same features contributed less heating in peripheral or well-ventilated zones. Similarly, the cooling efficacy of vegetation patches and water bodies was highly context-dependent, functioning most effectively when contiguous or located in suburban and ventilated areas, but often neutralized by surrounding anthropogenic heat in dense urban cores.

Overall, these findings provide actionable insights for climate-adaptive urban planning in plateau-basin cities, emphasizing the need for spatially targeted deployment of green infrastructure, strategic urban design, and management of localized heat sources to mitigate extreme thermal stress across both urban cores and peri-urban regions.