A Local Climate Zone-Based Seasonal Net-Benefit Assessment Model for the Urban Thermal Environment—A Case Study in a Cold-Region City

Abstract

The combined effects of urbanization and climate warming subject cold coastal cities to summer heatwaves and winter extreme cold, yet most studies emphasize built-environment modifications for summer overheating and lack evaluation methods and planning-oriented strategies to balance seasonal trade-offs. Using Dalian as a case study, we develop a seasonal net-benefit model that quantitatively characterizes and reconciles seasonally differentiated built-environment effects on land surface temperature (LST) and interprets urban heterogeneity within the Local Climate Zone (LCZ) framework. Summer LST is mainly governed by static factors such as greenspace configuration and topography, whereas winter LST is more sensitive to development intensity and locational factors, including building density and the Normalized Difference Built-up Index (NDBI). Coastal areas and mountainous green corridors are net-benefit zones performing well in both seasons, while dense industrial and compact low-rise areas account for ~80% of pronounced net-penalty zones. Compact mid- and high-rise neighborhoods show more favorable structural climatic conditions but with substantial retrofit potential (Retrofit Seasonal Net-Benefit Index (R-SNBI) markedly lower than Structural Seasonal Net-Benefit Index (S-SNBI) by ~3). Large low-rise problems mainly stem from an unfavorable structure rather than insufficient greenness, whereas industrial land has greater improvement potential via blue–green spaces. The framework supports refined climate adaptation, sustainability-oriented planning, and identifying urban renewal priority areas in cold-climate cities.

1. Introduction

Rapid urbanization, compounded by climate warming, has intensified urban thermal-environment challenges worldwide [1]. In mid- and high-latitude cities where space heating is essential, summer heatwaves and winter extreme cold coexist, creating a compound burden that links heat-related health risks, rising cooling energy use, and persistent winter heating demand [2,3,4,5]. The Intergovernmental Panel on Climate Change (IPCC)’s Sixth Assessment Report indicates that the 2011–2020 global mean temperature was approximately 1.1 °C higher than pre-industrial levels, accompanied by a broadly increasing frequency and intensity of extreme heat events [6]. Globally, an estimated ~5 million deaths per year are associated with exposure to non-optimal temperatures; heat-related mortality accounts for 0.91% of all-cause deaths and is increasing, whereas cold-related mortality accounts for ~8.52% [7]. Warming substantially boosts cooling demand and elevates peak loads on hot days [8]. By contrast, although heating demand shows an overall downward tendency, heating-climate cities still face non-negotiable requirements to maintain minimum heating services, resulting in a dual burden in which summer cooling pressure coexists with winter heating constraints [2,3].

The Local Climate Zone (LCZ) scheme provides a unified framework for representing fine-scale urban thermal environments; however, its applicability and classification approaches across different climate regimes and cities warrant further investigation [9,10]. By partitioning the urban surface into a set of built and natural classes and jointly encoding key attributes such as building height, building density, and surface cover, the LCZ framework has been widely used for urban heat-island assessments, ventilation-corridor identification, and climate-adaptive planning [9,11,12,13]. Nevertheless, widely used LCZ mapping pipelines often rely on World Urban Database and Access Portal Tools (WUDAPT)-based training data and random-forest classifiers. While this facilitates cross-city comparisons, classification accuracy and local suitability can be constrained in cities with pronounced topographic complexity, strong land–sea thermal contrasts, or distinctive climatic conditions [14,15,16,17]. Recent advances in deep-learning classifiers—particularly semi-supervised frameworks—enable a unified network architecture to leverage both a small set of high-quality labeled samples and large volumes of unlabeled remote-sensing imagery, thereby learning multi-scale spatial textures and semantic representations automatically. Compared with traditional machine-learning approaches, such methods typically exhibit stronger representational power and generalization in complex urban scenes [18,19,20]. In this context, adopting semi-supervised learning approaches that fully exploit local sample information for fine-scale LCZ mapping can substantially improve city-specific mapping accuracy and spatial expressiveness while preserving overall methodological consistency [21].

Machine learning has become a widely used approach for mapping the fine-scale spatial distribution of LST and probing its driving mechanisms based on indicators that describe the built environment and static surface attributes [22,23,24]. A range of multi-source predictors—including building height and density, the normalized difference vegetation index (NDVI), the normalized difference built-up index (NDBI), albedo, distance to the coastline, slope, and aspect—have been shown to be closely associated with seasonal variations in LST [25,26,27,28]. Ensemble learning models such as Random Forest and LightGBM are widely applied to LST modeling because they can effectively capture nonlinear relationships and accommodate high-dimensional feature spaces [22,29,30,31]. Meanwhile, SHapley Additive exPlanations (SHAP) provides a powerful tool for quantifying marginal contributions, identifying threshold effects, and revealing seasonal differences across predictors [32,33,34,35]. Despite these advances, much of the existing literature focuses on heat-risk identification and cooling-potential assessments under hot conditions, such as summer daytime extremes [36,37], or models different seasons separately and compares their drivers in parallel [38]. By comparison, studies that systematically examine and contrast summer- and winter-built environment–LST relationships within a unified modeling framework, and that explicitly characterize the mechanisms underlying their seasonal divergences, remain relatively limited [39,40].

In cold-region cities, a key unresolved challenge is how to integrate the differential impacts of built-form characteristics on winter versus summer thermal environments into a unified analytical framework and to conduct explicit trade-off assessments [41,42,43,44,45]. From an energy perspective, even subtle built-environment-induced changes in near-surface temperature can be disproportionately magnified in space-heating and space-cooling demand [2,3,46]. Evidence from heating/cooling degree-day analyses and building-energy studies indicates that, in mid- and high-latitude cities, a warming climate tends to reduce winter heating loads overall while markedly increasing summer cooling loads; moreover, the energy implications per unit temperature change are asymmetric between winter and summer [47]. Widely adopted climate-adaptation measures—such as urban greening, cool roofs, and high-albedo pavements—typically deliver substantial summer cooling and energy-saving benefits, yet in cold climates they may reduce solar heat gains in winter and thus increase heating demand [48,49,50]. Furthermore, compact high-density urban forms often exhibit a seasonal trade-off: while they provide beneficial shading in summer, they simultaneously restrict ventilation; conversely, their winter heat-retention capacity is offset by the challenges of increased heating demands. Although prior studies have systematically characterized seasonal LST patterns and identified key drivers—including LCZ, greenspace configuration, topography, and land–sea thermal contrasts [51,52,53,54]—few have integrated these effects into a unified framework that connects built-environment features to both seasonal LST fluctuations and the resulting building energy consumption and adaptation benefits [41,55].

The main objectives of this study are as follows: (1) to produce high-accuracy LCZ maps using semi-supervised learning, thereby establishing a refined morphological baseline for fine-scale urban analyses; (2) to systematically elucidate winter–summer built environment–LST relationships within a unified machine learning–SHAP framework, with particular attention to the seasonality of dominant drivers and their nonlinear threshold behaviors; and (3) to develop a seasonal net-benefit evaluation model that consolidates summer cooling benefits and winter warming effects into a unified metric framework, and—through integration with LCZ classes—identify the spatial patterns of seasonal net benefits across neighborhood types, providing quantitative evidence to support built-environment optimization and climate-adaptive renewal under a winter–summer trade-off perspective. Within a unified LCZ–machine-learning analytical framework, we further operationalize these seasonal mechanisms of built-environment impacts on LST into a seasonal net-benefit assessment and combine it with LCZ-stratified analyses, offering a quantitative pathway for reconciling winter–summer differences and optimizing urban thermal environments.

2. Materials

2.1. Study Area

The study area comprises the four core urban districts of Dalian—Zhongshan, Xigang, Shahekou, and Ganjingzi (Figure 1)—covering a total area of approximately 120.14 km². Dalian is situated at the southern tip of the Liaodong Peninsula and is surrounded by the sea on three sides; its topography descends from the southwest to the northeast, with a mosaic of mountainous–hilly terrain and coastal plains. Under the Köppen–Geiger classification, Dalian is categorized as Dwa, characterized by cold, dry winters and hot, humid summers with pronounced seasonal temperature contrasts [56]. According to the Seventh National Population Census, the study area hosts more than 2.8 million residents and concentrates the city’s population and construction activities (www.citypopulation.de (accessed on 16 December 2025)) [57]. The combined mountain–sea terrain and intensive urban development generate substantial local climatic heterogeneity, making this region a representative setting for examining how urban-form differences interact with seasonal thermal environments [58,59,60].

2.2. Data

The datasets used in this study include LCZ mapping inputs, LST, and a suite of built-environment and static surface indicators. All raster layers were resampled to a common spatial resolution of 30 m and integrated for spatial overlay analyses. This resolution matches the native sampling of the land surface temperature product, avoiding artificial downscaling of temperature data and reducing scale inconsistencies in multi-source raster overlays. It also provides a favorable trade-off between capturing neighborhood-scale thermal variability and limiting noise and computational burden. All computations were implemented in Python 3.10. Data sources and key dataset information are summarized in

Table 1. Data source.

Data	Resolution	Period	Source Link
Sentinel-1 SAR:VV;VH and Sentinel-2 MSI: B2 (490 nm); B3 (560 nm); B4 (665 nm); B8 (842 nm)	10 m	2024	https://developers.google.com/earth-engine/datasets/catalog/ (accessed on 5 June 2025)
High-resolution Google Earth imagery (for LCZ training samples)	1 m	2024	Google Earth Pro 7.3
Landsat 8/9 OLI/TIRS	30 m	Summer and winter clear scenes (July–August, January–February, 2024)	https://earthexplorer.usgs.gov/ (accessed on 8 June 2025)
NASADEM NASA/NASADEM_HGT/001	30 m	Static (~2000)	https://search.earthdata.nasa.gov/ (accessed on 6 June 2025)
JRC Global Surface Water	30 m	1984–2020	https://global-surface-water.appspot.com/ (accessed on 16 October 2025)
OpenStreetMap building footprints and roads	Vector	2024	https://www.openstreetmap.org/ (accessed on 18 October 2025)
VIIRS Nighttime Lights (VNL V2, DNB composite)	500 m	2020	https://eogdata.mines.edu/ (accessed on 16 October 2025)
WorldPop gridded population	100 m	2020	https://www.worldpop.org/ (accessed on 16 October 2025)

.

The LCZ mapping inputs consist of Sentinel-1 synthetic aperture radar (SAR) and Sentinel-2 optical imagery, together with expert-labeled training samples delineated on high-resolution Google Earth imagery (344 polygon samples in total). LST was obtained from thermal-infrared satellite products accessed via Google Earth Engine (GEE); scenes were filtered using quality-control criteria and then resampled to 30 m. The digital elevation model (DEM) was derived from global elevation data, clipped in GEE, and used to compute slope. NDVI, NDBI, MNDWI, and surface albedo were calculated from multispectral surface-reflectance products using standard formulations (

Table 2. Feature calculation formula.

Factors	Formula
NDVI [61]	$\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}(\mathrm{j})={\displaystyle \frac{{\mathsf{\rho }}_{\mathrm{N}\mathrm{I}\mathrm{R}}\left(\mathrm{j}\right)-{\mathsf{\rho }}_{\mathrm{R}\mathrm{e}\mathrm{d}}\left(\mathrm{j}\right)}{{\mathsf{\rho }}_{\mathrm{N}\mathrm{I}\mathrm{R}}\left(\mathrm{j}\right)+{\mathsf{\rho }}_{\mathrm{R}\mathrm{e}\mathrm{d}}\left(\mathrm{j}\right)}}$	${\mathsf{\rho }}_{\mathrm{N}\mathrm{I}\mathrm{R}}(\mathrm{j})$: surface reflectance of the near-infrared band in cell j ${\mathsf{\rho }}_{\mathrm{R}\mathrm{e}\mathrm{d}}\left(\mathrm{j}\right)$: surface reflectance of the red band in cell j
NDBI [62]	$\mathrm{N}\mathrm{D}\mathrm{B}\mathrm{I}(\mathrm{j})={\displaystyle \frac{{\mathsf{\rho }}_{\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}}\left(\mathrm{j}\right)-{\mathsf{\rho }}_{\mathrm{N}\mathrm{I}\mathrm{R}}\left(\mathrm{j}\right)}{{\mathsf{\rho }}_{\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}}\left(\mathrm{j}\right)+{\mathsf{\rho }}_{\mathrm{N}\mathrm{I}\mathrm{R}}\left(\mathrm{j}\right)}}$	${\mathsf{\rho }}_{\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}}\left(\mathrm{j}\right)$: surface reflectance of the short-wave infrared band in cell j ${\mathsf{\rho }}_{\mathrm{N}\mathrm{I}\mathrm{R}}\left(\mathrm{j}\right)$: surface reflectance of the near-infrared band in cell j
MNDWI [63]	$\mathrm{M}\mathrm{N}\mathrm{D}\mathrm{W}\mathrm{I}(\mathrm{j})={\displaystyle \frac{{\mathsf{\rho }}_{\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}}\left(\mathrm{j}\right)-{\mathsf{\rho }}_{\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}}\left(\mathrm{j}\right)}{{\mathsf{\rho }}_{\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}}\left(\mathrm{j}\right)+{\mathsf{\rho }}_{\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}}\left(\mathrm{j}\right)}}$	${\mathsf{\rho }}_{\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}}\left(\mathrm{j}\right)$: surface reflectance of the green band in cell j ${\mathsf{\rho }}_{\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}}\left(\mathrm{j}\right)$: surface reflectance of the short-wave infrared band in cell j
Albedo [64]	$\mathrm{A}\mathrm{l}\mathrm{b}\mathrm{e}\mathrm{d}\mathrm{o}(\mathrm{j})={\displaystyle \sum _{\mathrm{b}\in \mathrm{B}}}{\mathsf{\alpha }}_{\mathrm{b}}\hspace{0.17em}{\mathsf{\rho }}_{\mathrm{b}}(\mathrm{j})$	${\mathsf{\rho }}_{\mathrm{b}}\left(\mathrm{j}\right)$: surface reflectance of band b in cell j ${\mathsf{\alpha }}_{\mathrm{b}}$: weighting coefficient of band b in the broadband albedo calculation B: set of spectral bands used for albedo estimation
NGVI_270 [65]	${\mathrm{N}\mathrm{G}\mathrm{V}\mathrm{I}}_{270}(\mathrm{j})={\displaystyle \frac{1}{{\mathrm{N}}_{270}\left(\mathrm{j}\right)}}{\displaystyle \sum _{\mathrm{k}\in {\mathsf{\Omega }}_{270}\left(\mathrm{j}\right)}}{\mathrm{G}}_{\mathrm{k}}$	${\mathsf{\Omega }}_{270}\left(\mathrm{j}\right)$: set of pixels within 270 m of cell j ${\mathrm{N}}_{270}\left(\mathrm{j}\right)$: number of pixels in ${\mathsf{\Omega }}_{270}\left(\mathrm{j}\right)$ ${\mathrm{G}}_{\mathrm{k}}$: masked NDVI at pixel k
NGWI_150 [66]	${\mathrm{N}\mathrm{G}\mathrm{W}\mathrm{I}}_{150}(\mathrm{j})={\displaystyle \frac{1}{{\mathrm{N}}_{150}\left(\mathrm{j}\right)}}{\displaystyle \sum _{\mathrm{k}\in {\mathsf{\Omega }}_{150}\left(\mathrm{j}\right)}}{\mathrm{W}}_{\mathrm{k}}$	${\mathsf{\Omega }}_{150}\left(\mathrm{j}\right)$: set of pixels within 150 m of cell j ${\mathrm{N}}_{150}\left(\mathrm{j}\right)$: number of pixels in ${\mathsf{\Omega }}_{150}\left(\mathrm{j}\right)$ ${\mathrm{W}}_{\mathrm{k}}$: masked MNDWI at pixel k
BD [67]	${\mathrm{B}\mathrm{D}}_{\mathrm{L}}(\mathrm{j})={\displaystyle \frac{{\sum }_{\mathrm{i}\in {\mathsf{\Omega }}_{\mathrm{b}}\left(\mathrm{j}\right)}{\mathrm{A}}_{\mathrm{i}}}{{\mathrm{A}}_{\mathrm{c}\mathrm{e}\mathrm{l}\mathrm{l}}}}$	${\mathsf{\Omega }}_{\mathrm{b}}\left(\mathrm{j}\right)$: set of building footprints intersecting grid cell j ${\mathrm{A}}_{\mathrm{i}}$: plan area of building footprint i ${\mathrm{A}}_{\mathrm{c}\mathrm{e}\mathrm{l}\mathrm{l}}$: area of grid cell j
BH [68]	${\mathrm{B}\mathrm{H}}_{\mathrm{m}\mathrm{e}\mathrm{a}\mathrm{n}}(\mathrm{j})={\displaystyle \frac{{\sum }_{\mathrm{i}\in {\mathsf{\Omega }}_{\mathrm{b}}\left(\mathrm{j}\right)}{\mathrm{A}}_{\mathrm{i}}\hspace{0.17em}{\mathrm{h}}_{\mathrm{i}}}{{\sum }_{\mathrm{i}\in {\mathsf{\Omega }}_{\mathrm{b}}\left(\mathrm{j}\right)}{\mathrm{A}}_{\mathrm{i}}}}$	${\mathsf{\Omega }}_{\mathrm{b}}\left(\mathrm{j}\right)$: set of building footprints intersecting grid cell j ${\mathrm{A}}_{\mathrm{i}}$: plan area of building footprint i ${\mathrm{h}}_{\mathrm{i}}$: height of building i ${\sum }_{\mathrm{i}\in {\mathsf{\Omega }}_{\mathrm{b}}\left(\mathrm{j}\right)}{\mathrm{A}}_{\mathrm{i}}$: total building footprint area within grid cell j

). BH and BD were generated by rasterizing building footprints and associated height attributes from OpenStreetMap. Neighborhood metrics (NLBD, NSBD, NGVI, and NGWI) were computed from building or spectral rasters using neighborhood statistics. Euclidean distance layers—including the distance to water bodies, distance to green space, and distance to the city center—were derived from vector boundaries for water and green space and a city-center point. Nighttime light intensity was obtained from nighttime-light remote-sensing products, and population density was sourced from global gridded population datasets (e.g., WorldPop); both layers were resampled to 30 m and clipped to the study area.

3. Methods

3.1. Research Framework and Technical Workflow

Figure 2 illustrates the overall workflow of this study. The analysis comprises two components: (i) assessing the seasonal impacts of LCZ types on the urban thermal environment, and (ii) leveraging the inferred seasonal-difference mechanism to construct a Seasonal Net-Benefit Index. The workflow includes four stages. First, urban morphology is classified via semi-supervised learning. Second, remote-sensing and OSM datasets are integrated on a common 30 m grid to build an indicator library for the built environment and static surface attributes. Third, separate LightGBM regression models are developed for summer and winter LST, and SHAP is used to quantify predictor contributions and characterize nonlinear response patterns. Building on these results, SHAP outputs from both seasons are converted into O-SNBI, yielding R-SNBI and S-SNBI, which are then overlaid with LCZ classes for spatial analyses. Finally, this study derives practical strategies for planners and managers to optimize urban form and maximize seasonal net benefits.

3.2. LCZ Mapping Methodology

This study follows the LCZ scheme proposed by Stewart and Oke and delineates 14 LCZ classes within the study area, comprising 8 built classes and 6 natural classes [9]. LCZ7, LCZ9, and LCZC were excluded because they were rarely observed in the study area and/or insufficiently represented in the samples (

Table 3. LCZ classification results and descriptions.

Built-Up Types	Definition	Land-Cover Types	Definition
LCZ 1	Dense high-rise buildings; very high impervious fraction; little vegetation.	LCZ A	Continuous, densely wooded cover (closed canopy).
LCZ 2	Dense mid-rise buildings; high impervious fraction; little vegetation.	LCZ B	Discontinuous tree cover with significant open ground/grass.
LCZ 3	Dense low-rise buildings; high impervious fraction; little vegetation.	LCZ D	Grasses/crops and other low vegetation dominate.
LCZ 4	High-rise buildings with open spacing; more pervious/vegetated surfaces than compact classes.	LCZ E	Exposed rock or paved/built hard surfaces with minimal vegetation.
LCZ 5	Mid-rise buildings with open spacing; mixed impervious and pervious surfaces.	LCZ F	Unvegetated or sparsely vegetated soil/sand surfaces.
LCZ 6	Low-rise buildings with open spacing; relatively abundant pervious surfaces/vegetation.	LCZ G	Open water bodies (rivers, lakes, sea).
LCZ 8	Large-footprint low-rise buildings (e.g., warehouses, malls); extensive paved surfaces; low vegetation.
LCZ 10	Industrial facilities with large structures and paved/industrial surfaces; sparse vegetation.

).

The workflow is as follows. First, representative training polygons for each LCZ class were manually delineated on high-resolution Google Earth imagery (at least 24 polygons per class), and a 10 m optical–SAR multimodal image-patch dataset was constructed. Training polygons were selected using three explicit criteria: (i) class purity (homogeneous urban form/land cover, avoiding boundaries and mixed areas); (ii) sufficient size to support patch extraction at 10 m; and (iii) spatial representativeness across the study area to reduce clustering. Next, a self-supervised multimodal contrastive-learning approach was used to pre-train the model on large volumes of unlabeled patches, followed by supervised fine-tuning with the labeled samples to obtain a deep LCZ classifier. The trained model was then applied across the entire study area using sliding-window inference to generate an initial 10 m LCZ raster. Subsequently, pixels that were clearly inconsistent with the underlying urban form were manually corrected to produce the final LCZ map. Classification accuracy was assessed using independent validation samples, and the LCZ results were aggregated to a 30 m regular grid via a majority-vote rule. Finally, the summer and winter LST layers were overlaid with the LCZ map to characterize LST distributions and quantify deviations relative to the citywide mean.

3.3. Selection of Built-Environment Indicators and Static Surface Attributes

In this study, a built-environment and static land-surface indicator system was constructed from dimensions such as remote-sensing spectra, topography and location, and urban morphology and activity intensity, to characterize the relationships between summer and winter LST and underlying-surface characteristics in the central urban area of Dalian. Indicator selection followed three principles: first, they have been relatively maturely applied in urban thermal-environment studies; second, they have clear physical meanings and are convenient for interpretation and intervention in planning and renewal practices; third, multicollinearity was reduced as much as possible through correlation tests.

In this study, the features used are roughly divided into two categories: static land-surface indicators and built-environment indicators. The former mainly include surface and neighborhood greenness, surface albedo, proximity to water bodies and the coast, and topographic elevation and slope, which are used to characterize the cold-source pattern and basic energy-balance conditions; the latter include building height and density, degree of imperviousness, and distance to the city center, which are used to reflect the intensity of block morphology and locational differences [69,70,71]. Considering the existing mechanisms and data availability, a total of 15 indicators, including NDVI, NGVI, NDBI, building density, and albedo, were finally selected as model inputs. All indicators were calculated in GEE (web-based platform; accessed in October 2025) and ArcGIS Pro 3.1.5 based on a 30 m regular grid, and the indicators related to vegetation status were extracted independently based on summer and winter imagery, respectively [72]. The calculation formulas and physical meanings of each variable are shown in

Table 4. Details of the selected indicators.

Category	Indicator	Abbreviation	Attribute
Static-surface	Normalized Difference Vegetation Index	NDVI	Surface vegetation cover and evapotranspiration intensity
	Neighborhood Green Vegetation Index (270 m)	NGVI_270	Shading and cooling from nearby greenspaces
	Modified Normalized Difference Water Index	MNDWI	Presence of open water bodies
	Neighborhood Water Index (150 m)	NGWI_150	Influence of nearby water bodies within 150 m on local cooling
	Broadband surface albedo	Albedo	Fraction of short-wave radiation reflected by the surface
	Elevation	DEM	Topographic height
	Slope	Slope	Cold-air drainage
	Distance to coastline	Dist_sea	Position along sea–land breeze corridor
	Distance to major greenspace	Dist_green	Proximity to large urban parks/green corridors
Built environment	Building density (270 m)	BD_270	Planar compactness of buildings within 270 m neighborhood
	Mean building height	BH	Average building height
	Normalized Difference Built-up Index	NDBI	Imperviousness
	Distance to city center	Dist_CBD	Urbanization gradient
	Population density	PD	Residential population concentration
	Night-time light intensity	NTL	Human activity level and anthropogenic heat

.

3.4. Machine Learning and Statistical Analyses

3.4.1. LightGBM Modeling and SHAP-Based Interpretation

We employed LightGBM to model the nonlinear relationships between LST and predictors describing the built-environment and static surface attributes. Separate regression models were developed for summer and winter at the 30 m grid-cell level, and multicollinearity was controlled using variance inflation factors (VIF < 10). Samples were split into training and test sets using a 7:3 ratio. To manage model complexity and mitigate overfitting, we applied five-fold cross-validation with hyperparameter optimization, and evaluated predictive performance using R², RMSE, and MAE.

To improve interpretability, SHAP was used to attribute each grid-cell prediction to the additive contributions of individual features, yielding both the sign and magnitude of each contribution. We then computed global mean SHAP values for summer and winter to quantify predictor importance, and used feature value–SHAP response curves to characterize the nonlinear and threshold effects of individual predictors on LST.

3.4.2. Seasonal Net-Benefit Indices and Spatial Statistical Analyses

To jointly account for summer cooling gains and winter warming gains within a unified framework, we adopt the additive structure of the net-benefit concept widely used in health economics (Net Monetary Benefit, NMB) [73,74,75] and transform the LightGBM–SHAP outputs into a Seasonal Net-Benefit Index. For each grid cell $\mathrm{j}$ and feature $\mathrm{k}$, the SHAP values from the summer and winter models are denoted as ${\mathsf{\varphi }}_{\mathrm{j}\mathrm{k}}^{\mathrm{s}\mathrm{u}\mathrm{m}}$ and ${\mathsf{\varphi }}_{\mathrm{j}\mathrm{k}}^{\mathrm{w}\mathrm{i}\mathrm{n}}$, respectively (both measured as marginal contributions to LST, in °C). The seasonal net benefit (SNB) associated with this feature is defined as Equation (1):

$$\mathrm{S}\mathrm{N}{\mathrm{B}}_{\mathrm{j}\mathrm{k}}= \mathsf{\lambda }\hspace{0.17em}{\mathsf{\varphi }}_{\mathrm{j}\mathrm{k}}^{\mathrm{w}\mathrm{i}\mathrm{n}}-{\mathsf{\varphi }}_{\mathrm{j}\mathrm{k}}^{\mathrm{s}\mathrm{u}\mathrm{m}},$$

(1)

where $\mathsf{\lambda }=1.5$ represents, from an energy-consumption perspective, the relative weight assigned to winter warming gains compared to summer cooling gains. Summer cooling (${\mathsf{\varphi }}_{\mathrm{j}\mathrm{k}}^{\mathrm{s}\mathrm{u}\mathrm{m}}$ < 0) and winter warming (${\mathsf{\varphi }}_{\mathrm{j}\mathrm{k}}^{\mathrm{w}\mathrm{i}\mathrm{n}}$ > 0) jointly increase the ${\mathrm{S}\mathrm{N}\mathrm{B}}_{\mathrm{j}\mathrm{k}}$. For seasonal indicators such as NDVI and NGVI, SHAP values are extracted separately from the summer and winter models, matched by grid-cell ID, and then substituted into the above equation.

Finally, three indices are constructed by summing seasonal net-benefit values over different feature subsets: (1) Overall Seasonal Net-Benefit Index (O-SNBI): this summarizes the winter–summer integrated seasonal net benefit of each grid cell across all features, as defined in Equation (2);

$${\mathrm{O}-\mathrm{SNBI}}_j={{\sum }_{k\text{∈}K}\mathrm{SNB}}_{\mathit{jk}},$$

(2)

(2) Retrofit Seasonal Net-Benefit Index (R-SNBI): this aggregates seasonal net-benefit values only over short-term adjustable factors, where $C$ denotes the feature set consisting of NDVI, NGVI, and albedo, as defined in Equation (3).

$${\mathrm{R}-\mathrm{SNBI}}_j={{\sum }_{k\text{∈}C}\mathrm{SNB}}_{\mathit{jk}},$$

(3)

(3) Structural Seasonal Net-Benefit Index (S-SNBI): this aggregates seasonal net-benefit values only over static factors that are difficult to modify through engineering retrofits, where $S$ denotes the feature set excluding $C$, as defined in Equation (4).

$${\mathrm{S}-\mathrm{SNBI}}_j={{\sum }_{k\text{∈}S}\mathrm{SNB}}_{\mathit{jk}},$$

(4)

The three indices are classified into multiple benefit tiers using symmetric thresholds, and then overlaid with the LCZ map to quantify the area shares and index distributions of different built-type LCZ classes. Specifically, O-SNBI used symmetric breakpoints at ±1, ±3, and ±6; S-SNBI used ±4, ±2, 0, and 8; and R-SNBI used ±0.5, ±1.5, and ±2.5. This scheme uses 0 to preserve the direction of net benefit and establishes an interpretable mild–moderate–strong gradient within each index, enabling the identification of priority intervention neighborhoods that combine high structural suitability with strong retrofit potential.

4. Results

4.1. LCZ Classification Results

Using a semi-supervised deep learning framework, we classified the four central districts of Dalian into 14 LCZ classes. The overall accuracy is 0.93 and Cohen’s kappa is 0.91, indicating high classification performance.

The spatial heterogeneity of LCZ classes and their area shares are presented in Figure 3 and

Table 5. Area shares of LCZ classes in the study area.

Built types	LCZ1	LCZ2	LCZ3	LCZ4	LCZ5	LCZ6	LCZ8	LCZ10
Pixel count	5354	18,575	33,883	36,520	55,789	34,518	69,972	10,329
Share (%)	0.77%	2.69%	4.90%	5.28%	8.07%	4.99%	10.12%	1.49%
Natural types	LCZA	LCZB		LCZC	LCZD	LCZE	LCZF
Pixel count	177,969	49,514		115,875	12,078	41,578	29,227
Share (%)	25.75%	7.16%		16.76%	1.75%	6.02%	4.23%

, respectively. Built types account for 38.33% of the study area, whereas natural types account for 61.67%. Compact built types (LCZ 1–3) comprise 8.36% and are mainly concentrated in the urban core and along transportation corridors to the north. Open built types (LCZ 4–6) comprise 18.34%; mid- and high-rise areas are distributed in central and coastal residential zones, while low-rise areas are scattered toward the urban fringe. Large low-rise and heavy industry areas (LCZ 8 and LCZ 10) together account for approximately 11.61%, clustering in the northern logistics park and the eastern coastal industrial belt, respectively. Among natural types, dense forest (LCZ A) accounts for 25.75% and is primarily located in the western and southeastern mountainous areas, while LCZ B, LCZ C, LCZ E, and LCZ F occur as mosaics along the coastline, river valleys, and around large, developed land parcels.

4.2. Summer–Winter Contrasts in LST

Figure 4 presents the summer and winter LST distributions for each LCZ class and their departures from the citywide mean. In summer, LST in the built-up areas is markedly higher than that over natural surfaces. Dense built classes—such as compact mid-rise, large low-rise, and industrial LCZs—exhibit the largest positive departures, typically about 5 °C above the city average, with large low-rise showing the highest median LST. In contrast, open and low-rise residential classes are only slightly warmer than the mean. Natural LCZ classes, including water and dense forest, act as persistent cool areas, with LST generally below the city mean and exhibiting a tighter distribution. In winter, inter-class LST contrasts a narrow overall distribution: most built types are only slightly warmer than the city mean, and the negative departures of natural types weaken substantially. However, the relative ordering remains broadly consistent, with large low-rise (LCZ 8) areas showing the largest seasonal shift. The hottest dense built-up areas in summer remain relatively warm in winter, while water and forest continue to be cooler, although the warm–cool gradient is weaker than in summer.

4.3. Impacts of Indicators on Temperature in the Study Area

4.3.1. Overall Simulation Results

To analyze the impacts of indicators on LST and identify dominant factors, LightGBM regression models were trained separately based on summer and winter data. The summer model has an average R² of 0.92, an RMSE of 1.24 °C, and an MAE of 0.95 °C; the winter model has an average R² of 0.76, an RMSE of 0.88 °C, and an MAE of 0.68 °C. The summer model shows a higher R² and lower RMSE, indicating that summer LST is mainly controlled by underlying-surface characteristics; in winter, it is more easily disturbed by large-scale circulation and cold-air processes, and the explanatory power of static surface and built-environment factors is relatively weakened.

Figure 5 summarizes overall feature importance and effect directions using absolute SHAP values. In summer, greenness and proximity-to-green-space metrics dominate (Dist_green, NGVI_270, NDVI), indicating higher LSTs where large green spaces are distant and neighborhood greenness is low. Topographic and built-intensity variables (DEM, BD_270) and coastal proximity (Dist_sea) also contribute. High Dist_green/Dist_sea, BD_270, NDBI, and albedo values are associated with positive SHAP contributions (warming), whereas greenness and water metrics (NGVI_270, NDVI, MNDWI, NGWI_150) and terrain/coastal settings tend to show negative SHAP contributions (cooling).

In winter, urban-intensity and urbanization-gradient indicators become more prominent (NDBI, Dist_CBD, NGVI_270). Higher NDBI/BD_270/NTL values and greater distance from the sea generally correspond to positive SHAP contributions, while greenness and water metrics remain linked to negative contributions, albeit with weaker marginal cooling effects than in summer. Overall, summer LST is most sensitive to greenness and coastal–inland structure, whereas winter LST is more controlled by built intensity and urbanization gradients; the cooling effect of green space persists across seasons but is stronger in summer.

4.3.2. Nonlinear Effects of Static-Surface and Built-Environment Features on LST

Figure 6 illustrates representative nonlinear SHAP response curves for selected features in summer and winter; the full set of response curves is provided in Appendix A (Figure A1 and Figure A2). The x-axis denotes the feature value, and the y-axis shows the corresponding SHAP value, representing the marginal contribution of each indicator to LST across its value range.

Static-surface factors exhibit broadly consistent effect directions across seasons, but differ in threshold locations and effect magnitudes. Albedo shows similar responses in both seasons, with a common threshold around 0.15. DEM exhibits a markedly lower turning point in summer than in winter, while a unit change in DEM corresponds to a larger SHAP contribution in summer. Pixel-level NDVI displays opposite seasonal effects: a low NDVI value is associated with pronounced warming in summer but a weak cooling contribution in winter, and the summer transition point occurs at a higher NDVI value than in winter, consistent with local phenology and vegetation composition. Dist_green shows stronger variability and a clearer threshold in summer, indicating that areas farther from major green spaces are more prone to forming hot patches, whereas its winter effect is comparatively limited.

Built-environment indicators generally exhibit convex-shaped warming responses. For BD_270 and NDBI, SHAP values increase approximately monotonically in both seasons: effects are near zero below the median but amplify rapidly beyond mid-to-high quantiles, with a steeper increase in summer. Dist_CBD shows opposite seasonal patterns, reflecting an urban functional ring structure. BH has limited influence from low-rise to mid-rise ranges; when approaching or slightly exceeding the citywide mean height, it is associated with slight cooling in summer but shifts to weak warming in winter. PD shows broadly similar warming patterns across seasons, with SHAP contributions typically on the order of 0.1–0.4 °C, although the onset of positive contributions occurs at lower PD values in summer.

4.4. Seasonal Net-Benefit Indices and Their Spatial Patterns

4.4.1. Seasonal Net-Benefit Response Curves

Figure 7 presents feature-wise SNB response curves under λ = 1.5, together with mean-value trajectories for built-type LCZ classes. Most features show strongly nonlinear associations with SNB. Across built-type LCZ classes, curve shapes broadly follow the overall SNB trend, while amplitudes and threshold positions vary.

For static-surface features, SNB response curves can be summarized into three patterns: transitions from positive to negative, transitions from negative to positive, and fluctuations around zero. In general, the integrated gains from increasing vegetation at the pixel and neighborhood scales outweigh potential winter shading-related penalties. Specifically, the SNB increases almost monotonically with the NDVI and NGVI_270, with the NDVI shifting from a net penalty to a net benefit at approximately 0.65. Likewise, the NGWI_150 rises nearly linearly with an increasing water fraction. Location-related indicators display clear distance gradients: the SNB is positive at small Dist_sea and Dist_green, but drops rapidly with distance and becomes negative in the far-distance regime. Built-type LCZ mean trajectories for these static-surface factors align closely with the overall curves, with differences mainly emerging at the high- or low-SNB ends.

Built-environment features exhibit more pronounced nonlinear SNB patterns. Morphology and intensity variables show a clear multi-regime structure: for BD_270, the SNB is slightly positive at a low density, declines sharply as density increases and remains negative over a broad range, with a modest rebound at the extreme high-density tail. The SNB increases approximately monotonically with the NDBI and BH, switching from a net penalty to a net benefit around an NDBI ≈ −0.1 and a building height ≈ 12 m. Dist_CBD displays marked non-monotonicity: the SNB is lowest in intermediate rings and slightly positive in both the core and far suburbs (over approximately 8.4–25.5 km). Overall, built-type LCZ trajectories track the overall curves, and class differences become substantially larger only at extreme height/density levels and within high-NDBI segments.

4.4.2. Spatial Distribution of Seasonal Net-Benefit Indices

Figure 8 maps the seven-tier classifications of the three Seasonal Net-Benefit Indices. For the O-SNBI, approximately 60% of grid cells fall within −3 to 3 (weak penalty to weak benefit), while strongly penalized and strongly benefited cells together account for less than ~3%. Medium-to-high O-SNBI clusters occur along the coastline, in mountainous–hilly areas, and near major green corridors. In contrast, low-lying, highly impervious, and vegetation-sparse continuous built-up areas are mostly near-neutral or mildly penalized, and strongly penalized patches concentrate in the northern sector characterized by fragmented development and the co-location of industry and bare land. In terms of LCZ composition, strong penalty tiers are mainly associated with LCZ 8, LCZ 10, and parts of LCZ 3, whereas strong benefit tiers are dominated by LCZ A–C.

The S-SNBI captures inherent climatic advantages/disadvantages driven by topography, coastal–inland setting, and the overarching development structure. Its spatial pattern closely mirrors that of the O-SNBI: high values concentrate in coastal belts, mountainous green areas, and potential ventilation corridors, indicating favorable baseline conditions. The low-S-SNBI areas are more widespread; notably, some locations show positive R-SNBI values despite persistently low S-SNBI values, suggesting that short-term greening/albedo measures may help, but long-term limitations from terrain and location remain.

The R-SNBI represents short-term seasonal gains achievable through adjustments to vegetation and albedo. Its distribution is more concentrated, with most grid cells within −2 to 2. High-R-SNBI patches tend to appear at the urban fringe or on well-vegetated slopes, whereas low-R-SNBI cells are embedded in high-density built-up areas—especially compact high-rise zones and areas adjacent to industrial land—highlighting priority targets for near-term climate-adaptation retrofits. Strong penalty tiers are mainly linked to LCZ 1–3 and LCZ 8/10, while strong benefit tiers remain dominated by LCZ A–C.

4.4.3. Seasonal Net-Benefit Profiles of Built-Type LCZ Classes

Based on the mapped patterns of the three indices, we compiled LCZ-wise distributions as boxplots (Figure 9), with an emphasis on eight built-type LCZ classes (LCZ 1–6, LCZ 8, and LCZ 10). Overall, the median O-SNBI values for built LCZ classes are generally below zero, yet their spreads and extremes vary markedly, indicating combined differences in baseline structural conditions and retrofit-related opportunities.

LCZ 1 and LCZ 2 have O-SNBI medians close to zero with relatively narrow interquartile ranges, implying limited within-class variability. Their median S-SNBI is about 1–2, the highest among the built types, whereas R-SNBI is distinctly negative, suggesting favorable baseline conditions but substantial room for improvement in current greenness-related measures. For LCZ 4 and LCZ 5, the three indices show broadly similar distributions clustered around (and slightly below) zero; O-SNBI and S-SNBI are modestly lower than those of LCZ 1–2, and R-SNBI is near zero, placing them in a mid-range category for both overall seasonal net benefit and retrofit potential.

LCZ 3, characterized by contiguous low-rise blocks, lies in the lower range for both the O-SNBI and S-SNBI, indicating a weaker overall seasonal net benefit and structural suitability. Its R-SNBI median is slightly higher but with larger dispersion, suggesting limited average gains from greening, while retrofit potential differs substantially across neighborhoods. LCZ 8 exhibits the lowest O-SNBI median, with the lower quartile extending into the strong-penalty tier; the S-SNBI is also clearly negative, whereas the R-SNBI—though still negative—is higher than the S-SNBI, implying that the deficit is driven primarily by unfavorable structural conditions rather than a simple lack of greenness. In contrast, LCZ 10 shows a lower median R-SNBI than S-SNBI, suggesting greater scope for improvement through blue–green enhancement and surface-property retrofits, and therefore, merits priority intervention.

5. Discussion

5.1. Winter and Summer LST Characteristics and Urban Fabric

Within the LCZ framework, summer and winter LST exhibit broadly stable spatial patterns with clear seasonal contrasts. Built LCZ classes are consistently warmer than natural classes in both seasons, while bare ground shows an LST signature comparable to compact low-rise areas. In summer, highly impervious and dense classes—compact mid-/low-rise, large low-rise, and heavy industry—form the most pronounced heat-island cores, consistent with prior studies [76]. In winter, inter-class thermal gradients are strongly compressed [77], yet LCZ2 and LCZ10 remain relatively warm, suggesting that heat storage and delayed release persist across seasons [78]. Open built classes are only slightly above the city mean in summer and return to near-mean conditions in winter. Among the natural classes, forests/green spaces (LCZ A–D) and water bodies maintain the lowest LST in both seasons, forming persistent cool belts along mountainous and coastal settings [59,79]. By contrast, bare land and sparsely vegetated classes lack shading and evapotranspiration, resulting in a higher summer LST and weak winter cooling [80]. Overall, seasonal LST contrasts are robust, but their intensity and practical relevance differ: summer highlights mitigation targets and cooling sources, whereas winter presents a weaker heat-island signal with potential implications for heating demand.

5.2. Seasonal Effects of Selected Features and the Seasonal Net-Benefit Index

Selected features exhibit both shared directional effects on LST and a seasonal shift in dominant controls. In summer, greenness metrics, proximity to major green/water bodies, distance to the sea, and terrain variability contribute most, consistent with cooling from shading and evapotranspiration as well as enhanced coastal ventilation and moisture influences [69,81,82]. In winter, the importance of built-environment intensity indicators (e.g., NDBI and BD) increases, likely reflecting stronger anthropogenic heat and energy-use signals associated with dense development and district heating, in line with Chen et al. [83]. Overall, summer LST is more sensitive to the spatial configuration of cooling sources, whereas winter LST is more influenced by development intensity and energy use; notably, highly impervious, low-greenness areas far from cooling sources tend to exhibit warming in both seasons [84].

To synthesize these seasonal mechanisms for more targeted planning, we construct a Seasonal Net-Benefit Index (SNBI) based on SNB, enabling a quantitative comparison of combined feature effects. Greenness-related features exhibit threshold ranges where SNB shifts from a net penalty to a net benefit [85], while proximity to water bodies or the coastline is consistently associated with a positive SNB [86]. In contrast, building height, density, and imperviousness often show inverted U-shaped SNB responses, suggesting that moderate development intensity can yield more balanced seasonal performance. Together, the proposed framework offers a new lens for interpreting nonlinear seasonal coupling between environmental indicators and LST and for identifying neighborhoods with substantial regulation potential.

5.3. Limitations

This study has several limitations. First, the analysis was implemented on a 30 m regular grid; the robustness of the feature-importance rankings and SNBI estimates remains to be tested through multi-scale sensitivity analyses. Second, the Seasonal Net-Benefit Index adopts the additive structure of the NMB framework from health economics, which places benefits and costs on a common scale to enable summation and comparison [87]. Accordingly, we set the winter-weight parameter to λ = 1.5, guided by evidence from heating/cooling degree-day studies showing that, in mid- to high-latitude cold regions, heating demand exhibits a larger magnitude and temperature sensitivity than cooling demand for a unit temperature change [3,88]. Nonetheless, treating λ as a fixed constant is a simplifying assumption: it likely varies with region, building type, and climatic year. Future work could therefore estimate the plausible ranges of λ and quantify its uncertainty via parameter calibration and sensitivity analysis [89]. Finally, our predictors focus on relatively static surface and built-environment attributes and do not explicitly account for dynamic processes (e.g., wind fields) or household energy-use behaviors. Given that the empirical evidence is derived from a single cold coastal city and a limited set of years, the transferability of our findings to other climate regimes, urban morphologies, and longer-term climate contexts requires further validation using multi-city, multi-year comparisons and robustness checks.

5.4. Future Work

Future work will focus on three directions. First, we will test the robustness of the seasonal net-benefit patterns and the SHAP-based interpretations through multi-scale sensitivity analyses and alternative spatial supports. Second, we will calibrate the winter-weight parameter (λ), estimate plausible ranges, and quantify uncertainty using independent evidence such as proxies of heating and cooling demand, so that the index can better support planning interpretation. Third, we will extend the framework by incorporating additional dynamic factors such as wind and temporal variability, and evaluate transferability through applications across multiple cities and years.

6. Conclusions

In this study, we focus on the four central districts of Dalian and implement the analysis on a 30 m grid within the LCZ framework. We first produce an LCZ map using a semi-supervised learning pipeline, then apply machine-learning models to quantify nonlinear relationships between multi-source predictors and LST, and finally adopt the additive NMB structure to synthesize summer and winter effects and characterize their trade-offs through the SNBI. The main conclusions are as follows:

• Semi-supervised learning enables high-accuracy LCZ mapping. With self-supervised pretraining followed by fine-tuning using limited labeled samples, we achieve an overall accuracy of 0.93, capturing the morphological gradient across mountainous areas, coastal zones, and dense built-up districts.
• Multi-source predictors exhibit seasonally consistent yet asymmetric nonlinear effects on LST. Dense and highly impervious built LCZ classes remain pronounced heat-island cores year-round, whereas blue–green spaces and mountain–sea corridors act as persistent cooling sources. Summer LST is more sensitive to cooling-source factors (e.g., greenness configuration), while winter importance shifts toward development-intensity indicators (e.g., NDBI). Many predictors display clear threshold behaviors, consistent with prior evidence.
• The seasonal net-benefit framework integrates summer and winter effects into a unified metric system and, combined with LCZ classes, reveals neighborhood-specific seasonal performance. Compact/large low-rise and heavy-industry-related built types show consistently poor two-season outcomes dominated by structural penalties, whereas mid- to high-rise neighborhoods—despite only moderate baseline conditions—can still improve seasonal net benefits through near-term retrofit measures, such as increasing greenness and modifying surface properties.

Overall, building on established LCZ–thermal environment insights, this work highlights the value of explicitly accounting for summer–winter trade-offs via a seasonal net-benefit perspective. Although Dalian serves as a single case, the findings offer practical implications for fine-scale climate adaptation and urban renewal in cold-region coastal and mountainous cities, and provide a transferable analytical framework for future multi-city, multi-climate investigations of seasonal thermal trade-offs.