Introduction

Urbanization and population growth have transformed surface landscapes, triggering a series of ecological and environmental issues including reduced biodiversity, disrupted atmospheric balance, and environmental pollution1,2,3. These problems further intensify the Urban Heat Island (UHI) effect, a phenomenon in which urban areas experience significantly higher temperatures than the surrounding suburban regions. Current research on UHI primarily focuses on its associations with natural and anthropogenic factors. Key influencing variables include land use/land cover types4, urban form metrics5, Normalized Difference Vegetation Index (NDVI)6, landscape pattern configuration7, and urban green infrastructure8,9. Research scales have also expanded from individual cities to large and medium-sized urban agglomerations at the basin and regional levels10,11,12,13. However, against the backdrop of globally constrained urban land expansion, the core scientific question — how green space morphology regulates heat island intensity — remains insufficiently addressed. Existing studies mostly emphasize landscape composition rather than functional morphology and rely on linear analysis methods, which struggle to capture the complex nonlinear relationships within urban thermal systems. This research gap urgently requires attention.

Green spaces serve as critical carriers for mitigating UHI, and the regulatory role of their spatial patterns on thermal environments is widely recognized — increased green space coverage is typically directly associated with reduced heat island intensity14. Traditional green space cooling strategies have focused on expanding area, adjusting tree species structure, or enriching landscape types15. Nevertheless, the scarcity of urban land resources and competing demands from housing, infrastructure, and agricultural development render unrestricted green space expansion impractical. This practical dilemma has driven a paradigm shift in green space planning from “quantity-oriented’ to “configuration optimization”, and the spatial morphological layout of green space elements has gradually become the core of thermal environment regulation16,17,18,19. Recent studies have further revealed the synergistic effects of green space morphology in UHI mitigation: Gupta and De utilized geospatial technologies to demonstrate that strategic use of urban green spaces, enhanced vegetation continuity, and optimized landscape geometric morphology can significantly reduce heat island intensity, providing a feasible pathway for improving thermal environments in high-density cities20. Lin et al. confirmed through machine learning that the morphological pattern of green spaces has a far greater impact on UHI than traditional landscape indices, and argued that core green space patches exhibit superior cooling efficiency compared to fragmented small isolated patches21.Simultaneously, Gupta and De clarified that the area, morphology, and spatial connectivity of urban blue spaces influence cooling effects22. These studies indicate that focusing solely on green space area or single landscape indices can no longer meet the needs of precise regulation, highlighting the urgent need to integrate green space morphological characteristics into the UHI research framework.

Existing research on the relationship between green spaces and UHI primarily relies on landscape pattern indices (e.g., Percentage of Landscape (PLAND), Mean Patch Area (AREA_MN), Aggregation Index (AI), Landscape Shape Index (LSI)16,17,18,19 to characterize green space composition and structure. Although these studies have summarized certain patterns (e.g., PLAND and AREA_MN are negatively correlated with Land Surface Temperature (LST), while Patch Density (PD) and Edge Density (ED) are positively correlated with LST23, significant limitations remain. On one hand, research results exhibit substantial regional and seasonal heterogeneity: for example, AI and LSI are negatively correlated with LST at the regional scale but show no significant association at the local scale, and some indices even display seasonal reversal (negative correlation in summer, positive correlation in winter), greatly increasing the difficulty of result interpretation. On the other hand, traditional landscape indices are inherently abstract quantitative metrics that emphasize compositional characteristics at the statistical level rather than spatial morphology, failing to distinguish landscape units with key ecological functions such as core green spaces and linear corridors24. With the intensification of green space fragmentation during urbanization, these indices also struggle to capture the degradation of cooling functions caused by morphological fragmentation — this flaw underscores the urgent need for a more refined, function-oriented morphological analysis framework.

Morphological Spatial Pattern Analysis (MSPA) is a landscape analysis method based on mathematical morphology25 that can accurately quantify the area, morphology, and connectivity of green space patches. Currently, MSPA has been widely applied in forest fragmentation assessment, landscape connectivity analysis, and ecological network construction26,27,28,29, but its application in UHI research remains limited30. Additionally, most MSPA-based studies on green space-UHI relationships rely on linear analysis methods or traditional spatial regression models31,32,33,34, which cannot effectively capture the inherent nonlinear relationships in urban thermal systems35,36. In recent years, the application of machine learning techniques (e.g., Random Forest (RF), Support Vector Machine (SVM)) in UHI research has provided a new approach — these models can more accurately simulate nonlinear dynamic processes by integrating large-scale LST data with multi-source auxiliary variables37,38,39.

In view of the aforementioned research gaps, this study takes Dali City as the research area and systematically explores the coupling mechanism between green space morphology and UHI by integrating MSPA and machine learning methods. Compared with existing research, the innovations of this study are mainly reflected in three aspects: (1) it systematically analyzes the impacts of seven MSPA-classified green space morphological types on UHI intensity, realizing a transformation from “compositional research” to “functional morphological research”; (2) it adopts Random Forest to construct a nonlinear relationship model, overcoming the limitations of linear and spatial regression models; (3) it compares five analytical frameworks (Ordinary Least Squares (OLS), Spatial Error Model (SEM), Spatial Lag Model (SLM), Geographically Weighted Regression (GWR), RF) to verify the superiority of machine learning in capturing the complex dynamic relationships between green spaces and UHI. This study focuses on addressing three core scientific questions: (1) What are the differences between the impacts of green space morphological factors and non-morphological factors on UHI intensity in Dali City? (2) Among the seven MSPA green space morphological types, which ones have the most significant regulatory effects on UHI, and what are their intrinsic mechanisms? (3) Among the five model types, which one can optimally explain the relationship between green space morphology and UHI, and what are its performance advantages? By answering these questions, this study aims to provide new insights for UHI mitigation and promote urban sustainable development.

Materials and methods

Overview of the study area

Dali City is located in the western part of Yunnan Province, in the middle of Dali Bai Autonomous Prefecture, at longitude 99°58’−100°27’ East and latitude 25°25’−25°58’ North, covering an area of 1,815 km² (Fig. 1). The city’s topography is complex; the terrain is high in the west and low in the east, with an average elevation of 2090 m. The landscape is mainly mountainous, with mountains accounting for 70.5% of the total area. It is a subtropical highland monsoon climate type, with an average temperature of 16℃. Dali City has an overall forest area of 808.9 km2, a forest coverage rate of more than 46%, an urbanization rate of 71.22%, and a built-up green space rate of 35.35%. According to studies, the temperature in Dali City increased by 0.58 °C from 2000 to 201940,41, and the thermal environment problem has become increasingly prominent, requiring targeted measures for improvement.

Fig. 1
Fig. 1The alternative text for this image may have been generated using AI.
Full size image

Location of the study area.

Data sources

The remote sensing images and digital elevation models involved in this study were obtained from Geospatial Data Cloud (https://www.gscloud.cn/, accessed on 10 June 2024) with a spatial resolution of 30 m. The four periods of remote sensing imageries with fewer clouds in the study area were selected to minimize the influence of weather conditions on the inversion results. They are 16 January 2020, 20 March 2020, 27 August 2020, and 28 September 2020, with strip number 131/42 and spatial resolution of 30 m, respectively. In addition, land use data were obtained from Esri Land Cover (https://www.arcgis.com/, accessed on 10 June 2024) with a spatial resolution of 10 m. The green space data were obtained through the interpretation of Google Earth high-resolution imagery, with an image class of 19 and a spatial resolution of 0.54 m, green space identification results were evaluated by randomly selecting 200 sample points as test samples. The overall accuracy was 91.12% with a Kappa coefficient of 0.85, which meets the research needs. The Normalized Difference Vegetation Index (NDVI) and Normalized Difference Built-up Index (NDBI) were calculated by calculating the average value of the four periods of remote sensing imageries, and the building height data were taken from Wu42 shared dataset with 10 m spatial resolution. Road data were obtained from Digital Earth Open Platform (https://open.geovisearth.com, accessed on 11 June 2024), and population data were collected from World Pop (https://hub.worldpop.org/, accessed on 12 June 2024) with a resolution of 100 m. The study area boundaries were obtained from the Dali City Territorial Spatial Master Plan (2021–2035) (https://www.yn.gov.cn/, accessed on 12 June 2024) (Fig. 2). To simplify the computation, land use data, building height data, and population data were reclassified, and their spatial resolution was adjusted to 30 m × 30 m. All images and data were unified into the WGS_1984_UTM_Zone_47N coordinate system.

Fig. 2
Fig. 2The alternative text for this image may have been generated using AI.
Full size image

Spatial data used in the study.

Methodology

Figure 3 shows the methodological framework of this study. Firstly, eCognition software was used for processing Google’s high-definition images in order to collect land use data. Guidos Toolbox was used to analyze the morphological spatial pattern analysis (MSPA) to make seven different types of landscapes. Secondly, the LST, NDBI, and NDVI data for Dali City were acquired using the four periods of remote sensing data. Thirdly, OLS, SEM, SLM and GWR models were employed to test the relationship between MSPA types and LST. Fourthly, the RF model was utilized to create the benchmark model and recommended model for investigating the nonlinear relationship between the MSPA types and LST in green space. Finally, based on the above study, an idea for green space planning is suggested.

Fig. 3
Fig. 3The alternative text for this image may have been generated using AI.
Full size image

Methodologic framework of this study.

Land surface temperature inversion

In this study, the radiative transfer equation was used for the surface temperature inversion13, and ENVI software was used to obtain the LST of Dali City. In order to verify the accuracy of the inversion results, the LST was compared with the MODIS (Moderate-Resolution Imaging Spectroradiometer) LST data of the same period, and the results showed that the root-mean-square error of this LST product is less than 2 K. The inversion results can meet the study needs, and the data are usable. The equation for calculating the surface temperature from the TIRS 10 band of Landsat 8 is as follows41:

$$\:{\text{L}}_{\lambda\:}=\left({M}_{L}*{Q}_{Cal}\right)+{\text{A}}_{L}$$
(1)

where Lλ is the top of atmosphere radiance (W⋅m− 2 ⋅sr− 1 ⋅µm− 1); ML is the band-specific multiplicative rescaling factor (0.0003342 for TIRS band 10); QCal is the quantized calibrated pixel value; AL is the band specific additive rescaling factor (0.1 for TIRS band 10.7).

$$\:{L}_{T}=\frac{{L}_{\lambda\:}-{L}_{\mu\:}^{\uparrow\:}-\tau\:(1-\epsilon\:){L}_{d}^{\downarrow\:}}{\tau\:\epsilon\:}$$
(2)

where LT denotes the blackbody radiance derived from Planck’s law; Lλ represents the spectral radiance of thermal infrared radiation received by satellite sensor (W⋅m− 2⋅µm− 1⋅sr− 1); ε is the land surface emissivity; τ is the atmosphere transmittance; \(\:{L}_{\mu\:}^{\uparrow\:}\) and \(\:{L}_{d}^{\downarrow\:}\) are the atmospheric upwelling and downwelling radiance43. τ, \(\:{L}_{\mu\:}^{\uparrow\:}\), and \(\:{L}_{d}^{\downarrow\:}\) can be obtained by inputting the date, time, and central latitude and longitude of the remote sensing image into the atmospheric correction parameter calculator provided by NASA (https://www.nasa.gov/, accessed on 10 June 2024).

$$\:{\text{T}}_{s}=\frac{{\text{K}}_{2}}{\text{ln}\left(\frac{{\text{k}}_{1}}{{\text{L}}_{\lambda\:}}+1\right)}-273.15$$
(3)

where Ts is the LST (◦C); K1 and K2 are the radiance constants, for TIRS band 10, K1 is 774.89 (W⋅m− 2 ⋅sr− 1 ⋅µm− 1) and K2 is 1321.08 K44:

The UHI is calculated for surface temperatures using the following Equations45

$$\:\text{U}\text{H}\text{I}=\frac{{T}_{S}-{T}_{M}}{SD}$$
(4)

Where: Ts = Land Surface Temperature (◦C).

Tm = Mean of the LST.

SD = Standard deviation.

Estimation of NDVI and NDBI

Normalized Difference Vegetation Index (NDVI) is a remote sensing metric used to monitor the state of vegetation growth and vegetation cover. It reflects the lushness of vegetation by calculating the difference in reflectance between the near-infrared (NIR) and red (R) bands, with values between − 1 and 1. Negative values usually indicate that the ground is covered with objects that are highly reflective of visible light, such as water or snow; values close to 0 may indicate the presence of rocks or bare soil, and positive values indicate the presence of vegetation, with larger values indicating more lush vegetation. The formula is as follows6:

$$\:NDVI=\frac{NIR-R}{NIR+R}$$
(5)

where NIR, R are the surface reflectance of band 5 and band 4, respectively.

Normalized Difference Built-up Index (NDBI) is an important indicator to understand the UHI effect; the index mainly uses the ratio of the mid-infrared (MIR) band to the near-infrared (NIR) band to identify urban built-up land (mostly impervious surfaces), which can reflect the information of the built-up land more accurately. The larger the value, the higher the proportion of built-up land and the higher the density of buildings, and the value ranges from − 1 to + 1, with positive values indicating built-up areas and negative values indicating water bodies and vegetation. The formula is46:

$$\:NDBI=\frac{\rho\:MIR-\rho\:NIR}{\rho\:MIR+\rho\:NIR}$$
(6)

where ρMIR and ρNIR represent the reflectance values for the mid-infrared and near-infrared bands of Landsat images.

Morphological Spatial pattern analysis

Morphological Spatial Pattern Analysis (MSPA) is an image processing method that measures, identifies, and segments the spatial patterns of raster images based on mathematical morphology principles, such as erosion, dilation, opening, and closing47. The method mostly relies on land use data, by extracting specific land use types as the foreground and other land types as the background, and then using Guidos software to classify the foreground into seven types according to the morphology. In this study, the land use was converted into binary TIFF data with a spatial resolution of 30 m. The green space was assigned as foreground data with a value of 1, and the non-green space was assigned as background with a value of 2. Secondly, MSPA classification was performed using Guidos Toolbox with the following parameters: (1) input data: binary TIFF of green space (30 m resolution), with green space as foreground (value 1) and non-green space as background (value 2); (2) neighborhood rule: 8-neighborhood (to capture fine-scale connectivity); (3) edge width: 1 pixel (corresponding to 30 m in field, determined by the average size of green patches in Dali); (4) minimum core area: 0.01 km² (to exclude small noise patches). A total of 150 sample points (≥ 20 per morphology type) were randomly selected for accuracy verification, with an overall accuracy of 89.5% and Kappa coefficient of 0.82. The core (92.3%) and edge (90.1%) had the highest accuracy, while the perforation (81.5%) had lower accuracy due to small area, which still met the research needs. and the MSPA analysis module was applied to obtain 7 types of morphology types: core, edge, branch, islet, bridge, loop, and perforation48 (Table 1). Finally, obtain the 7 types of landscape morphology type maps.

Table 1 Definitions of the seven morphological categories in MSPA.
Full size table

Regression analysis

Ordinary least squares (OLS)

The OLS model, based on linear predictive models, is mostly used for parameter estimation in linear regression7. The OLS model assumes that observations are independent of one another, that errors are normally distributed for all variables, that the dependent and independent variables have a linear relationship, and that the dependent variable’s variance is consistent across independent variable levels. The equation is as follows31:

$$\:\gamma\:={b}_{0}+{b}_{1x1}+{b}_{2x2}+\cdots\:+{b}_{kxk}$$
(7)

where b0 is the constant term and b1, b2 …, bk are the partial regression coefficients. The partial regression coefficient is the value of y that changes for every unit change in this independent variable when all other independent variables are fixed.

Spatial error model (SEM)

Because SEM is similar to the serial correlation problem in time series, it is also known as the spatial autocorrelation model (SA). When performing spatial regression model analysis, the independent variables may have endogeneity problems, and the traditional OLS estimation method will lead to biased or invalid coefficient estimates, in which case it is necessary to deal with the error term in the formula so as to use the spatial error model to perform regression analysis. The formulas are as follows32:

$$\:Y=X\beta\:+\epsilon$$
(8)

where is the dependent variable; is the explanatory variable; is an estimated independent coefficient; ε is a random error term.

Spatial lag model (SLM)

SLM is mainly used to describe and predict spatial correlations and spatial dependencies in economic, environmental and social phenomena. When the spatial data is irregularly shaped, the spatial matrix can be used to represent the lag included, modeled as follows32:

$$\:y=\rho\:Wy+X\beta\:+\epsilon$$
(9)

where is the dependent variable; ρ is the spatial regression coefficient; is the spatial weight matrix; is the explanatory variable; is an estimated independent coefficient; ε is a random error term.

Geographically weighted regression (GWR)

GWR is a spatial analysis technique widely used in geography and related disciplines31. At the heart of the GWR model is the spatial weight matrix, which defines weights based on the spatial distance between observations32. Common spatial weight functions include the distance threshold method, the inverse distance method, and the Gauss function method. In this study, the Akaike (AICc) information criterion was used to determine the range of the kernel, and FIXED was used to determine the distance. The model is as follows33:

$$\:{y}_{i}={\beta\:}_{0}\left({u}_{i},{v}_{i}\right)+{\sum\:}_{j=1}^{k}{\beta\:}_{j}\left({u}_{i},{v}_{i}\right)\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:$$
(10)

where\(\:\:\left({u}_{i},{v}_{i}\right)\) denotes the spatial coordinates of sample point i;\(\:{\:\beta\:}_{0}\left({u}_{i},{v}_{i}\right)\) and\(\:{\:\beta\:}_{j}\left({u}_{i},{v}_{i}\right)\) denote the regression coefficient and the jth regression parameter at point i, respectively;\(\:\:{\epsilon\:}_{i}\) is the independent error term that conforms to normal distribution (mean 0, variance\(\:{\delta\:}^{2}\)).

Random forest (RF)

The RF model is a machine learning method that uses multiple decision trees to train, classify, and predict sample data, then aggregates their predictions to obtain a final result. The advantages of the RF model are that it can handle high-dimensional data without feature selection or dimensionality reduction, it is robust to noise and outliers in the input data, and it can resist noise interference to a certain extent. Meanwhile, since the training process of individual decision trees is independent of each other, RF is well suited for parallelized computation, which can significantly improve the computational efficiency and evaluate the importance of each feature to the model49. In this study, LST was set as a predictor, and other factors were used as explanatory variables. Through repeated experimental comparisons, the number of trees is 350, the minimum leaf size is 5, and the maximum tree depth is 35, and 30% of the data are used for verification.

Model accuracy validation

To validate the model’s accuracy, this study evaluates the accuracy of OLS, SEM, SLM, GWR, and RF models using R2, MSE, and RMSE50. The coefficient of determination (R2) is the proportion of the degree of variation of the independent variables to the total degree of variation; the larger the R2, the better the model’s fitting ability. Mean Squared Error (MSE) is a measure of the difference between the predicted value and the true value of the model; the smaller the value of MSE, the closer the predicted value of the model to the actual value, i.e., the better the performance of the model. Root Mean Squared Error (RMSE) is the arithmetic square root of the standard error, which is used to measure the deviation between the predicted value and the real value, and the smaller the value of RMSE, the better the predictive ability of the model. The specific formula is as follows51:

$$\:R^{2} = \frac{{\sum {\:_{{i = 1}}^{m} } \:\left( {\hat{y} - \mathop y\limits^{ - } } \right)^{2} }}{{\sum {\:_{{i = 1}}^{m} } \:\left( {y_{i} - \mathop y\limits^{ - } } \right)^{2} }}$$
(11)
$$\:\text{MSE=}\frac{1}{m}\sum\:_{i=1}^{m}\:({y}_{i}-{\widehat{y}}_{i}{)}^{2}$$
(12)
$$\:\text{RMSE=}\sqrt{\frac{1}{m}\sum\:_{i=1}^{m}\:({y}_{i}-{\widehat{y})}^{2}}$$
(13)

where \(\:\widehat{y}\) is the predicted value of the model; \(\:\stackrel{-}{y}\) is the mean of the true value; \(\:{y}_{i}\) is the true value; m is the sample size.

Results and analyses

The spatial pattern of LST

As can be seen in Fig. 4, the high-value areas of heat island intensity in Dali city are mainly concentrated in the southwest and north of the city, which is due to the fact that the area is mainly industrial parks, and the Dali Hub Bus Terminal and Airport are also located here, with low vegetation cover and mostly bare ground, which exacerbates the heat island effect in the city. In the west and at the edge of Erhai Lake is a low-value area; this is due to the high proportion of green space in this area, and the water body also reduces the UHI intensity.

Fig. 4
Fig. 4The alternative text for this image may have been generated using AI.
Full size image

Intensity of UHI effect in study area.

Spatial characteristics of green space morphological patterns

From the perspective of the spatial distribution of the seven types of MSPA in the study area (Fig. 5; Table 2), core accounted for the highest proportion, accounting for 30.58% of the total green space area, mainly composed of park green space, affiliated green space and protective green space, and was mostly distributed in the edge of the study area. Edge and islet have a high proportion and are distributed in most areas of the study area. Branch and bridge are widely distributed and serve as connecting the core area, accounting for 9.71% and 7.70% of the green space area, respectively. Perforation and loop were mainly distributed in areas with low density, accounting for 1.22% and 3.46%, respectively.

Fig. 5
Fig. 5The alternative text for this image may have been generated using AI.
Full size image

Morphological spatial pattern of green space in Dali City.

Table 2 Proportion of the MSPA classes.
Full size table

Linear regression analysis of UHI with factors

Due to the significant differences in the spatial distribution of UHI intensity, this study used multiple linear regression analysis to explain the linear relationship between UHI and the factors. Regression analyses were done with the UHI effect as the dependent variable and altitude, NDVI, NDBI, average building height, population, roads, and the seven MSPA types as the independent variables. Linear regression analyses were performed using Origin 2022 software to obtain the model results and as a benchmark for comparison with other models (Table 3). The results show that R2 is 0.78, which indicates that the model can explain 78% of the variation of the dependent variable, and the model fit is good. As shown by the Pearson’s correlation coefficient (Table 4), the UHI intensity is significantly negatively correlated with altitude, NDVI, core, average building height, islet, bridge, branch, and edge. The higher these influencing factors (e.g., altitude, NDVI), the lower the LST. This is related to the geographical location of the study area: the local altitude is high, mountain vegetation is abundant, resulting in a high NDVI value, and the transpiration and cooling effects of vegetation contribute to a low LST. UHI intensity was positively correlated with NDBI, population, road, perforation and loop. These factors are not straightforward to dissipate heat; they increase the LST. Among the green space morphology, the islet had the highest correlation coefficient with UHI at −0.033, followed by the core area at −0.016. This is because there is a lot of construction land in the study area, and the green space is fragmented, forming many small islands. Therefore, mitigation of UHI should be based on local altitude, vegetation type, and rational urban space planning.

Table 3 OLS, SLM, SEM and GWR model analysis results.
Full size table
Table 4 Linear relationships between UHI intensity and influencing factors.
Full size table

Spatial analysis of UHI and factors

According to the Pearson correlation test, there is no multicollinearity problem in linear regression, but there is a spatial autocorrelation problem. In this study, Moran’s I (error) = 0.237, p = 0.000 for the linear regression model, so spatial autocorrelation analysis was performed. Referring to previous studies32, SLM and SEM were selected for fitting analysis in this study, and the results are shown in Table 3. By comparing the results of SLM and SEM, the R2 of SLM and SEM are 0.84 and 0.83, respectively, and the AIC and SC of SLM are smaller than those of SEM and OLS, but the R2 increases, which indicates that the SLM model fits the global structure of the data better and it can explain the spatial autocorrelation problem between the UHI and the factors well. Therefore, SLM was used in this study for the analysis. SLM assumed that the effects of altitude, NDVI, NDBI, average building height, population, roads, and the seven types of green space patterns on UHI were the same across the different study areas. However, UHI is influenced by multiple factors, such as land use, altitude, and human activities, and the effects of these factors vary in different locations. In order to deeply analyze the spatial variability of these factors on UHI, GWR analysis was used in this study. The results show that the AIC of the GWR model is 1716.47, which is smaller than that of the OLS model (1737.65) (Table 3). Meanwhile, the R2 is 0.79, which is larger than that of the OLS model (0.78), which indicates that the GWR model fits better. This indicates that the GWR model can better deal with the spatial heterogeneity of green space morphology and the UHI effect.

Non-linear regression analysis of UIH with factors

Focusing on the nonlinear relationship between green space morphology and land surface temperature, this study employed a random forest (RF) algorithm to construct a predictive model. Compared with the traditional ordinary least squares (OLS) regression, RF significantly improved the fitting capacity for complex green space morphological factors through an ensemble learning mechanism of decision trees. Model comparison experiments showed that the benchmark RF model achieved a coefficient of determination (R² = 0.84) and a mean squared error (MSE = 2.34), representing a 7.7% increase in goodness-of-fit and a 28% reduction in error compared to the OLS model (R² = 0.78, MSE = 3.21) (Table 5). The optimized RF proposed model, after incorporating green space spatial patterns, further enhanced R² to 0.88 and reduced MSE to 2.06, corresponding to a 36% error decrease relative to OLS. These results fully validate the necessity and effectiveness of nonlinear algorithms in such modeling. In contrast to spatial error model (SEM), spatial lag model (SLM), and geographically weighted regression (GWR), the RF model demonstrated optimal performance in key metrics: R² =0.88, MSE = 2.06, and root mean squared error (RMSE = 1.44). Specifically, compared with OLS (R² = 0.78, MSE = 3.21, RMSE = 1.79), SEM (R² = 0.83, MSE = 3.05, RMSE = 1.50), SLM (R² = 0.84, MSE = 3.02, RMSE = 1.75), and GWR (R² = 0.79, MSE = 2.30, RMSE = 1.52), RF exhibited remarkable advantages in interpreting the relationship between green space morphology and urban heat island (UHI) effect, enabling more accurate characterization of their complex nonlinear relationships.

Table 5 Quantitative metrics of the benchmark models and proposed models.
Full size table

The contribution of each factor to the UHI in the benchmark and proposed models is shown in Figs. 6 and 7. The comparison reveals that altitude, NDVI and population have the greatest influence on the UHI, contributing 35.67%, 34.61% and 21.15% in the benchmark model, respectively. In the proposed model, the contribution of core, branch, islet and edge exceeds that of average building height and NDBI. After normalizing the area of each green space morphological type and overlaying them with LST data, we found that core areas exhibited the highest cooling contribution rate, while perforation areas showed the lowest. This disparity is attributable to differences in landscape morphology and spatial configuration. Core areas, predominantly located at urban fringes with minimal anthropogenic disturbance, feature concentrated forest patches and high landscape connectivity. These conditions facilitate efficient heat conduction and dissipation across the region, thereby accelerating cooling rates and maximizing per-unit-area cooling efficacy. In contrast, perforation areas—characterized by small, fragmented patches within core zones—exhibit limited heat transfer pathways. The resulting thermal isolation impedes effective heat dispersion, leading to localized heat accumulation and diminished cooling performance.

Fig. 6
Fig. 6The alternative text for this image may have been generated using AI.
Full size image

Contribution of influencing factor to UHI intensity (benchmark model).

Fig. 7
Fig. 7The alternative text for this image may have been generated using AI.
Full size image

Contribution of influencing factor to UHI intensity (proposed model).

The above study shows that the UHI effect is influenced by many factors; the most influential one is altitude, followed by NDVI, while the road and NDBI have less influence on UHI. This is related to the geographical location of the study area. Dali City is located in a high-altitude area, and altitude is the most important factor affecting UHI. At the same time, the drop of the mountain makes the vegetation abundant, forming a high NDVI value. Furthermore, the comparison between the benchmark model and the proposed model reveals that the green space morphology has a non-negligible importance on the UHI effect. Therefore, this means that urban planners and managers in Dali City should make full use of the local topography and vegetation growth to rationally plan the urban layout during the urbanization process. Meanwhile, more attention is paid to the urban green landscape layout to alleviate the UHI effect.

Discussion

This study systematically explored the coupling mechanism between green space morphology and urban heat island (UHI) effect in Dali City through the combination of morphological spatial pattern analysis (MSPA) and machine learning. It revealed the regulatory role and nonlinear influence laws of green space morphology on the thermal environment. The results not only provide a new perspective for understanding the urban green space-UHI relationship but also offer empirical support for the optimal planning of urban thermal environments.

Core mechanism explanation and practical value

This study found that green space morphological types (e.g., core, branch, islet, and edge) have significantly higher explanatory power for UHI intensity than average building height and normalized difference building index. This result is consistent with the view proposed by Lin et al that “green space functional morphology can more accurately characterize the cooling effect than traditional landscape indices”22. Among these types, core green space—accounting for the highest proportion—achieves significant cooling by enhancing evapotranspiration and cold air circulation, relying on its concentrated vegetation coverage and high connectivity. This aligns with the conclusion observed by Zhang et al that “large continuous green spaces have better cooling effects than fragmented ones”52. Notably, although the single-factor correlation between green space morphology and UHI intensity is weak, the cumulative effect of these “weak correlations” is crucial in the context of urban land resource constraints. Particularly against the background of Dali’s high urbanization rate of 71.22%, achieving cooling effects by optimizing green space morphology (e.g., reducing perforations in core areas and connecting islet green spaces) is more feasible than unrestrained green space expansion. This is highly consistent with the viewpoint put forward by Chen et al that ‘compact green space morphology is the optimal choice for UHI mitigation under land constraints”48.

Furthermore, the study confirmed that improving green space connectivity can significantly enhance cooling efficiency, a mechanism consistent with the research conclusion emphasized by Gupta and De that “vegetation continuity improves cooling efficiency by promoting cold air diffusion”20. In contrast, the positive impacts of perforations and loops on UHIs may stem from their destruction of the integrity of core green spaces, which hinders the continuous diffusion of cold air. This is in line with the phenomenon discovered by Liu et al that “landscape fragmentation weakens the cooling function of green spaces”14. These findings indicate that the regulation of UHIs by green space morphology is not a simple linear relationship but functions through a chain reaction of “morphology-connectivity-function,” providing a key basis for the precise formulation of UHI mitigation strategies.

Connection and extension to existing research

The random forest (RF) model used in this study significantly outperforms traditional methods such as ordinary least squares (OLS) and spatial error model (SEM) in fitting accuracy and predictive performance. This result verifies the advantage of machine learning in capturing complex nonlinear relationships—it is consistent with the research conclusion of Ma et al. who used the RF model to analyze the nonlinear relationship between landscape patterns and ecological resilience31, and also echoes the discovery of Lin et al. who quantified the nonlinear contribution of green space morphology to UHIs via the RF model21. Meanwhile, the explanatory power of the RF model was enhanced (R²=0.88) after incorporating green space morphological factors, indicating the interpretation bias of traditional models caused by ignoring morphological characteristics. This aligns with the view pointed out by Qi et al. that “the lack of landscape morphological indicators reduces the accuracy of UHI simulation”53.

From the perspective of regional characteristics, the UHI pattern in Dali City is most significantly affected by elevation and NDVI. This is consistent with the conclusion of Shen et al. in their study on tropical cities that “topography and vegetation are the basic regulatory factors of the thermal environment”54, and also confirms the universal law proposed by Yuan and Bauer that “NDVI is significantly negatively correlated with land surface temperature (LST)”6. Additionally, the cooling effect of green space morphology exhibits spatial heterogeneity—for example, islet areas show a negative correlation but limited cooling effect. This is consistent with the cross-regional research result of Masoudi and Tan that “the cooling effect of green spaces depends on morphological integrity and connectivity”18, suggesting that the regulatory role of green space morphology needs to be comprehensively considered in combination with regional topography and vegetation background.

Policy implications and planning recommendations

Based on the research results, the following strategies are proposed for UHI mitigation planning in Dali and similar mountainous cities: Firstly, prioritize the protection of the integrity of core green spaces to avoid perforation and fragmentation caused by urban construction. This is consistent with the suggestion put forward by Wang et al. that “core green space protection is the key to the sustainable regulation of the thermal environment”55, and also conforms to the planning logic that morphological optimization is more efficient than area expansion21. Secondly, enhance the connectivity of fragmented green spaces (e.g., islet areas) by constructing ecological connectors such as corridors and branches. This strategy aligns with the idea proposed by Bai et al. that “ecological network construction strengthens the synergistic cooling effect of green spaces"56. Finally, integrate ground vegetation, rooftop greening, and vertical greening to form a multi-level greening system, combined with Dali’s mountainous topographical characteristics. Meanwhile, blue space planning can be integrated to achieve the synergistic cooling of green-blue infrastructure.

Limitations and future research directions

Despite the above findings, this study still has several limitations. Firstly, the 30 m spatial resolution remote sensing data used may fail to accurately capture the fine morphology of small green spaces and narrow corridors. Secondly, the differences in cooling effects caused by vegetation types and green space age were not distinguished. Thirdly, the model validation did not combine ground meteorological observation data, relying only on remotely sensed LST, which may deviate from near-surface air temperature. “Future research can be conducted from three aspects: first, adopt higher-resolution remote sensing data combined with unmanned aerial vehicle (UAV) observations to capture the morphological details of small green spaces; second, distinguish vegetation types and growth status to quantify the cooling efficiency of different green space morphology-vegetation combinations; third, integrate ground meteorological station data and microclimate observations to improve the practical application value of the model.

Conclusion

Mitigating the urban heat island (UHI) effect is pivotal for enhancing urban livability, yet conventional research has largely neglected the regulatory role of green space morphological patterns in thermal environments. This study integrated morphological spatial pattern analysis (MSPA) with machine learning to investigate the coupling relationship between green space morphology and UHI in Dali City. Our findings demonstrate that green space morphological patterns exert a significant influence on UHI intensity. Among the seven MSPA-derived categories, core green areas exhibited the strongest cooling effect, showing a significant negative correlation with UHI. This is attributed to their concentrated vegetation coverage and high landscape connectivity, which facilitate efficient heat dissipation. Conversely, perforations and loops were positively correlated with UHI, as they disrupt the integrity of core green areas and impede heat transfer. Beyond morphological factors, altitude and the Normalized Difference Vegetation Index (NDVI) emerged as the most impactful non-morphological drivers of UHI—reflecting Dali’s high-altitude topography and abundant vegetation—whereas the Normalized Difference Built-up Index (NDBI), road density, and population density exacerbated UHI by increasing impervious surfaces and anthropogenic heat emissions. Model comparisons confirmed the superiority of the random forest (RF) algorithm in capturing the non-linear relationship between green space morphology and UHI, achieving a coefficient of determination (R²) of 0.88—outperforming ordinary least squares (OLS), spatial error model (SEM), spatial lag model (SLM), and geographically weighted regression (GWR). Incorporating green space morphological factors further improved the RF model’s accuracy relative to a benchmark model excluding these variables, highlighting the necessity of integrating morphological indicators in UHI research. For urban planning in Dali and similar mountainous cities, planners should prioritize the protection of core green spaces to preserve their integrity. Additionally, constructing ecological corridors to connect fragmented islet-like green patches can enhance overall green space connectivity while avoiding excessive subdivision of core areas via perforations. Complementary multi-layered greening strategies—integrating ground-level vegetation, roof greening, and vertical greening—can further optimize the spatial dimension of urban greenery. This study provides a novel framework for UHI mitigation by emphasizing the role of green space morphology, offering actionable insights for sustainable urban development.