The Impact of Urban Forest Landscape on Thermal Environment Based on Deep Learning: A Case of Three Main Cities in Southeastern China

Abstract

Accelerated urbanization has exacerbated the urban heat island phenomenon, and urban forests have been recognized as an effective strategy for modulating thermal environments. Nevertheless, there remains a dearth of systematic investigations into the nonlinear associations between the detailed spatial configurations of urban forests and thermal conditions. We proposed a deep learning-based approach to extract forest data, utilizing multisource high-resolution remote sensing data with relative radiometric correction. Subsequently, we employed deep neural networks (DNNs) to quantify the linkages between urban forest landscape patterns and land surface temperature (LST) in summer and winter across Fuzhou, Xiamen, and Zhangzhou in Fujian Province, China. Our findings indicate the following: (1) Our extraction approach outperforms DeepLabv3+, FCN_8S, and SegNet in terms of extraction precision and adaptability, achieving an overall accuracy (OA) of 87.57%; furthermore, the implementation of relative radiometric correction enhances both the extraction precision and model generalizability, improving OA by 0.05%. (2) Geographic and seasonal differences influence the urban forests’ cooling effects, with more pronounced cooling in summer, particularly in Zhangzhou. (3) The significance of forest landscape composition and configuration in affecting the thermal environment varies seasonally; landscape configuration plays a more pivotal role in modulating surface temperatures across the three cities, with a more critical role in winter than in summer. (4) Seasonal and city-specific variations in forest spatial patterns influence LST. Adopting the appropriate forest structures tailored to specific seasons, cities, and scales can optimize cooling effects. These results offer quantitative insights into urban heat island dynamics and carry significant implications for urban planning strategies.

1. Introduction

As urbanization advances, the urban heat island phenomenon progressively intensifies [1,2]. Extensive research has demonstrated that the expansion of urban forest areas can effectively modulate the regional climate and alleviate elevated urban temperatures [3,4]. Concurrently, the configuration of urban forest landscape patterns, such as patch size and connectivity, exerts a significant influence on the urban thermal environment [5,6]. However, urban land resources in China are constrained, and the financial implications of expanding forest areas are prohibitive. Consequently, optimizing cooling effects within confined spaces has become a critical focus in the study of urban thermal environments [7]. Current research methods on urban green spaces’ thermal environments mainly include field measurements, numerical simulations, and remote sensing inversion. Field measurements utilize meteorological stations or mobile devices, offering the advantages of data authenticity and proximity to the ground; however, they are limited by human and material resources, making their large-scale implementation difficult [8]. Although numerical simulations have been increasingly applied in the study of green spaces’ thermal environments in recent years, the precision and reliability of the models still require further evaluation [9]. In comparison to the previous two methods, remote sensing inversion provides surface temperature data that can reflect the spatial distribution characteristics of the thermal environment in urban green spaces in real time. Additionally, it effectively integrates various landscape pattern analysis methods, making it a crucial approach in current research on urban green spaces’ thermal environments [10].

A comprehensive body of scholarly work has been devoted to examining the quantitative relationship between urban forest spatial patterns and thermal environments from a landscape perspective. This encompasses investigations into landscape composition and configuration, contributing to a substantial corpus of the literature. Nevertheless, certain studies have yielded disparate results [11]. For example, Yao et al. [12] and Zhou et al. [13] argued that more intricately shaped forest patches with enhanced connectivity yielded superior cooling effects, whereas Wang et al. [14] contended that simpler, less fragmented urban forest patches have proven to be more efficacious in cooling, highlighting the need for further inquiry in this domain. Moreover, to more precisely delineate the impact of urban forests on thermal environments, several studies have undertaken fine-grained analyses of temperature fluctuations within urban forests [15,16], although selecting surface data with a suitable resolution presents challenges. Numerous investigations have employed MODIS data; for instance, Liu et al. [17] analyzed the spatiotemporal variations in vegetation and land surface temperature from 2003 to 2021 based on MODIS data, yet their coarse spatial resolution was inadequate for distinguishing heterogeneous land cover or for accurately gauging temperature shifts in specific locales. Thus, high-resolution remote sensing imagery is essential for accurately determining the relationship between urban forest configurations and thermal environments [18], although only a limited number of studies presently incorporate such detailed imagery for analysis. For instance, Liu et al. [19] investigated the competitive effects of urban vegetation and artificial surfaces on land surface temperature (LST) by integrating ZY-3 and Landsat 8 data.

Extracting valuable data from high-resolution imagery has constituted a considerable challenge. Traditional manual interpretation techniques for urban forests are precise yet demand an inordinate amount of labor and time, rendering them inefficient. Zylshal et al. [20] combined support vector machines (SVMs) and object-oriented similarity measurement methods to develop a framework for urban green space extraction, reducing subjectivity and enhancing classification efficiency. Most of these methods are based on the analysis of pixel spectral feature differences, which can lead to confusion of spectrally similar objects during classification. Research indicates that contextual features contain rich spatial information. Xu et al. [21] proposed an automatic layered extraction architecture for farmland using a gray-level co-occurrence matrix, while Zujovic et al. [22] incorporated texture metrics into image classification, eliminating the salt-and-pepper noise in the results. However, these methods all rely on human–computer interaction and have complex parameter settings. Deep learning models, renowned for their robust self-learning and generalization capabilities, circumvent the cumbersome manual feature extraction and preprocessing required by conventional machine learning methods, thereby pioneering new strategies and methodologies for urban forest extraction. Ronneberger et al. [23] introduced an innovative semantic segmentation model called U-Net, which integrates a deconvolution network with a bridging structure to balance classification accuracy with efficiency. This model has garnered significant achievements in domains such as remote sensing image classification [24,25], although U-Net necessitates high scale uniformity, complicating the direct attainment of optimal depth and width. To address these constraints, Zhou et al. [26] developed U-Net++, which links feature maps across various scales to promote cross-layer information flow and integration of high-level semantic insights, mitigating issues associated with information bottlenecks and semantic information loss, albeit at the expense of augmented computational complexity and an increase in parameter volume. Additionally, to enhance the network structure, certain researchers have introduced novel techniques to refine the U-Net model. For instance, Shi et al. [27] incorporated a multi-scale fusion attention module (MSFAM) into U-Net to counteract the insensitivity to local details engendered by global pooling operations in attention modules, proposing a network architecture dubbed A-MSFAM-UNet for high-resolution image water-body segmentation. In this context, for the extraction of urban green spaces, scholars predominantly consider data from a single type of sensor and a specific location for model training, neglecting the integration of high-resolution satellite data from different models and the variations in model transferability across different locations and satellite data. Therefore, proposing a highly applicable and accurate urban forest extraction model is essential.

Here, we address the following two primary concerns: one related to high-precision urban forest extraction, and the other concerning the exploration of the interplay between landscape and thermal environments. For the extraction of large-scale urban forests, models necessitate enhanced data generalizability. Contemporary research often restricts itself to utilizing data from a single sensor type at a specific locale for model training, neglecting the variations in model transferability across diverse locations and types of satellite data. To tackle these challenges, we developed a dataset from multisource high-resolution satellite remote sensing imagery, employing relative radiometric correction for preprocessing to mitigate radiometric discrepancies in the images and bolster the model’s transferability. An advanced U-Net model approach is proposed in this paper for delineating urban forest distribution. Following this, the influence of urban forest landscape patterns on the urban land surface temperature (LST) is examined in Fuzhou, Xiamen, and Zhangzhou, which are situated in Fujian Province, which boasts the highest forest coverage rate nationally. Our research objectives encompass the following: (1) Forecasting an urban forest semantic segmentation model using multisource high-resolution remote sensing imagery, relative radiometric correction, and an advanced U-Net network, while evaluating its classification accuracy and transferability. (2) Utilizing the urban forest semantic segmentation model to generate distribution maps of urban forests in the main urban areas of the three cities and comparing the seasonal cooling variations among these cities. (3) Investigating the relative contributions of urban forest landscape pattern indices to seasonal LSTs in the three cities using the DNN model. (4) Exploring the marginal effects of urban forest landscape pattern indices on seasonal LSTs in the three cities. The goal is to furnish references for the management and design of urban forests, with a focus on modulating urban thermal environments.

2. Methods

2.1. Study Area

2.1.1. Application Area

Fujian Province is situated along the southeastern coast of China, characterized by high terrain in the northwest and low terrain in the southeast, predominantly comprising mountains and hills. The region displays distinct water and thermal conditions and exhibits vertical zonation, with an average annual temperature ranging from 17 to 22 °C, featuring warm winters and hot summers. According to the ninth national forest resource survey, Fujian Province maintained the highest forest coverage rate in China for the 44th consecutive year in 2022. However, over the past two decades, Fujian Province has experienced rapid economic advancement and urban expansion, significantly transforming the land cover and progressively exacerbating the urban heat island effect, thereby necessitating urgent adjustments and optimization. Compared to suburban areas, studying the thermal environment in the principal urban zones carries greater practical relevance [28]. Consequently, we selected the main urban areas of Fuzhou, Xiamen, and Zhangzhou in Fujian Province as the study locales. Fuzhou, the provincial capital, also serves as the political and economic hub of the province, boasting the highest economic growth rate in the first three quarters of 2023. Xiamen’s main urban area is the largest in the province and exhibits a high level of urbanization. Zhangzhou’s main urban area, although smaller compared to Fuzhou and Xiamen, has undergone rapid economic growth in recent years, with an economic growth rate of 4.96% in the first three quarters of 2023, ranking third in the province. The imagery originates from the National Comprehensive Earth Observation Data Sharing Platform (https://www.chinageoss.cn/ (accessed on 1 February 2024)), with images of Fuzhou captured by the Gaofen-6 (GF-6) satellite and those of Xiamen and Zhangzhou by the Jilin-1 (JL-1) satellite. The delimitation of the main urban areas is based on the 2020 built-up area data masks of Chinese cities [29], with the processed images as depicted in Figure 1.

2.1.2. Training Sample Area

Current research on urban forests frequently involves generating training samples based on specific application plots, which exhibit limited transferability. Alternatively, some approaches utilize large-scale data for model initialization, followed by pre-training on plot-specific data, yet this strategy demands high data consistency and incurs significant computational and temporal costs. To develop a resource-efficient, high-performance, and highly transferable urban forest extraction model with national applicability across different sensors, we delineated the training sites from the application sites. The sites were extracted from four provinces in the northern, southern, eastern, and western regions of China, covering a wide range of the country’s geographical conditions, with imaging times encompassing the four seasons of spring, summer, autumn, and winter. The regions, as depicted in Figure 2, were sourced from the China Resources Satellite Application Center (https://www.cresda.com/ (accessed on 23 June 2023)). The provinces sampled sequentially included the following: Dongkeng town, Dongguan city, Guangdong Province; Wenshang County, Tai’an city, Shandong Province; Menyuan Hui Autonomous County, Haibei Tibetan Autonomous Prefecture, Qinghai Province; Haikou town, Yuxi city, Yunnan Province; and Yiliang County, Kunming city, Yunnan Province, subsequently referred to as Zones A to E. Data for Zones A and B were captured by the Gaofen-1 (GF-1) satellite, Zones C and D by the Gaofen-2 (GF-2) satellite, and Zone E by the Gaofen-7 (GF-7) satellite. Data from Zones A to D were employed for model training, The data from Zone E, which were not involved in model training and originated from a different satellite sensor than those in the other four zones, were selected to evaluate the transferability of the proposed model and to serve as a benchmark for relative radiometric correction of the other images.

All data sources used in this study are shown in

Table 1. Data sources.

Data	Spatial Resolution	Date		Usage	Differences
GF-6	2 m	04/13/2020		Fuzhou city main urban area imagery	Multispectral sensor covering blue, green, red, and near-infrared bands, suitable for agricultural and environmental monitoring.
JL-1	5 m	04/08/2022		Xiamen and Zhangzhou city main urban area imagery	Sensor covers visible and near-infrared bands, suitable for urban planning and resource management.
GF-1	2 m	06/04/2021 07/23/2021		Training sample experimental areas: Areas A and B (from Guangdong Province and Shandong Province)	High-resolution panchromatic and multispectral sensors covering blue, green, red, and near-infrared bands.
GF-2	1 m	10/18/2016 12/13/2020		Training sample experimental areas: Areas C and D (from Qinghai Province and Yunnan Province)	Panchromatic resolution better than 1 m, multispectral resolution better than 4 m, covering visible and near-infrared bands.
GF-7	0.65 m	04/14/2021		Training sample experimental area: Area E (from Yunnan Province)	Capable of stereo mapping, panchromatic resolution better than 0.8 m, multispectral resolution of 3.2 m.
Landsat 8	100 m	06/01/2020–09/01/2020	12/01/2020–03/01/2021	Constructing summer and winter LST for Fuzhou, Xiamen, and Zhangzhou	Eleven bands covering visible, near-infrared, shortwave infrared, and thermal infrared bands.
		06/01/2022–09/01/2022	12/01/2022–03/01/2023

.

The overall approach of this research is illustrated in Figure 3.

2.2. Deep Learning Model Framework-Based Urban Forest Extraction

2.2.1. ResSE-UNet Model Architecture

U-Net, a convolutional neural network, is widely employed for image segmentation tasks [23]; however, its limited receptive field presents challenges in processing large-scale objects within remote sensing imagery, and it often lacks sufficient contextual information, resulting in inadequate edge feature representation. To overcome these limitations, residual network modules and attention mechanisms are integrated to enhance U-Net’s efficacy in remote sensing image segmentation tasks.

(1)

Integration of residual networks:

The Deep Residual Network (ResNet), introduced by He et al. [30], adjusts the output of each layer by adding the input to its transformed output. Residual modules utilizing shortcut connections are implemented to combat gradient vanishing and explosion issues.

(2)

Incorporating attention mechanisms:

Although residual networks enhance the capture of urban forest feature information, they can overlook the contextual relationships around target pixels during pixel classification. To emphasize significant features and accurately classify different pixel types, the Squeeze-and-Excitation (SE) [31] module can be integrated into the U-Net framework. This process involves two phases, namely squeeze and excitation. The squeeze phase consolidates spatial information within the feature map using global average pooling, whereas the excitation phase accentuates critical features and diminishes lesser ones by learning inter-channel relationships.

(3)

Overall framework:

In the enhanced U-Net, traditional convolution blocks are replaced with residual modules, each comprising two convolution layers. To maintain network depth without introducing superfluous parameters, the number of convolution layers per level in the base model is reduced from two to one, preserving depth consistency with the original U-Net configuration. The kernel size remains 3 × 3, and the dimensions and channel count of the feature maps align with those of the initial U-Net. The starting number of kernels is 64, with a stride of 1, employing the ReLU activation function for nonlinear transformation of the feature maps. Furthermore, the SE module is embedded within the skip connection modules of each layer in the U-Net network, facilitating the amplification of pertinent features and enhancing the network performance and precision. The architecture that amalgamates residual modules and SE attention modules, termed ResSE-UNet, is illustrated in Figure 4, where rectangles symbolize feature maps, superscripts denote channel counts, and edge numbers represent feature map dimensions.

2.2.2. Relative Radiometric Correction and Other Preprocessing

Currently, many studies have employed the fusion of various satellite data [32,33], yet they often overlook the fact that variations in lighting conditions, sensor types, and spatial resolutions across different satellite remote sensing images can cause fluctuations in the dynamic range, thereby undermining the generalizability of the trained models. Therefore, we applied relative radiometric correction to standardize the radiance values of disparate images to a uniform range. The corrected outcomes closely resemble reference images captured with the same sensor under similar atmospheric and lighting conditions. Radiometric correction mitigates variations in radiance due to inconsistent acquisition conditions [34,35].

Panchromatic multispectral sensors (PMSs) mounted on various satellites utilize bands B4 (NIR), B3 (R), and B2 (G) to generate standard false-color images. Following the acquisition of the original ground images through orthorectification and image cropping, relative radiometric correction was executed on these images using ENVI 5.6. Data from Zone E, which were excluded from the training phase, served as the reference for calibrating the gain and bias necessary for the other images through linear regression methods, culminating in 8-bit true-value ground images, as illustrated in Figure 5. It can be observed that the corrected image exhibits noticeable differences in color brightness compared to the reference image. However, after applying relative radiometric correction, the image after correction displays similar color brightness to the reference image. This similarity is also evident in their corresponding histograms, where the gray-level range of the corrected image is normalized after relative radiometric correction to match the reference image’s range.

The process of urbanization induces interactions between urban forests and cultivated land, manifesting significant differences in their rates of change and spatiotemporal variations, which complicates the classification task. Consequently, the selected classification categories were urban forests, cultivated land, and other land features, organized into three distinct groups. Deep learning necessitates substantial data volumes for training models, and the available open-source data are insufficient for the classification demands of this study. Visual interpretation was employed to augment the dataset. Utilizing Google Earth (https://earth.google.com/ (accessed on 15 July 2023)) and high-resolution satellite remote sensing imagery, all land features within the designated area were manually annotated, and the outcomes were saved as 8-bit raster images, designated as real ground labels.

The processed images and corresponding label data from Zones A to D were partitioned into training and testing sets at a 3:1 area ratio, ensuring that there was no overlap between the training and testing datasets. Techniques such as random cropping and random data augmentation were implemented to create additional data samples [36], thereby expanding the size of the training set. This approach resulted in the generation of 2400 training samples and 800 testing samples (Figure 6).

2.2.3. Experimental Setup

All training samples consisted of 256 × 256 pixel images alongside their corresponding labels, comprising 2400 training samples and 800 testing samples. “BATCH_SIZE” refers to the batch size, determining the number of samples used in each training iteration. In this study, BATCH_SIZE was set to 4 to enhance the training efficiency while ensuring sufficient data diversity to maintain a reasonable allocation of memory resources. “OPTIMIZER” denotes the algorithm used to update network parameters. We selected the Adam optimizer due to its excellent performance in the field of deep learning. “LEARNING_RATE” indicates the initial learning rate, a critical hyperparameter in the optimization algorithm, which was set to 0.00001 in this study; this determines the magnitude of parameter updates in each iteration. “MAX_EPOCH” represents the maximum number of iterations, indicating the number of times the dataset is traversed during the entire training process. Based on experimental observations, we chose 100 as the maximum number of iterations. For the loss function, we selected the “Cross Entropy Function” because it effectively handles probability distributions, ensuring that the model’s output probability distribution closely approximates the true distribution.

The system configuration details are presented in

Table 2. Configuration of operating environment.

Experimental Environment	Detailed Information
Software environment	Programming language: Python 3.8
	Deep learning framework: Keras 2.10.0, TensorFlow 2.10.0
	Development environment: Anaconda, PyCharm 2021
	Results visualization: ArcGIS 10.7
Hardware environment	CPU: AMD Ryzen 7 5800H with Radeon Graphics
	GPU: NVIDIA GeForce RTX 3080 Laptop GPU

.

The accuracy evaluation metrics were based on the confusion matrix [37], as follows:

$$OA={\displaystyle \sum _i^n\frac{x_{ii}}{N}}$$

(1)

$$MIoU={\displaystyle \frac{{\displaystyle \sum _i^n\frac{x_{ii}}{x_{i+}+x_{+i}+x_{ii}}}}{N}}$$

(2)

$$kappa={\displaystyle \frac{N{\displaystyle \sum _i^n{x}_{ii}}-{\displaystyle \sum _i^n\left(x_{i+}{x}_{+i}\right)}}{N^2-{\displaystyle \sum _i^n\left(x_{i+}{x}_{+i}\right)}}}$$

(3)

where OA denotes the overall accuracy, signifying the ratio of correctly classified pixels to the total pixel count; this metric reflects the precision of the entire classification output. MIoU, or mean intersection over union, is employed to assess the accuracy of the image segmentation model, indicating the extent of overlap between the forecasted segmentation outcomes and the actual segmentation outcomes. The kappa coefficient is utilized to evaluate the predictive accuracy of the classification model, illustrating the discrepancy between the actual accuracy of the classification model and chance-based predictions, where $x_{ii}$ represents the count of correctly predicted pixels within the image, $x_{i+}$ and $x_{+i}$ correspond to the aggregate sums of the rows and columns in the confusion matrix, respectively, and $N$ denotes the total pixel count.

2.3. Land Surface Temperature Retrieval

We selected data from Landsat 8 for our calculations. Although using high-resolution temperature data or field measurement methods can provide more precise results, the cost of such methods is prohibitive for large-scale urban studies. Specifically, obtaining and processing high-resolution temperature data requires significant time and computational resources, making it impractical for extensive and long-term urban applications. Additionally, many studies have combined high-resolution imagery with Landsat 8 TIRS data to investigate the impact of urban forests on LST [38,39,40], demonstrating the feasibility of this approach. Therefore, we chose to use Landsat data to reduce costs, enabling the acquisition of broader spatial coverage and extended temporal data. The formula employed to ascertain the LST of the three main urban zones involved calculations using the T1-level surface reflectance data product (Thermal Infrared Sensor, TIRS) from Landsat remote sensing images on the GEE platform (https://earthengine.google.com/ (accessed on 17 February 2024)), featuring a spatial resolution of 100 m, and was already adjusted for radiometric, atmospheric, and geometric corrections [41]. The calculation formula is as follows:

$$LST=0.00341802\times DN+149-273.15$$

(4)

where $DN$ is the grayscale value of the image pixel in the LC2L2ST_B10 band corresponding to the TIRS sensor.

To calculate the average LST for each season, we selected the summer and winter months of each year according to the months and the high-resolution imaging schedules of the three cities. The average LST for all images within the defined area was computed for each season, and the spatial resolution was subsequently resampled to 30 m.

2.4. Landscape Pattern Analysis

Landscape pattern indices, given their extensive encapsulation of landscape pattern data, are extensively applied in quantitative assessments of urban landscape patterns and thermal environments. Informed by prior studies [42,43], we categorized two principal types of landscape indices—landscape composition and configuration—utilizing the moving-window technique provided by Fragstats 4.2 to calculate the landscape pattern indices, as outlined in

Table 3. Landscape pattern indices.

Landscape Pattern Index	Description	Unit
Aggregation index (AI)	The connectivity of a given patch type within the landscape, reflecting the degree of patch aggregation	%
Edge density (ED)	The ratio of patch boundary length to area within the landscape, reflecting the edge effect of landscape patches	m/ha
Mean Euclidean nearest-neighbor distance (ENN_MN)	The shortest straight-line distance between a focal patch and its nearest neighbor, reflecting the connectivity of landscape patches	m
Landscape shape index (LSI)	The complexity of patch shapes compared to a simple geometric shape, reflecting landscape complexity	m
Percentage of landscape (PLAND)	The percentage of total landscape area covered by a specific patch type, reflecting the scale of landscape patches	%

. Landscape composition pertains to the size and diversity of patch types, with the prevalent landscape metric being the percentage of landscape (PLAND), whereas landscape configuration concerns the spatial structure and proximity relationships among patches, with typical metrics including the Aggregation Index (AI), edge density (ED), mean Euclidean nearest-neighbor distance (ENN_MN), and Landscape Shape Index (LSI). It is well known that the resolution of remote sensing data in landscape pattern analysis is scale-dependent, necessitating precise determination of the relationship between urban forest spatial patterns and thermal environments. We selected four moving-window sizes when constructing indices in Fragstats 4.2: 0.3 × 0.3 km, 0.6 × 0.6 km, 0.9 × 0.9 km, and 1.2 × 1.2 km.

2.5. Deep Neural Network Regression Model

We employed deep neural networks (DNNs) to analyze the relationship between urban forest spatial patterns and the thermal environment. We used the fishnet tool in ArcGIS 10.7 to extract the spatial patterns of urban forests and surface temperature. To correspond with the moving-window sizes set for constructing landscape pattern data, we similarly set the fishnet grid sizes from 0.3 to 1.2 km to determine the optimal grid size. We adopted the root mean square error (RMSE) as the evaluation metric to gauge the accuracy of the DNN regression model. RMSE calculates the square root of the average of squared differences between predicted values and actual values, effectively quantifying the deviation between these values and exhibiting sensitivity to outliers in the dataset. The RMSE formula is as follows:

$$RMSE=\sqrt{{\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^n{\left(Y_i-f\left(x_i\right)\right)}^2}}$$

(5)

where Yi represents the actual regression values and f(xi) represents the model-predicted values.

Subsequently, utilizing the DNN regression model, we explored the specific relationships between urban forest spatial patterns and the thermal environment under various grid scales to identify the optimal grid scale. Following this, based on the determined optimal grid scale, the relative contributions and marginal effects of each landscape index were examined. One of the advantages of DNNs, in contrast to traditional regression models, is their ability to autonomously learn high-level features from data without the necessity for manual design or selection, and they impose fewer assumptions and constraints when modeling nonlinear, high-dimensional, complex functional relationships. The term “relative contribution” refers to the extent to which each variable within the DNN regression model affects the dependent variable, reflecting the explanatory power or significance of different independent variables. However, prevailing methods for calculating relative contributions, such as permutation importance, face challenges in capturing the nonlinear interactions and relationships between features. Consequently, this study adopted the SHAP (SHapley Additive exPlanations) [44] method based on cooperative game theory for importance interpretation. SHAP is a game-theoretic approach to explaining the output of any machine learning model; utilizing classic game theory, Shapley values, and their related extensions, it combines optimal credit allocation with local explanations, capable of identifying and quantifying interactions between features and providing explanations at the individual sample level.

The marginal effect refers to the variation in the prediction of the target variable induced by a feature variable when other feature variables are held constant. This effect is calculated by observing the changes in the predicted outcomes of the feature variable under varying values; its changing curve can intuitively illustrate the relationship and threshold characteristics of each independent variable with respect to the dependent variable. The DNN model for this study was developed in a Python environment, with software and hardware settings consistent with those listed in Table 1. After multiple experiments, the finalized settings included a learning rate of 0.01, a learning decay rate of 0.005, a batch size of 512, a training-to-validation set ratio of 8:2, hidden-layer neuron counts of 64, 128, 256, 512, 512, 1024, 1024, with the optimizer set to Adam, and 200 training epochs.

3. Results

3.1. Accuracy Assessments

3.1.1. Ablation Experiments

To evaluate the efficacy of the enhanced modules in the ResSE-UNet model, ablation studies were performed on the residual module, SE attention module, and relative radiometric correction. These studies assessed their transfer capabilities and classification accuracy specifically in the untrained Zone E. The results of the ablation experiments with different configurations are presented in Figure 7 and

Table 4. Quantitative comparison of different configurations of the bottleneck.

Feature Separation			OA/%	MIoU/%	Kappa
Res Model	SE Model	Relative Radiometric Correction
×	×	√	86.51	67.49	0.77
√	×	√	87.00	68.27	0.78
×	√	√	87.20	68.97	0.78
√	√	×	87.52	69.00	0.79
√	√	√	87.57	69.47	0.79

Note: × √ indicates whether the feature was included in the model, with bold font indicating the best results in each column.

. Here, “Unet_Res” signifies the sole incorporation of the residual module, “Unet_SE” represents only the addition of the SE attention module.

As illustrated in Figure 7, the base U-Net model captured detailed edge information but failed to identify many dispersed forest areas. The incorporation of the residual module mitigated the issue of missed classifications, although it led to some misclassifications, such as urban areas in the upper left corner of the image being erroneously classified as farmland. Introducing only the SE attention mechanism enhanced the detection of scattered forests, yet it also resulted in an increase in misclassifications, likely because the attention mechanism, while emphasizing contextual details, inadvertently incorporated many irrelevant features. The ResSE-UNet model, which amalgamates both modules, delivered the most accurate classification outcomes, effectively discerning subtle urban forests and minimizing misclassifications. Both Unet_Res and Unet_SE, which individually integrated the residual module and SE attention mechanism, respectively, demonstrated enhanced transfer accuracy compared to the base U-Net model. Integrating both modules into U-Net achieved a higher average precision, with an OA increase of 1.06 percentage points relative to the original U-Net. When comparing the impact of relative radiometric correction, the ResSE-UNet model without relative radiometric correction exhibited inferior performance across all accuracy metrics relative to the ResSE-UNet with correction, underscoring that mitigating radiometric discrepancies between sensors can significantly bolster the network’s transfer capabilities in multisource remote sensing images.

3.1.2. Comparative Experiments

To illustrate the transfer advantages of the model proposed in this study, a comparative accuracy analysis was performed with three well-known semantic segmentation networks (SegNet, FCN_8S, and DeepLabv3+) in Zone E. The extraction outcomes and accuracy metrics are detailed in Figure 8 and

Table 5. Quantitative comparison of ResSE-UNet and other deep learning models.

Method	OA/%	MIoU/%	Kappa
SegNet	84.21	62.50	0.73
FCN_8S	86.65	66.07	0.77
DeepLabv3+	80.97	58.87	0.68
ResSE-UNet	87.57	69.47	0.79

Note: The bold font represents the optimal result for each column.

.

As depicted in Figure 8, on an entirely different high-resolution remote sensing image, both SegNet and FCN_8S missed significant details along the edges of objects, being unable to clearly delineate the diamond shapes of the roads and farmlands intermixed with the urban forests. DeepLabv3+ displayed considerable salt-and-pepper noise across extensive farmland areas, proving ineffective at accurately extracting the urban forests. In contrast, the visual classification performance of ResSE-UNet was notably the most satisfactory among these models. According to Table 5, the ResSE-UNet model achieved the highest accuracy compared to SegNet, FCN_8S, and DeepLabv3+. The OA, ranked from highest to lowest, was as follows: ResSE-UNet (87.57%), FCN_8S (86.65%), SegNet (84.21%), and DeepLabv3+ (80.97%), demonstrating that the ResSE-UNet model holds a substantial advantage in classification accuracy and transfer capability over existing popular semantic segmentation models. The OA, MIoU, and kappa coefficient for ResSE-UNet were recorded at 87.57%, 69.47%, and 0.79, respectively, fulfilling the accuracy requirements of this study. We also conducted accuracy comparisons in the training Zones A–D. For the comparison results, please refer to Figure S1 in the Supplementary Materials, and for the accuracy results, see Table S1.

3.2. Distribution Characteristics of Urban Forests and Thermal Environment in Three Cities

The spatial distribution of urban forests within the main urban area of Fuzhou in 2020, as delineated by the ResSE-UNet model, is depicted in Figure 9. The urban forests cover an area of 27.54 km², constituting approximately 26.73% of the main urban zone. The LST during summer varies from 33.33 to 62.25 °C, while in winter, it ranges from 11.50 to 30.05 °C. The spatial distribution patterns of LST in both seasons are relatively similar, with higher temperatures predominantly located in the northern sector of the city. The urban forest coverage in the main urban area of Xiamen in 2022 is illustrated in Figure 10. The total area of these forests amounts to 31.39 km², representing ~27.53% of the main urban zone. The summer LST fluctuates between 27.93 and 67.11 °C, and the winter LST between 9.87 and 35.23 °C. The highest and lowest LSTs in both seasons surpass those in Fuzhou, likely due to Xiamen’s proximity to the sea, which influences its climate, causing high LSTs mainly on the western side of the city, opposite the sea. The layout of urban forests in the main urban area of Zhangzhou is displayed in Figure 11. These forests occupy a total area of 5.39 km², roughly 29.41% of the main urban zone. The summer LST spans from 25.29 to 57.66 °C, and the winter LST from 16.29 to 32.75 °C. The highest summer LST is observed on the western side of the city, while in winter, it shifts to the eastern side.

Table 6. Cooling intensity of urban forests in different seasons.

Temperature Regulation (°C)	Summer	Winter
The average surface temperature of the Fuzhou city area	47.59	19.70
The average surface temperature of urban forests in Fuzhou city	46.31	19.28
The cooling intensity of Fuzhou city	1.29	0.41
The average surface temperature of the Xiamen city area	44.34	21.69
The average surface temperature of urban forests in Xiamen city	43.21	21.34
The cooling intensity of Xiamen city	1.13	0.35
The average surface temperature of the Zhangzhou city area	44.00	22.11
The average surface temperature of urban forests in Zhangzhou city	42.66	21.50
The cooling intensity of Zhangzhou city	1.34	0.61

outlines the regulatory impacts of urban forests on LST across different seasons, indicated by the disparity between the average surface temperature of the region and that of the urban forests. In both summer and winter, the average LSTs in the three cities demonstrate that impervious surfaces exhibit higher temperatures than urban forests, with the cooling effect in summer (1.29 °C, 1.13 °C, 1.34 °C) being more pronounced than in winter (0.41 °C, 0.35 °C, 0.61 °C), highlighting that the spatial variability in LST in summer is significantly greater than in winter. From a city comparison perspective, the cooling effects in summer and winter were ranked as follows: Zhangzhou > Fuzhou > Xiamen.

3.3. The Relative Contributions of Urban Forest Landscape Pattern Indices to Seasonal LSTs in the Three Cities

Existing research demonstrates a significant dependence on scale in the analysis of urban forest thermal environments. Figure 12 illustrates the fitting precision of landscape pattern indices and LST in Fuzhou across various scale units within the training and testing datasets of the DNN regression model. Among the four grid dimensions analyzed, the 300 m grid size displayed a notably lower RMSE relative to other grid dimensions, indicating that larger grid dimensions may impede the model’s fitting efficacy. Consequently, a 300 × 300 m grid size was determined to be the optimal scale, and subsequent statistical analyses were carried out based on this optimal scale.

Utilizing the optimal grid dimensions of 300×300 m and the DNN regression algorithm, the relative contributions of various landscape pattern indices to LST were determined (Figure 13), revealing distinct seasonal and geographical variations between forest spatial configurations and the urban thermal environment. Specifically, the five landscape indices were ranked from the highest to the lowest average contributions as follows: ENN_MN (29.54%), AI (22.00%), PLAND (18.23%), LSI (17.78%), and ED (12.45%), with the ENN_MN index demonstrating the most substantial contribution across all three cities. Across different periods, the relative contributions of landscape configuration to the urban thermal environment exceeded those of landscape composition in all three cities; in winter, the impact of landscape configuration surpassed that in summer, while the reverse was observed for landscape composition. Overall, landscape configuration emerged as a critical factor influencing urban thermal environments in the three cities, followed by landscape composition.

3.4. The Marginal Effects of Urban Forest Landscape Pattern Indices on Seasonal LSTs in the Three Cities

Figure 14 further elucidates the marginal effects of urban forest landscape pattern indices on seasonal LST in the three cities. For the results of the DNN-fitted function, the marginal effect of each predictor on LST was ascertained by maintaining all other variables at their average values. A positive slope of the fitted function’s curve suggests a positive correlation between the landscape index and LST; conversely, a negative slope indicates a negative correlation. When the slope approaches zero, it signifies that the correlation between the landscape index and LST is negligible. Data within parentheses reflect the magnitude of influence exerted by each factor on LST.

Our findings indicate that the marginal effects of urban forest landscape patterns on LST also exhibit geographical and seasonal variability. In summer, the primary factors in Fuzhou are ENN_MN (1.47 °C), AI (0.79 °C), and PLAND (0.42 °C), with ENN_MN displaying a nonlinear positive correlation with LST, PLAND demonstrating a negative correlation, and AI exhibiting a transition from negative to positive correlation at a threshold of 62. In winter, the principal factors are ENN_MN (0.99 °C), AI (0.82 °C), and LSI (0.86 °C), with ENN_MN showing a positive correlation with LST, while AI and LSI are negatively correlated. In Xiamen, during summer, the dominant factors are ENN_MN (1.55 °C), PLAND (0.64 °C), and AI (0.53 °C), with both ENN_MN and AI positively correlated with LST, while PLAND is negatively correlated; in winter, the significant factors are ENN_MN (0.89 °C), AI (0.60 °C), and PLAND (0.43 °C), with AI positively correlated, PLAND generally negatively correlated, and ENN_MN exhibiting a shift from positive to negative correlation at a threshold of 79. The marginal effects in Zhangzhou mirror these findings, with summer’s dominant factors being ENN_MN (1.65 °C), PLAND (1.43 °C), and AI (0.55 °C), and with ENN_MN positively correlated with LST, PLAND negatively correlated, and AI transitioning from negative to positive correlation at 57; in winter, the dominant factors are ENN_MN (1.93 °C), AI (0.63 °C), and PLAND (0.38 °C).

From the perspective of landscape pattern indices, in terms of landscape configuration, except for winter in Xiamen, ENN_MN is consistently positively correlated with LST, showing the broadest adjustment range (0.89 °C to 1.93 °C); AI and LSI exhibit positive and negative correlations with LST in winter, respectively, and alternate between positive and negative correlations in summer upon reaching specific thresholds. It was observed that the initial positive or negative correlation is dependent on a lower PLAND index, while the reverse correlation manifests under a higher PLAND index. In simpler terms, during summer, only when urban forest areas exceed a certain size do AI and LSI display consistent warming and cooling effects as observed in winter. Increasing these indices under an inadequately scaled forest leads to the opposite effect on LST; ED generally shows a negative correlation with LST across most conditions, except in winter in Zhangzhou. In terms of landscape composition, PLAND is negatively correlated with LST in summer. In winter, below a certain threshold (16.8 in Fuzhou, 22.1 in Xiamen, and 21.8 in Zhangzhou), the cooling effect of PLAND is reduced and may even lead to a warming effect (as observed in Fuzhou and Xiamen), but exceeding this threshold reinstates the same negative correlation as in summer.

4. Discussion

4.1. The ResSE-UNet Model

Currently, the fine classification of urban forests globally is often restricted to data obtained from identical sensors, neglecting the model’s capacity for transfer and generalization across regions captured by distinct sensors [45,46]. Furthermore, most urban forest distribution mapping relies solely on imagery data from a specific study area, curtailing its broader applicability. To address this limitation, this paper introduces a resource-efficient, effective, and highly transferable urban forest extraction model that is applicable nationwide across different sensors—ResSE-UNet. A dataset was compiled using images from four cities across China, captured by the GF-1/2 satellites, and the model’s transferability was confirmed using images from the GF-7 satellite. To ensure uniformity across multisource data, relative radiometric correction was implemented to mitigate spectral discrepancies among different remote sensing images, thereby minimizing pixel confusion and enhancing the classification accuracy and generalizability. Based on the U-Net network, residual modules were integrated, and SE attention modules were added at cross-layer connections. An end-to-end approach was adopted, employing feature map inputs and outputs to identify and extract urban forests and farmland from multisource high-resolution remote sensing images.

The results demonstrate that, compared to current prevalent semantic segmentation models and U-Net networks with a single module added, ResSE-UNet achieves superior forest extraction precision and generalizability, exhibiting significant improvements across all metrics, with its extraction accuracy, MIoU, and kappa coefficient reaching 87.57%, 69.47%, and 0.79, respectively. Relative to the original U-Net, the overall classification accuracy increased by 1.06 percentage points, and relative to U-Net networks incorporating either the residual module or the SE attention module alone, the accuracy improved by 0.57 and 0.37 percentage points, respectively, indicating that the combination of residual modules and SE attention mechanisms can synergistically enhance the network’s performance. Compared to the three prevailing classification methods—SegNet, FCN_8S, and DeepLabv3+—the overall accuracy improved by 3.36, 0.92, and 6.60 percentage points, respectively. Compared to models trained without relative radiometric correction, the post-correction transfer accuracy increased by 0.05 percentage points, confirming that eliminating radiometric discrepancies between samples can bolster model performance and robustness, thereby augmenting generalizability under diverse conditions. Thus, relative radiometric correction proves to be an efficacious preprocessing method for constructing multisource satellite datasets.

4.2. Effect of Urban Forest Landscape Pattern Index on LST

Research conducted in the main urban areas of Fuzhou, Xiamen, and Zhangzhou within Fujian Province revealed that the influence of urban forests on the thermal environment is subject to seasonal and geographical variations, with Zhangzhou demonstrating the most pronounced cooling effect among the three cities. In both summer and winter, urban forests exert a cooling influence on the thermal environment, achieving the greatest temperature reduction during the summer and the least in winter, corroborating findings reported by Liu et al. [47]. By integrating DNN regression models with Pearson’s correlation analyses, it was established that the optimal research scale for landscape indices derived from urban forest distribution maps using the ResSE-UNet model, and their correlation with the thermal environment, was 300 m. This contrasts with the conclusions of some scholars [28,48]. Multiple factors may account for these results, including the following: (1) The landscape indices in this study were generated from high-resolution imagery of urban forest distribution, providing greater spatial resolution than the predominantly used 30 m Landsat images, which are susceptible to pixel confusion and obscured forest details. (2) The main urban area selected in this study excludes most wastelands and suburbs, whereas other studies define their urban extents based on urban distribution maps from planning agencies, encompassing a significant amount of non-urban pixel data.

In both summer and winter, landscape configuration exerts the most substantial influence on LST, followed by landscape composition. The relative contribution of landscape configuration to the urban thermal environment is more pronounced in winter than in summer, with the reverse being true for landscape composition. Specifically, considering different indices representing forest spatial patterns, we elucidated the nonlinear relationships between various influencing factors and LST. Regarding landscape configuration, some indices display consistent threshold characteristics across various seasons. ENN_MN emerges as the pivotal factor affecting LST among all indices, predominantly exhibiting a positive correlation with LST in most experimental settings. AI is positively correlated with LST in winter and shifts to a positive correlation upon surpassing a threshold in summer, reflecting the connectivity and aggregation of urban forests. ED generally shows a negative correlation with LST across most studies, while LSI is positively correlated with LST in winter but transitions to a negative correlation after a certain threshold in summer, reflecting the complexity and fragmentation of forest shapes. Pertaining to landscape composition indices, PLAND is negatively correlated with LST in summer. However, in winter, if the threshold is not surpassed, PLAND’s cooling effect may decrease or even induce a warming effect (in Fuzhou and Xiamen), and exceeding this threshold reinstates the same negative correlation as in summer. This suggests that larger-scale urban forests effectively reduce LST in summer, but smaller-scale forests are less effective in cooling during winter.

In summary, during the summer, when urban forests are smaller in scale, the strategic clustering of patches promotes urban cooling. However, once the forest reaches a certain size, modifying its shape and fragmentation can significantly enhance the cooling effect. This is consistent with the findings of Zhou et al. [49], who posited that the morphological structure of forests mirrors their opportunities for energy exchange with the external environment. A more complex structure augments the forest’s interaction with the external environment and increases its shading area, thereby reducing the LST. When forests are smaller, overly intricate and fragmented structures lead to increased internal disturbances from the surrounding thermal environment, diminishing their cooling effect. Conversely, when forests are larger, the internal environment becomes relatively stable, and enhancing the complexity and fragmentation of the forest shape can indeed aid in reducing LST [50]. In winter, the influence of landscape composition on LST becomes less pronounced, and smaller forests are less effective in reducing LST. Once the forest scale surpasses a certain threshold, the pattern aligns with that observed in summer. During winter, many plants enter a dormant state and shed leaves, diminishing the vegetation’s capacity to absorb solar radiation and provide shade, which can result in a reduced ability of vegetation to modulate LST, coupled with lower winter temperatures and diminished solar radiation intensity, thereby weakening the urban heat island effect and lessening the impact of landscape composition on LST. An increase in PLAND can effectively reduce LST, indicating that cultivating urban green spaces can consistently and effectively control the urban thermal environment, consistent with the findings of previous studies [42,51,52]. The optimal grid scale identified in this study differs from other research, possibly due to our use of high-resolution imagery for constructing landscape indices, whereas other studies often use imagery with 30 m or 10 m spatial resolution [53,54].

It should be highlighted that although ENN_MN and ED generally exhibit positive and negative correlations with LST, respectively, ENN_MN in Xiamen and ED in Zhangzhou both display positive and negative correlations around the threshold during winter. This indicates that, in winter, when the contribution of landscape composition is diminished and the distribution of urban forests is more extensive, an overly dense clustering of patches can also reduce LST. In Zhangzhou, where the urban area is more constrained, increasing the complexity of patches when forest size is minimal can impede the regulation of LST.

4.3. Implications and Limitations

Urbanization is characterized by the development of the urban heat island effect. This suggests that climate-conscious urban planning can contribute to balancing the urban energy budget and mitigating thermal environments. Therefore, we propose addressing the urban heat island effect from two perspectives: Firstly, seasonally, given that the need to optimize the urban thermal environment primarily occurs in the summer or transitional seasons, enhancing the urban thermal environment through optimizing the landscape composition proves more effective than modifying the landscape configuration. Nevertheless, it must be acknowledged that, in highly urbanized areas with limited forest resources, it is challenging to significantly increase the forest size. Thus, it is crucial to judiciously adjust the forest structure to optimize the urban thermal environment. Assuming a feasible increase in the size of urban forests, modifying the structural morphology of forests based on the threshold characteristics of different indices to maximize the reduction in regional temperatures represents a significant direction for future urban planning and forest construction. For winter, the focus should shift to cultivating small forest patches that have not reached the threshold size and maintaining forest patches at this threshold.

Secondly, from the perspective of the distribution area of principal urban zones, taking Zhangzhou as an example, for cities with smaller main urban areas, in addition to the suggestions above, it becomes more crucial in winter to concentrate on cultivating and maintaining smaller forest patches. Conversely, using Xiamen as an example, for highly urbanized larger main urban areas, it is vital in winter to focus on the distribution and structure of large forest patches.

The following limitations of this study must be acknowledged:

(1)

In this study, we focused solely on urban forests, but urban forest composition—e.g., tree species diversity, age structure, canopy coverage, and the mix of trees and shrubs—also significantly impacts the urban thermal environment. Future research will need to incorporate urban forest composition into the analysis.

(2)

The conclusions drawn from this study, which are based on remote sensing inversion of the urban green space thermal environment, have certain limitations. These limitations are not due to the accuracy of remote sensing inversion but stem from the fundamental differences between surface temperature and air temperature [55]. On the one hand, the mechanisms of surface temperature and air temperature differ; air temperature changes more rapidly than surface temperature and is more influenced by atmospheric movements, weather, and radiative convection. Consequently, it is necessary to investigate whether studies based on air temperature will yield similar or opposite conclusions to those based on surface temperature, such as optimal research scales, relative importance, and threshold characteristics. The conclusions of these two types of studies should not be conflated and require further exploration. On the other hand, as thermal comfort research advances, human comfort has become an important metric for assessing the cooling effects of green spaces. Air temperature serves as a direct indicator in this context and shows greater research value compared to surface temperature. For instance, Schatz et al. [56] used continuous temperature measurements from 151 fixed sensors to characterize the thermal environment of different land cover types throughout the year. However, conducting large-scale thermal environment studies based on air temperature is challenging due to the time, manpower, and financial costs involved. Notably, some researchers have developed models to estimate near-surface air temperature from surface temperature, combining remote sensing with field measurements [57]. This method could facilitate a more comprehensive understanding of the interactions and mechanisms between air temperature and the thermal environment of green spaces.

(3)

This study selected only one region with data collected from different sensors for validation, resulting in a limited number of validation areas. Future research should focus on examining the model’s transferability across more validation areas with different sensor types and geographical factors. Additionally, this study utilized visual interpretation instead of field surveys. Although field surveys for recording urban forest labels require a significant amount of labor, they offer greater accuracy and rigor compared to our visual interpretation. In the future, we will consider incorporating this method for label construction.

(4)

In regions of low spectral value and similar spectral zones, the ResSE-UNet architecture proposed in this paper, like other popular semantic segmentation models, exhibits some misclassifications and omissions. The performance of deep learning models is intimately connected to the quality of sample data. Considering the effective number of bands in high-resolution satellites, future efforts might involve augmenting the number of sample data channels, such as incorporating vegetation features and texture characteristics to enhance models’ extraction accuracy.

(5)

We focused solely on the summer (June to September) and winter (December to March) seasons as the periods of interest. In the future, we plan to include spring and autumn, representing the transitional seasons from cold to warm and vice versa, respectively.

(6)

We did not consider factors such as vertical greening and rooftop gardens—i.e., small green spaces—within the context of limited urban land resources. Properly establishing these small green spaces could increase the total urban forest cover and, combined with ventilation corridors and hydrological conditions, amplify the cooling effects of forests, thereby benefiting the optimization of urban heat environments. Finally, we only considered the impact of urban forests on the thermal environment. Moving forward, we plan to extend our research to encompass other elements, such as bodies of water.

5. Conclusions

Utilizing the cities of Fuzhou, Xiamen, and Zhangzhou as case studies, we investigated the application of deep learning to the spatial patterns of urban forests and their impacts on the urban thermal environment. Our principal conclusions are as follows: (1) This study developed a deep learning-based urban forest semantic segmentation model—ResSE-UNet—that is adaptable across different cities, sensors, and spatial resolutions. The verification results confirm that this model surpasses existing popular semantic segmentation methods in terms of classification accuracy and transferability. (2) Employing ResSE-UNet, we successfully obtained precise urban forest coverage data for Fuzhou, Xiamen, and Zhangzhou. An analysis of the urban forest landscape patterns’ impact on the urban thermal environment indicated that urban forests significantly moderate the thermal environment, with Zhangzhou demonstrating the most substantial cooling effect among the cities, and the cooling effect being more pronounced in summer than in winter. (3) The relative significance of forest landscape composition and configuration varies seasonally. Overall, landscape configuration markedly influences the surface temperatures of the three cities, more so than landscape composition. The relative contribution of landscape configuration to the urban thermal environment is more substantial in winter than in summer, whereas the reverse was observed for landscape composition. (4) Different forest landscape configurations and composition indices exhibit diverse marginal effects on the thermal environment. In terms of landscape configuration, ENN_MN and winter AI are predominantly positively correlated with LST, while ED and winter LSI are predominantly negatively correlated with LST. In summer, AI and LSI demonstrate positive–negative and negative–positive correlations around a threshold, respectively. Regarding landscape composition, PLAND is negatively correlated with LST in summer; however, its cooling effect diminishes or even reverses in winter if the threshold is not achieved, and surpassing the threshold reinstates the same negative correlation. Consequently, in summer, employing structures with low fragmentation and high aggregation when the forest size is small, and using complex shapes with high fragmentation and aggregation when the forest size is large, proves more beneficial for reducing local LST. In winter, small forest sizes are ineffective at reducing LST, necessitating enhanced monitoring and cultivation of small-scale forest patches. In conclusion, the deep learning model introduced in this paper provides a novel tool for managing and monitoring urban forests in other cities. The insights and methodologies derived from analyzing the cases of Fuzhou, Xiamen, and Zhangzhou offer innovative approaches for urban planning and design aimed at optimizing the urban thermal environment.