Machine Learning Application in Investigating Cooling Effect of Urban Blue–Green Infrastructure: A Systematic Review

Abstract

Blue–green infrastructure (BGI) is an important nature-based solution for mitigating urban overheating, and machine learning (ML) techniques offer strong potential in capturing the complex, nonlinear relationships between influencing factors and BGI cooling effects. This study presents a systematic review of 54 journal articles published between 2015 and 2025 that applied ML to assess the cooling performance of BGI. We classified BGI into seven types across three spatial scales, evaluated the metrics and indicators used to characterize the BGI cooling effects, and examined the ML algorithms, model performance, and data sources adopted in reviewed studies. We found that 2D BGI morphology metrics and BGI physical and biological metrics are most frequently used as model inputs, whereas 3D morphology metrics are used less. Surface temperature-related indicators dominate model outputs, while air temperature and human heat stress indicators, which better reflect heat health hazards, remain underutilized. In terms of methodology, ML applications are primarily based on artificial neural networks for temporal regression and tree ensemble algorithms for spatial regression. Building on these findings, we identify key research gaps and advocate that future research should develop comprehensive multi-source and multi-scale BGI databases, improve model transparency and generalizability, and integrate energy-balance information into BGI-ML studies. Finally, we propose a systematic BGI-ML modeling framework to guide future research in this field.

1. Introduction

The urban population accounted for 55.3% of the global population in 2018, and by 2050, this proportion will expand to 68.4% [1]. The rapid growth of the urban population has led to a continuous increase in people‘s demand for living and working space, triggering urban expansion. The process of urbanization has led to the deterioration of the urban thermal environment, giving rise to urban heat problems such as the urban heat island (UHI). The total human-caused global surface temperature increase from 1850–1900 to 2010–2019 is 1.07 °C [2]. Urban heat causes significant environmental, social, economic, and health consequences [3,4]. According to the Lancet Countdown, in 2023, heat-related mortality of people older than 65 years increased by a record-breaking 167% [5]. In the summer of 2022, excessively high temperatures led to 61,672 additional deaths in Europe [6]. Climate change has led to increasingly frequent extreme weather events, and the resulting economic losses are also prominent. The average annual economic losses from weather-related extreme events increased by 23% from 2010–2014 to 2019–2023 to US$227 billion [5]. The rise in urban temperatures will lead to an increase in cooling energy consumption. In eastern China, daily peak cooling loads for urban buildings during heatwaves increased by 21% to 62% [7]. In Rome, Italy, building cooling energy consumption increased by 12% in peripheral neighborhoods and by 46% in the city center [8]. The rise in electricity demand for cooling presents significant challenges for energy supply, which further causes an increase in greenhouse gas (GHG) emissions.

Blue–green infrastructure (BGI), which includes both water (pools, ponds, lakes, rivers, etc.) and vegetative (green spaces, trees, green roofs, green walls, etc.) elements, is increasingly recognized for its potential to mitigate the impact of urban overheating and decrease energy consumption [9,10,11]. The green component of BGI can regulate urban heat and may promote a more comfortable and cooler urban environment [12,13] through mechanisms of evaporation, transpiration, shading, and thermal insulation [14,15], while the blue component cools the surrounding area mainly through evaporation and radiation [16]. Different forms of BGI have been found to be effective in keeping urban environments cool. In the Taipei Metropolitan Area, natural forests can form a cooling core with a maximum air temperature (AT) decrease of 3.0 °C [13]. A study in the US emphasized the synergistic cooling effect of urban BGI, with average land surface temperature (LST) reductions of up to 3 °C in parks with dense vegetation [10]. In a study of the 2021 summer heatwave in Moscow, green spaces larger than 250 ha reduced the AT within green spaces by 1.7 °C, while green spaces between 1 and 18 ha reduced only 0.5 °C [17]. A global investigation found that green roofs provided consistent energy savings from 1.1 to 7.3 kWh/m² a year, and the highest pedestrian-level AT reduction of 0.80 °C [18].

Currently, the research on the BGI cooling effect mainly focuses on quantifying the cooling impact [18,19,20,21], discovering the important factors affecting the cooling effect [10,22,23], and the BGI optimization design [23,24,25]. The previous research can be divided into three themes [26,27,28]: (1) on-site measurement to directly collect meteorological data; (2) remote sensing (RS) techniques used to calculate cooling indices, such as cooling distance and cooling intensity; and (3) estimation and prediction of urban temperature through modeling. Frequently, simple and multiple linear regression were employed to evaluate the association between urban climate and BGI variables [26].

In the fields of ecology and urban climatology, artificial intelligence (AI)-based data-driven predictive models are receiving increasing attention due to their advantages in handling complex, nonlinear relationships. However, existing reviews have not yet systematically synthesized the application of ML to the cooling effects of BGI. Tao et al. [29] explored the integration of computational fluid dynamics (CFD) and ML in understanding the synergistic effects of urban green infrastructure for urban heat mitigation and air quality improvement. Shaamala et al. [30] summarized AI-driven green infrastructure optimization research for addressing climate change from six perspectives: air quality, biodiversity and ecosystem security, energy efficiency, public health, heat island mitigation, and water management. Wang et al. [31] investigated the application of ML for predicting air temperature in urban canopies, concluding that tree-based models are suitable for spatial predictions, while neural network models excel in temporal predictions. Li et al. [32] reviewed studies on landscape metrics that reflect the impact of spatial configuration on urban green spaces, noting that while some studies used ML methods, they did not conduct a systematic analysis of ML techniques.

Although previous review studies have discussed AI-driven urban climate research or BGI from broader perspectives, the rapidly expanding body of ML-based empirical studies on the cooling effects of urban BGI—green roofs [33,34,35], street trees [36,37], urban parks [38,39,40], and urban green space networks [41,42,43]—has not yet been systematically synthesized across different BGI types and spatial scales. These studies have provided new insights into the mechanisms underlying BGI’s cooling effects in urban environments. However, several common challenges persist across this body of research: (1) Research on BGI at similar scales has not received sufficient attention. Although BGI types are diverse, those of similar scales tend to exhibit comparable cooling effects, and many research methodologies can be mutually informative. (2) The factors influencing BGI’s cooling effects are not effectively screened, requiring the development of more efficient input datasets to improve predictive accuracy and prevent overfitting. Additionally, the evaluation metrics for cooling effect lack multi-dimensionality and tend to be simplistic. (3) The selection of ML algorithms and the model-building process is often arbitrary, lacking a systematic guiding framework.

To address these issues, we conducted a systematic review of ML-based studies on the cooling effects of urban BGI, focusing on three aspects: (1) A scale-based classification of BGI, including their area, cooling range, and analytical methods; (2) a synthesis of influencing variables and evaluation metrics to inform the construction of ML model inputs and outputs; and (3) a review of ML algorithms, model performance, and data sources used in existing studies.

In the Discussion section, we advocate the development of comprehensive multi-source and multi-scale BGI databases, underscore the importance of improving model transparency and generalizability, and identify the limited incorporation of energy-balance mechanisms in current studies. Finally, we propose a systematic BGI-ML modeling framework that aligns ML tasks with the study of cooling effects across various types of BGI. This framework emphasizes classifying BGI based on spatial scales and outlines key research steps, such as defining input and output parameters and selecting the appropriate ML models. It is designed to provide a comprehensive structure for future research, facilitating the use of ML to more effectively model the cooling effects of BGI.

2. Materials and Methods

Following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidance [44], we searched the English peer-reviewed articles published from 1 January 2015 to 1 August 2025 from the Web of Science database. The retrieval keywords of the database were divided into three groups:

(1)

ML: “machine learning” OR “artificial intelligence” OR “deep learning”.

(2)

Cooling effect: “cooling effect” OR “urban heat mitigation” OR “cooling island” OR “thermal comfort” OR “air temperature” OR “surface temperature” OR “energy saving” OR “cooling energy” OR “cooling load”.

(3)

Urban blue–green infrastructure: “blue-green infrastructure” OR “urban green infrastructure” OR “urban green space” OR “urban blue infrastructure” OR “urban blue space” OR “vertical greening system” OR “green façade” OR “living wall” OR “green roof” OR “street tree” OR “urban park” OR “urban forest” OR “water body” OR “urban river”.

The first group determines the key research methods. The second group determines the assessment methods and indicators of the cooling effect. The third group determines BGI as the research object, including BGIs in different forms and scales. These keywords were intended to identify studies that have investigated the BGI cooling effect through ML models. We entered these three groups of search strings in the “topic” query field, which were connected with the “AND” logical operator. Duplicates were removed from the initial search records. The screening process, eligibility, and exclusion criteria are detailed in Figure 1. The last search was performed on 30 September 2025.

In the initial screening, the filter of Web of Science for document types was utilized to select research articles. Types such as reviews and conference papers were excluded. Then, through title and abstract screening, non-English-written articles were excluded, and articles that met the relevant themes were initially included. During the full-text deep screening, the articles that met the requirements for subsequent analysis were identified. The research object must be a specific BGI. The research must use ML to predict, evaluate, or classify the BGI cooling effect. Additionally, our review focused on empirical studies and took a cautious approach to the application of simulation techniques. The targeted application of simulation methods can compensate for the limitations of empirical data and enhance data quality. However, when simulation data serves as the primary dataset, the input and output parameters of ML models are highly dependent on the design of the simulation software. This reliance on simulated data may reinforce prior assumptions, potentially overshadowing the influence of empirical scientific evidence. While studies relying solely on simulation data contribute to exploring ML application approaches, they fall outside the scope of this review. Therefore, 8 studies were excluded. A total of 6 articles meeting the eligibility criteria were added through backward and forward searching. Finally, we obtained 54 studies for further analysis (

Table A1. A list of the reviewed studies.

No.	Year	Author	City (Nation)	Climate Zone	BGI Types
1	2025	Xie et al. [76]	Shenzhen (CN)	Cwa	RG
2	2025	Yan et al. [55]	Xi’An (CN)	Cwa	GR
3	2025	Yin et al. [59]	421 parks in Henan Province (CN)		BGS
4	2025	Liu et al. [72]	24 cities (CN)		BGS
5	2025	Ling et al. [82]	Guangzhou (CN)	Cfa	BGS
6	2025	Sanchez-Cordero et al. [57]	Granada (ES)	Csa	GR
7	2025	Liu and Li [66]	Hangzhou (CN)	Cfa	BGS
8	2025	Lai et al. [77]	Guangzhou (CN)	Cfa	GR
9	2025	Liu and Qian [88]	Shijiazhuang (CN)	Dwa	BGIN
10	2025	Wang et al. [95]	Hangzhou (CN)	Cfa	RG
11	2025	Yu et al. [45]	229 global cities		BGIN
12	2025	Sheng et al. [49]	8 cities in the Yangtze River Delta region (CN)		RG
13	2025	Li and Cheng [83]	Beijing (CN)	Dwa	RG
14	2025	Shen et al. [69]	Xiamen (CN)	Cfa	WB
15	2025	Feng et al. [75]	Beijing (CN)	Dwa	BGIN
16	2025	Zhong et al. [89]	Shanghai (CN)	Cfa	BGIN
17	2025	Zhang et al. [81]	Beijing (CN)	Dwa	BGIN
18	2025	Jato-Espino et al. [47]	A Valencian Community (ES)		BGIN
19	2024	Yan et al. [64]	Beijing (CN)	Dwa	BGIN
20	2024	Wu et al. [41]	Paris (FR)	Cfb	BGIN
21	2024	Stumpe and Marschner [74]	Ruhr Metropolitan Region (DE)		GS
22	2024	Kafy et al. [58]	Austin (US)	Cfa	GR
23	2024	Wang et al. [33]	Guangzhou (CN)	Cfa	GR
24	2024	Sun et al. [65]	Shanghai (CN)	Cfa	BGIN
25	2024	Zhang et al. [96]	Shanghai (CN)	Cfa	BGS
26	2024	Islam et al. [38]	Kolkata (IN)	Am	BGS
27	2024	Ibsen et al. [42]	8 cities (US)		BGIN
28	2024	Chen et al. [85]	Urumqi (CN)	Bwk	BGIN
29	2024	Daemei et al. [54]	Rasht (IR)	Cfa	GW
30	2024	Zhang et al. [97]	3 cities in Fujian Province (CN)		BGIN
31	2024	Wang et al. [48]	six cities near 30°N (CN)		WB
32	2024	Stumpe et al. [51]	5 cities (DE)		GS
33	2024	Li et al. [35]	Shanghai (CN)	Cfa	GR
34	2024	He et al. [40]	596 global cities		BGIN
35	2023	Yang et al. [67]	Tianjin (CN)	Dwa	BGIN
36	2023	Liu et al. [39]	Zhejiang Province (CN)		BGS
37	2023	Kang et al. [71]	Nanjing (CN)	Cfa	WB
38	2023	Zhao et al. [46]	806 global cities		BGIN
39	2023	Lyu et al. [70]	Yinchuan (CN)	BWk	BGIN
40	2023	Lin et al. [43]	Shenzhen (CN)	Cwa	BGIN
41	2022	Liu et al. [36]	Wuhan (CN)	Cfa	RG
42	2022	Zhang et al. [68]	Urumqi (CN)	Bwk	BGIN
43	2022	Chen et al. [86]	Urumqi (CN)	Bwk	BGIN
44	2022	Kraemer and Kabisch [63]	Leipzig (DE)	Cfb	GS
45	2022	Wei et al. [79]	Chengdu (CN)	Cfa	RG
46	2021	McCarty et al. [61]	Dallas (US)	Cfa	GS
47	2021	Daemei et al. [53]	Rasht (IR)	Cfa	GW
48	2021	Sun et al. [73]	Shanghai (CN)	Cfa	GS
49	2020	Wei et al. [52]	Wuyishan (CN)	Cwa	GR
50	2020	Asadi et al. [56]	Austin (US)	Cfa	GR
51	2020	Helletsgruber et al. [37]	4 European cities		RG
52	2019	Osbornea and Alvares-Sanches [98]	Southampton (UK)	Cfb	BGIN
53	2019	Duncan et al. [90]	Perth (AU)	Csa	BGIN
54	2012	Pandey et al. [78]	Ujjain (IN)	Aw	GR

).

The following information was extracted from the articles: (1) The city and Köppen climate zone where the study site was located; (2) the type, area, cooling range, and analysis method of the studied BGI; (3) influencing variables and evaluation metrics of BGI cooling effect; (4) the ML task and applied algorithm; (5) accuracy evaluation of the model; and (6) the sources of BGI data and output data. All information extracted from the articles can be found in Table S1.

3. Results

3.1. Overview of Studies

Firstly, we examined the geographical distribution of 54 studies. Figure 2 shows the spatial distribution of the studied cities from the reviewed articles.

Most of the studies are from Asia (39, 72.2%), followed by Europe (8, 14.8%), North America (3, 5.8%), and South America (1, 1.9%), while the rest are global studies (3, 5.8%). Chinese cities are mostly studied (34, 63%). A total of 40 studies were conducted based on sites from individual cities, while 14 studies were carried out on sites from multiple cities, countries, or even worldwide. Since most of the multi-city studies did not provide detailed information on climate zones, only the climate zones of the single-city studies were counted. These studies mainly focused on temperate climates, particularly humid subtropical climates (Cfa) (19, 35.2%), followed by Dwa (6, 11.1%). The multi-city studies took into account the differences in climate zones. Three global studies [40,45,46] and a study in the U.S. [42] covered multiple climate zones. Other multi-city studies used administrative boundaries or geographical concepts when defining the scope of the research to obtain sufficient representative research samples and to make the research results more universal. For instance, a study was conducted in the Valencian Community of Spain, which comprises 542 municipalities [47]. A study in China explored the cooling effect of water bodies in six cities along the Yangtze River [48]. A study concentrated on eight cities in the lower Yangtze River Delta that exhibited a typical subtropical monsoon climate [49].

3.2. Scale-Based Classification of BGI

We categorized the GBI in reviewed studies into seven types, under three scales (Figure 3): (1) micro-scale (18, 33.3%), including green wall (GW, 2, 3.7%), green roof (GR, 9, 16.7%, and roadside greening (RG, 7, 13.0%); (2) local-scale (15, 27.8%), including green space (GS, 5, 9.3%), water body (WB, 3, 5.6%), and blue and green space (BGS, 7, 13.0%); and (3) city-scale (21, 38.9%), represented by the blue and green infrastructure network (BGIN, 21, 38.9%). The classification of different BGIs across scales is based on Oke’s scale analysis of urban elements [50].

GW and GR are types of BGI integrating vegetation into walls and roofs. RG refers to vegetation along urban roads and streets. GS refers to land predominantly covered with vegetation, such as parks, urban forests, and farmlands. WB includes rivers and lakes within urban areas. BGS refers to land covered with vegetation and water bodies, such as river parks and wetlands. The key factor in distinguishing BGS from GS and WB is whether BGI includes both water bodies and vegetation elements, and whether the study simultaneously considers the impact of both. BGIN refers to the system encompassing all blue–green patches identified within a city. The cooling effect of BGIN is associated with the interconnectedness (i.e., adjacency) of various blue–green patches, which distinguishes BGIN from other types of BGI, such as individual GS, WB, and BGS.

The cooling range of BGI in this study refers to the spatial extent used to estimate its cooling effect, determined by the analysis method of BGI cooling performance. The cooling range can serve both as an indicator for evaluating the cooling effect of BGI and as a means to define the scope of BGI research. In urban environments, focusing on the cooling effect of BGI on its surrounding built area is particularly relevant for research. Therefore, it is crucial to define a reasonable influence range for the cooling effect of BGI. We summarized the methods for defining the cooling range and its values when studying BGI across different types (Figure 3), providing references for related studies. Overall, the grid-based method has been widely adopted, as it meets the need for large sample sizes in ML applications. For studies that consider only the cooling effect within the BGI coverage [33,35,51,52], the cooling range is not included in the synthesis.

Research on GW is limited and primarily based on field measurements. Studies on GW typically employ control groups, such as varying distances from measurement points to the GW or comparisons between green and bare walls, to assess the impact of GW on outdoor or indoor air temperature [53,54]. In contrast, research on GR is more abundant and diverse, covering both individual building-level observations and the potential for large-scale implementation at the city level. At the individual building level, most GR studies focus on small-scale extensive herbaceous roofs, typically around 10 m² in area. These studies generally examine the impact of GR on the covered area, including indoor and outdoor building surface temperatures (BST) [33,34,35]. From the perspective of large-scale GR implementation in urban areas, some studies investigated the effect of GR on land surface temperature (LST). For instance, Yan et al. [55] identified all GRs in Xi’an from the year 2024, which have an average area of 604 m² and a total area of 1.41 km². Their study found that GRs reduced LST by an average of 2.49 °C compared to surrounding bare roofs. Additionally, some studies have mapped building rooftops and modified indices, such as the normalized difference vegetation index (NDVI) and normalized difference built-up index (NDBI), to predict the temperature reduction after the implementation of GR [56,57,58].

Research on RG has been conducted on both urban roads and internal paths within urban blocks (e.g., campuses, parks). Studies on urban roads generally obtain spatial form data of the built environment through street-view images along urban roads at fixed intervals, such as 100 m or 80 m, as analysis samples. Studies on internal paths of urban blocks typically set sampling points and collect data through field measurements and on-site surveys. RG studies primarily focus on the impact of plant configuration along roadsides on the surrounding environment. For example, Sheng et al. [49] collected samples from areas of 400 m² in parks, residential areas, squares, and roads, and found that the arbor-grass structure and mixed planting of coniferous and broad-leaf trees provided more favorable cooling and humidification effects.

In studies of GS, WB, and BGS, a minimum threshold of 1 ha is commonly set to identify significant cooling effects of these types of BGI [39,40,48,59]. The average BGI area in most studies typically ranges from 10 to 100 ha (Figure 3). To assess BGI’s cooling performance, these studies predominantly employ buffer zone methods. BGI serves as a heat sink, reducing surrounding temperatures. Generally, as the distance from the BGI boundary increases, a corresponding increase in thermal indicators (e.g., LST, AT) is observed. The point at which this temperature reaches a turning point is typically considered the effective cooling range of the BGI [60].

Although the turning point method provides a precise measure of cooling effects, it can be time-consuming. Therefore, alternative approaches, such as the fixed radius method [61], equal area method [59], and equal radius method [62], are frequently used. These methods strike a balance between the accuracy of the evaluation and the simplicity of their implementation. The buffer zone scales for LST evaluation usually range from several hundred meters, with larger scales applied for large water bodies. For instance, Wang et al. [48] used a 2500 m buffer zone for rivers exceeding 70 m in width. The buffer zone for AT evaluation is typically smaller in scale, such as 50 m [63].

Research has shown that local-scale BGI can lead to a reduction in LST of approximately 2 °C. In line with this, Liu et al. [39] observed that most parks in Zhejiang Province, China, exhibited a cooling effect, with an average park cooling intensity (PCI) of 1.98 °C. Furthermore, He et al. [40] found that LST in tree-covered areas was an average of 2.13 °C lower than in the surrounding built-up areas during the summer. BGI at the local scale can also effectively reduce AT, though more scientific evidence is needed. Kraemer and Kabisch found that urban parks could lower the AT of the surrounding area by about 1 °C [63].

Studies on BGIN primarily focus on the spatial morphology of BGI patches, such as their connectivity, aggregation, and fragmentation. These studies often analyze BGIN by dividing it into grid units; hence, the grid-scale effect must be considered [64]. Grid scales vary widely, ranging from 30 m to 3 km (Figure 3). For studies examining the BGIN impact at the community level, grid scales typically fall within the range of several hundred meters [65,66]. At the urban level, researchers must consider various factors, including the cooling distance of multiple blue–green patches [67], the evaluation scale of urban livability [68], and the sufficiency of information for morphological analysis [43]. These factors often necessitate the use of grid scales spanning several kilometers. To capture a more refined cooling gradient, some studies adopt finer grid resolutions. For instance, Wu et al. [41] assessed the cooling efficiency of green space networks using a 70 m grid, while Zhao et al. [46] employed a 30 m grid to analyze the cooling efficiency of urban tree networks.

3.3. Influencing Metrics and Evaluation Indicators of BGI Cooling Effect

To investigate the nonlinear relationships between influencing metrics and the cooling effects of BGI, a well-structured parameter set is essential for ML analysis. This requires the selection of appropriate evaluation indicators as model outputs and relevant influencing metrics as inputs. In practice, beyond the primary variables of interest, studies often incorporate additional commonly recognized factors, such as meteorological metrics, to improve predictive performance. In this review, we synthesized the influencing metrics and evaluation indicators adopted in existing studies, categorized them, and analyzed their frequency of use across spatial scales, thereby summarizing the characteristic parameter configurations employed at different scales.

3.3.1. Influencing Metrics of BGI Cooling Effect

We summarized six categories of influencing metrics identified in the reviewed studies: (1) meteorological metrics (22, 40.7%); (2) geographical metrics (13, 24.1%); (3) land cover and land use metrics (21, 38.9%); (4) human activities and physiological metrics (9, 16.7%); (5) blue–green infrastructure (BGI)-based metrics, which are subdivided into 2D BGI morphology metrics (34, 63.0%), 3D BGI morphology metrics (13, 24.1%), and BGI physical and biological metrics (30, 55.6%); and (6) built environment (BE)-based metrics, which are also subdivided into 2D BE morphology metrics (17, 31.5%), 3D BE morphology metrics (18, 33.3%), and BE physical metrics (14, 25.9%). Figure 4 shows the frequency with which these influencing metrics were used as input parameters and their distribution across different scales.

Overall, researchers typically prioritize the selection of input parameters in the following order: BGI-based metrics, followed by BE-based metrics, and then the remaining four categories. The dual focus on natural elements and the built environment is a common structure in the input parameter set [41,69,70], while other parameters provide supplementary metrics with more diverse characteristics.

BGI-based metrics, the most frequently used input parameters, emphasize the morphological, physical, and biological characteristics of BGI. Among these, 2D BGI morphology metrics are the most commonly studied, followed by BGI physical and biological metrics, while 3D BGI morphology metrics are less commonly used. The coverage and geometric characteristics of vegetation and water bodies—such as landscape shape index, edge density, and patch density—are crucial in determining the extent and intensity of the cooling effect. These factors have consistently emerged as the most important influencing factors in various studies [39,40,41,48,64,67,71,72]. Additionally, BGI physical and biological metrics, such as NDVI and NDBI, which reflect the distribution of vegetation and water bodies as well as plant growth status, along with more specific parameters like tree age and species, which describe the biological characteristics of vegetation, are also widely applied. 3D BGI morphology metrics, which capture the morphological characteristics and biomass of vegetation more precisely, are gaining attention. Therefore, the contribution of 3D BGI metrics to predictive models is crucial [37,40,41,51].

The second most frequently used metrics are BE-based metrics. The focus on 3D BE morphology metrics and 2D BE morphology metrics is relatively similar, followed by BE physical metrics. Both BEM-2D parameters (e.g., building density and building area) and BEM-3D parameters (e.g., building height and floor area ratio) reflect construction intensity and complex urban morphology, which indirectly indicate the size and distribution of BGI. Moreover, the complex and variable urban climate (e.g., temperature, wind direction, and wind speed) further affects BGI’s cooling effect [42,73]. Additionally, high-density built environments, with more impervious surfaces and less BGI coverage, indirectly affect the cooling potential of BGI. In central urban areas, where thermal issues are more pronounced, this can lead to a more prominent cold island effect associated with BGI [48]. Among BE physical metrics, NDBI, which reflects the distribution of buildings and impervious surfaces, is widely applied. Other metrics, such as surface albedo and street orientation, are used to assess solar radiation exposure in urban areas [37].

Among the remaining four categories, meteorological and land cover/land use metrics are the most frequently used, while geographical and human activities and physiological metrics were found in only a few studies. Meteorological metrics are often used as input parameters to improve prediction accuracy. For example, solar radiation was used for predicting LST [56]. Similarly, land cover and land use metrics, reflecting the distribution of pervious and impervious surfaces and urban functional zones (e.g., residential, commercial, and industrial), are widely used in studies [41,42,61]. Geographical metrics are employed to describe terrain features and the location of the study area, especially in studies involving BGI with mountainous characteristics [74]. Human activity metrics, such as points of interest and nightlight data, which indirectly reflect urbanization, have also been applied. For instance, studies on heat exposure or heat equity at the city scale have used human activity factors such as population density and house prices as input parameters [66,75]. Additionally, a study on thermal comfort incorporated human physiological metrics measured by wearable devices [76].

For studies at the micro scale, meteorological metrics, along with BGI physical and biological metrics, are the most commonly used metrics. At the local and city scales, the most frequently utilized metrics include 2D BGI morphology metrics, as well as BGI physical and biological metrics, and land cover and land use metrics. The key distinction is that studies at the city scale place a greater emphasis on the application of 2D BGI morphology metrics.

3.3.2. Evaluation Indicators of BGI Cooling Effect

The indicators used to evaluate the cooling effect of BGI can be classified into three categories: (1) Cooling capacity indicators that include cooling intensity indicators (13, 24.1%) and cooling efficiency indicators (4, 7.4%); (2) thermal environmental indicators that include surface temperature (ST)-related indicators (29, 53.7%), air temperature (AT)-related indicators (6, 11.1%), and urban heat island (UHI) indicators (2, 3.7%); and (3) heat health indicators that include thermal comfort indicators (3, 5.6%) and heat risk indicators (2, 3.7%). The specific descriptions of indicators and their distribution at different scales are shown in Figure 5.

Overall, ST-related indicators and cooling intensity indicators are the two most commonly used evaluation indicators. This is largely due to the widespread use of LST as a key indicator, with most studies deriving the cooling intensity indicators based on LST.

For studies at the micro scale, ST-related indicators, including both building surface temperature and LST, are the most commonly used. Measured building surface temperature has been applied in small-scale experimental GR studies [33,35,52,77], while LST is used to assess the cooling effects of GR implementation in urban areas [56,57,58]. For instance, one study compared the LST of GRs with that of the surrounding bare roofs [55]. Additionally, heat flow and air temperature are also measured as evaluation indicators in some observational studies. For example, one study evaluated heat reduction between experimental GRs and bare roofs [78]. Moreover, two studies on GW used outdoor and indoor air temperature as indicators [53,54]. Additionally, some studies on RG have incorporated heat health indicators. For example, two studies examined subjective thermal comfort along pathways within a park and a campus [76,79].

For studies at the local-scale, cooling intensity indicators and ST-related indicators are mostly used. Specifically, park cooling intensity (PCI) and LST are frequently employed as evaluation indicators. PCI is generally calculated by comparing temperature indices (e.g., LST, AT) between the BGI and its surrounding buffer zone [39,59,72]. Relative indicators like PCI can more intuitively represent the cooling effect of BGI. Another commonly used approach is to directly use LST within the BGI coverage area or its buffer zone as an evaluation indicator [38,48,51,74].

For studies at the city-scale, LST is more commonly used than other indicators. Spatial patterns of LST in the studied city area are derived, based on which the average LST of each grid unit is calculated. In addition to directly using LST as an evaluation indicator, some studies also used cooling efficiency and urban heat island (UHI) indicators. Cooling efficiency indicators are typically calculated by determining the negative regression coefficient between BGI coverage and LST within the grid units [41,45,46]. UHI indicators involve urban heat island intensity and a binary classification for identifying the presence or absence of UHI phenomena [43,47]. Furthermore, a few studies have employed AT and UTCI as evaluation indicators [42,66], as these indicators offer a more accurate representation of human perception of the environment.

3.4. ML Task Types and Models

ML is well-suited for capturing nonlinear relationships and handling high-dimensional variables with complex interactions, which has motivated its increasing use in studies examining the drivers of BGI cooling effects. In existing research on multi-scale BGI cooling performance, ML applications mainly involve regression (51, 94.4%) and classification (6, 11.1%) tasks, with regression dominating current practice. Regression analyses are primarily conducted using supervised learning algorithms, while classification tasks are addressed using both supervised methods and unsupervised approaches, the latter typically implemented through clustering techniques.

Regression tasks include temporal regression (7, 13.0%) and spatial regression (44, 81.5%). Temporal regression is used to predict the future thermal performance of BGI based on historical data, which is commonly applied in studies conducting long-term continuous observation. Spatial regression considers the spatial variation in input and output parameters. Studies on spatial regression obtained spatial patterns of landscape features and cooling performance indicators of discrete or contiguous areas, and used them as input and output data, which were integrated based on geographical locations, whereas classification was used less, mainly in studies that graded thermal parameters or cooling effect (e.g., Kang et al. classified cooling effect of WB into four classes [71]) or carried out binary classification (e.g., comfortable vs. uncomfortable [76], UHI vs. non-UHI [47]).

The used frequencies of ML algorithms for different tasks are shown in Figure 6. For temporal regression, all studies employed artificial neural networks, including Multilayer Perceptron (MLP, 2), Convolutional Neural Network–Long Short-Term Memory Network (CNN-LSTM, 1), Graph Convolutional Network–Fully Connected Network (GCN-FCN, 1), Genetic Model–Back Propagation Neural Network (GA-BP, 1), Temporal Convolutional Network–Gated Recurrent Unit (TCN-GRN, 1), and Artificial Neural Network (ANN, 1). In this article, all unspecified artificial neural networks are classified under the term ANN.

For spatial regression, the tree ensemble algorithms account for the majority: Random Forest (RF, 27), Extreme Gradient Boosting (XGBoost, 6), Boosted Regression Tree (BRT, 5), Categorical Boosting (CatBoost, 2), Extremely Randomized Trees (ERT, 2), and Gradient Boosting Decision Tree (GBDT, 2). A few studies utilized artificial neural networks: ANN (3).

The algorithms used for classification include: K-means clustering (K-means, 3), Bayesian Network (BN, 1), Linear Discrimination (LD, 1), Support Vector Machine (SVM, 1), and XGBoost (1).

Overall, the prevalence of particular algorithms largely reflects a task-data fit in BGI cooling studies. Temporal regression typically draws on dense time-series observations, where neural networks, especially LSTM-based models, are well-suited to capture nonlinear temporal dependencies. In contrast, spatial regression is usually built on heterogeneous tabular geospatial predictors linked to thermal indicators. In this setting, tree ensemble models are often preferred because they handle nonlinear threshold responses and are relatively robust to multicollinearity. They also provide variable-importance measures that facilitate model interpretation.

We summarized the dataset sizes used for modeling with different algorithms (Figure A2). The dataset sizes for artificial neural networks range from 200 to 40,305, with a median of 2160. For tree ensemble algorithms, the sizes range from 28 to 53,319, with a median of 430. Additionally, one study using LD and K-means had a dataset size of 35 [71]. Most datasets are split into 80% training set and 20% testing set, or 70% training set and 30% testing set.

3.4.1. Temporal Regression

Temporal regressions are applied exclusively at the micro scale. The data used for training temporal regression models are typically collected by fixed meteorological stations, with a time resolution of either hourly or sub-hourly intervals. The historical data typically spans a period of one season or one year to ensure statistical representativeness. After excluding the effects of weather conditions and random variability, sufficient data samples are available for model training. For example, a study conducted a year-long observational experiment with data recorded every hour, and selected a continuous seven-day sunny dataset from each month as the input data for their model, which resulted in a total of 2016 datasets [33]. Additionally, a few studies have employed continuous multi-day observational data to train models for short-term predictions [53,54].

Most studies that perform temporal regressions employ hybrid artificial neural network algorithms. Due to the ability of artificial neural networks to perform end-to-end processing and self-optimization through backpropagation, it is common to construct hybrid neural network algorithms to leverage the strengths of various models and optimize performance. Wang et al. applied a CNN-LSTM model, suitable for capturing spatiotemporal features in sequential data [33]. Li et al. combined GCN with a traditional four-layer neural network, enabling the model to more effectively focus on the factors that significantly influence the final output and their relationships with other factors [35]. Wei et al. employed a GA-BP neural network to improve the optimization process of predicting the impact of different soil conditions on the outer building surface temperature [52]. Lai et al. developed a TCN-GRU model that captures both short-term and long-term features, designed to forecast both the inner and outer building surface temperature of GR [77].

3.4.2. Spatial Regression

Spatial regression is commonly used in studies at the local and city scales, where decision tree-based ensemble algorithms are most frequently applied. Decision trees recursively partition data into subsets to identify influential features and threshold values that impact thermal indicators [31]. These tree ensembles include bagging (e.g., RF, ERT) and boosting (e.g., XGBoost, BRT, CatBoost, and GBDT). Bagging trains trees independently and aggregates their predictions, ensuring computational efficiency and robustness against overfitting. In contrast, boosting focuses on misclassified instances, iteratively enhancing prediction accuracy by minimizing residual errors through weighted votes from each tree [80]. The majority of studies on spatial regression employed bagging-family algorithms (29, 53.7%), and RF is the most frequently used algorithm (27, 50.0%), far exceeding other algorithms. More than a quarter of the studies used boosting-family algorithms (16, 29.6%), in which XGBoost (6, 11.1%) and BRT (5, 9.3%) are the most frequent two.

The reliable predictive performance of RF has garnered considerable attention from researchers. It has shown high accuracy across a range of predictive tasks in different contexts. For example, a study comparing the cool island effect of WB across six cities found RF and ERT to be the top performers [48]. Liu and Li [66] demonstrated that RF outperformed XGBoost and LightGBM in predicting UTCI distribution at a 30 m resolution. Furthermore, the application of geographically weighted optimization enhanced RF’s ability to capture spatial variations, outperforming traditional RF models [81].

XGBoost has been utilized to model the impact of influential factors on the cooling effects of BGI due to its excellent fitting performance. For instance, one study employed XGBoost to explore the nonlinear relationships of cooling effects represented by UTCI across different housing price tiers [66]. Liu et al. used XGBoost to spatialize the importance of key factors affecting park cooling intensity [39]. Additionally, other boosting-family algorithms have also demonstrated strong model performance and have been applied. Several studies employed BRT for its ability to mitigate overfitting and enhance predictive performance through the boosting technique [41,46,65]. Ling et al. modeled the relationships between PCI and potential influencing factors under different UHI gradients using CatBoost [82]. In another study, GBDT was used to analyze the influence of factors on PCI across 24 Chinese cities [72].

Some other studies on spatial regression have utilized artificial neural networks (3, 5.6%). When sufficient training data is available, artificial neural networks can also exhibit high predictive performance in spatial regression. For example, two studies exploring the impact of GR on LST used more than 10,000 samples to train ANN models, with the final models achieving prediction accuracies (R²) of 0.77 and 0.78, respectively [56,58].

3.4.3. Classification

Various algorithms have been applied to classification tasks in the reviewed studies, encompassing both unsupervised and supervised methods. Unsupervised algorithms, such as K-means, are commonly used for clustering. For example, Kang et al. [71] employed K-means to cluster WB cooling effects and subsequently trained LD to develop a classification prediction model. Clustering by K-means is also valuable for identifying outliers in datasets, which is crucial for defining a reliable statistical foundation [51,74]. Supervised algorithms are used for analyzing specific types of urban environments or classifying meteorological phenomena. BN has been applied in scenario analysis, as shown by Lyu et al. [70], who utilized BN to examine the combinations of factors that could result in lower LST scenarios. Moreover, some algorithms are preferred by researchers for their specific advantages. For instance, Jato-Espino et al. [47] selected SVM for its effectiveness in spatial data classification to distinguish between UHI and non-UHI in grid units. Similarly, Xie et al. [76] employed XGBoost for binary predictions of thermal comfort and discomfort, owing to its strong interpretability.

3.5. Model Evaluation

R² and RMSE are commonly used as indicators for evaluating model performance in most studies.

Table 1. Model accuracies in different BGI scales.

Scale	R2 and RMSE of Surface Temperature: Mean [Minimum, Maximum Values]
	R2	RMSE (°C)
Micro-scale	0.867 [0.69, 0.982]	0.745 [0.26, 1.83]
Local-scale	0.702 [0.477, 0.99]	1.111 [0.065, 2.63]
City-scale	0.619 [0.304, 0.898]	0.897 [0.44, 1.36]

summarizes the prediction accuracies of models using surface temperature (the most frequently used output parameter) as the output variable across different scales. Studies at all scales have the potential to achieve high model accuracy, but research at the local and city scales faces an increased risk of lower model performance. This may be due to the larger number of influencing factors on BGI cooling effects at these scales, making it difficult to determine the optimal input parameter structure. Additionally, temporal regression models at the micro scale tend to achieve higher accuracy. Therefore, the overall trend is that model performance decreases as the BGI scale increases. The R² values of studies at the micro, local, and city scales are 0.867, 0.702, and 0.619, respectively. The RMSE values for models at the local and city scales are also higher than those at the micro scale.

3.6. Data Source

We compiled the sources of the BGI data and the evaluation indicators’ data in the reviewed studies. Overall, sources of BGI data include field survey (12, 22.2%), street imagery (3, 5.6%), urban statistical data (7, 13.0%), open and official map data (19, 35.2%), aerial data (7, 13.0%), and remote sensing (34, 63.0%), as well as simulation and re-analyzed data (2, 3.7%). Sources of evaluation indicators’ data include field measurement (14, 25.9%), questionnaire (2, 3.7%), remote sensing (38, 70.4%), and simulation/re-analyzed data (2, 3.7%). Figure 7 shows the distribution of data sources in studies at different scales.

Most micro-scale studies rely on field observation experiments, where BGI data and evaluation indicators are primarily obtained through field surveys or measurements [33,35,52,53]. The collected BGI data typically include biological characteristics such as vegetation type, leaf area index, and vegetation height, as well as physical properties such as leaf emissivity and reflectivity. Additionally, studies on roadside greening often use street imagery and open maps to identify and locate greenery along urban roads [36,83]. Common evaluation indicators include air temperature near the vegetation canopy and building surface temperature. Some studies examining the large-scale application of GR explore their cooling potential across different urban areas using airborne lidar and remote sensing data [56,57,58]. In these studies, more abstract parameters, such as NDVI and NDBI, are used to represent GR characteristics, replacing specific biological parameters. In contrast, LST is employed as the primary indicator to assess the cooling effect.

Local-scale studies primarily extract BGI information by using RS and map data. RS data includes various products (e.g., images, spectral data, and elevation data), which allow for the extraction of abundant BGI data such as land cover, landscape indices, and tree canopy height [40,69]. Map data and urban statistical data are used to accurately delineate the boundaries of BGIs. For example, Liu et al. [72] obtained park lists from official city websites and extracted park boundaries from open online maps, subsequently deriving the morphological and biological characteristics of BGI from RS data. Most studies evaluated the cooling effect of BGIs by deriving LST within their coverage or influence range from RS data [38,39,84]. In contrast, only one study adopted a more detailed approach by measuring air temperature using fixed stations at a finer scale [63].

City-scale studies primarily rely on RS methods to obtain BGI data and evaluation indicator data, with some studies using RS data exclusively for their research [46,64,65,67,68,81,85,86,87,88,89]. Existing land cover products offer standardized classifications, facilitating the rapid identification of BGI distributions. For instance, Ibsen et al. [42] collected land cover data for eight U.S. cities, specifically extracting tree and grassland data for their analysis. Additionally, more detailed aerial data can offer finer classifications of BGIs. For instance, a study classified urban vegetation into grass, shrubs, and trees using aerial photography, which provides vegetation height data with a vertical resolution of 1 mm and a horizontal resolution of 20 cm [90]. The use of multi-source and multi-scale data is relatively rare in city-scale studies, often due to data availability and integration challenges. One comprehensive study integrated field survey data, airborne lidar data, and RS data for BGI analysis, offering detailed insights into how landscape configuration can influence LST [70]. Most studies rely on LSTs derived from RS as evaluation indicators. Air temperature is less commonly used as an evaluation indicator, primarily due to the limited distribution of meteorological stations, which makes it difficult to obtain reliable data at the city scale. Only one study successfully measured AT by deploying a network of sensors across multiple urban areas [42]. Additionally, some studies have employed machine learning and simulation techniques to derive higher-resolution temperature data from a limited set of observation stations [66,75].

4. Discussion

Based on an analysis of 54 BGI-ML studies, this research shows that ML has been widely applied across diverse BGI types and spatial scales, demonstrating a strong capacity for capturing nonlinear relationships and identifying key drivers of the cooling effect. Despite these advances, several aspects still warrant further attention to enhance the robustness and applicability of BGI-ML research. In response, we outline a systematic modeling framework to guide future studies in this field (Figure 8). We offer a concise checklist that translates the framework steps into recommended protocols and key reporting items, to make the framework more actionable and practically useful. (Checklist S1).

4.1. The Applicability and Potential Bias of the Results

We summarized the distribution characteristics of studies across different scales (Figure A1) and evaluated the applicability and potential bias of the results. Research on micro-scale BGI is based entirely on individual cities, with the majority located in temperate climates. Therefore, the applicability of the results regarding micro-scale BGI to other climate zones requires further investigation. Nearly half of the local-scale studies are based on multiple cities, while the remaining studies, based on individual cities, are primarily located in temperate climates, with a smaller number in tropical and continental climates. As a result, the results of studies on local-scale BGI are relatively robust, but are primarily applicable to temperate regions. For city-scale studies, about a quarter are based on multiple cities, while studies based on individual cities are evenly distributed across arid, temperate, and continental climates. The results of city-scale studies are generally applicable, but there is limited evidence from tropical and other climate zones.

4.2. Building Multi-Scale and Multi-Source BGI Databases

A large, high-quality database is essential for advancing ML applications to BGI, yet such data remain scarce and uneven across cities. Most studies rely on freely available remote sensing products, which capture BGI morphology and spatial heterogeneity but still provide only a top-down and 2D view that overlooks many local and 3D characteristics shaping urban thermal conditions. Only a small number of cities maintain detailed official BGI inventories. For example, several European cities, such as Frankfurt in Germany and the Valencian Community in Spain, have developed urban tree databases that include information on species, age, trunk dimensions, and canopy structure [47,51]. Although self-collected datasets using aerial imagery or airborne lidar offer richer spatial detail, they require specialized equipment and substantially increase research costs. Recent computer-vision methods now allow the extraction of fine-scale information, such as sky-view factor and green-view index, from street-view images, providing a low-cost complement to remote sensing. Despite these advances, the overall data landscape remains fragmented, underscoring the need for coordinated efforts to develop open, multi-source, and multi-scale BGI databases that can support consistent modeling across diverse urban contexts.

Although remote sensing dominates large-scale BGI cooling analyses, it introduces uncertainties such as spatial-resolution limits, incomplete coverage, weather-related contamination (e.g., clouds), and atmospheric-correction and retrieval errors. These uncertainties can propagate through ML models and may shift the feature-importance ranking of key drivers (e.g., BGI area or morphology metrics). Future studies should strengthen uncertainty-aware evaluation by using multi-temporal imagery to derive representative LST, cross-checking LST products from different sensors, conducting sensitivity checks, and reporting key data-quality information and plausible uncertainty ranges to reduce their influence on interpretation.

4.3. Evaluating Model Transparency and Generalizability

Our review shows that issues of transparency and generalizability are insufficiently addressed in current BGI-ML studies. Transparent reporting is essential for assessing methodological rigor, yet many studies provide only limited information on training data characteristics, sample sizes, hyperparameter tuning procedures, and comparative model performance. Given the inherently opaque nature (black box) of ML methods, these details must be clearly disclosed to enhance the credibility of results and enable meaningful cross-study comparison. For example, Stumpe et al. provided the average values of LST, tree characteristics, and spectral parameters of GS samples in multiple cities in Germany [51]. Generalizability is an equally important but often overlooked concern. ML models are highly sensitive to the climatic and geographic characteristics of the training data, meaning models developed in a single city rarely transfer well beyond their local context without further validation [62,65]. Recent efforts to compile datasets from cities across multiple climate zones [40,45,46] offer a promising direction for producing more broadly applicable insights.

4.4. Advancing ML from Correlation to Mechanism

Current BGI-ML studies remain focused on correlations in microclimate outcomes, yet a mechanism-oriented shift is needed to understand how cooling is actually produced. Most models treat microclimate variables or thermal comfort indices as response metrics and compare predictor importance, but such correlations cannot reveal the physical processes behind the cooling effect. In reality, BGI cools the thermal environment by reshaping local energy exchanges rather than directly lowering air temperature, mainly through changes in radiative fluxes and the balance between latent and sensible heat [24]. ML models trained only on final microclimate outputs, without explicitly considering these energy pathways, therefore face clear limits in interpretability. Incorporating energy-balance information directly into BGI-ML frameworks may be a promising direction for future research. Integrating such physical mechanisms into ML models could help examine how different drivers influence radiative, latent, and sensible heat fluxes and reveal whether synergies or trade-offs exist among vegetation and water elements [91]. Building on this approach, combining energy-flux analysis, ML, and diverse observational data would enable the field to move beyond outcome prediction toward a more mechanistic understanding of BGI cooling, advancing research from empirical correlations to a physical-data-integrated paradigm.

4.5. A Systematic Framework for BGI-ML Modeling

4.5.1. Matching ML Tasks with BGI Types and Indicator Characteristics

A systematic BGI-ML modeling framework should begin by ensuring that the selected ML tasks align with the characteristics of both BGI elements and thermal environment indicators. Different types of BGI, such as street trees, parks, green roofs, and blue–green combinations, operate at varying spatial scales and are influenced by distinct biophysical factors. These differences determine not only which input variables are relevant but also how cooling effects should be measured across scales. In addition to the seven common BGIs discussed in the reviewed studies, our research indicates that the study of a newly emerging BGI facility—water spraying systems [92]—remains to be explored in the context of ML techniques. Likewise, the nature of the thermal environment indicator determines the appropriate ML task. Time-series measurements from continuous observations are suited to temporal regression [33,35], spatially continuous or discrete environmental surfaces are suited to spatial regression [40,41], and subjective thermal sensation votes are suited to classification tasks [76,79].

When a single ML task cannot adequately capture the multi-dimensional interactions between BGI and urban thermal conditions, a staged multi-task modeling strategy is often more effective. Typically, an initial task, such as clustering [51] or regression [66], is used to structure the data or refine key input features, followed by a second task focused on scenario prediction or correlation analysis. For example, Lyu et al. [70] first used Random Forest to identify the landscape metrics most strongly associated with LST, and then applied a Bayesian Network to simulate cooling scenarios and support spatial decisions. Similarly, Liu and Li [66] generated high-resolution UTCI through ML regression and subsequently used MGWR and SHAP-XGBoost to analyze spatial drivers and equity implications. Such staged modeling pathways enable targeted use of ML tools at each step and support a more nuanced understanding of BGI-driven cooling processes.

4.5.2. Defining Input and Output Parameters in BGI-ML

Defining appropriate input and output parameters is fundamental to constructing reliable BGI-ML models. Common thermal environment indicators include LST, AT, other microclimatic parameters (e.g., RH), and composite thermal comfort indices (e.g., UTCI, PET). Compared with the widespread availability of high-resolution LST data, ground-based meteorological observations (e.g., AT) were less frequently used due to limited spatial coverage. To obtain higher-resolution AT surfaces, interpolation or regression-based approaches—including ML-based nonlinear regression—can expand sparse observations into continuous urban-scale datasets. For example, Ibsen et al. [42] predicted AT for all 60 m grid cells using ML models trained on sensor-based AT measurements. Similar techniques have been used to downscale LST [83,91], such as in Yan et al. [55], who refined 30 m LST to 10 m resolution to capture roof-scale temperature variations.

Given the widespread use of LST as an outcome variable in large-scale BGI cooling studies, its interpretation warrants caution. LST characterizes surface thermal conditions and does not directly capture pedestrian-level exposure or thermal comfort. Human-centered conclusions are therefore better supported by pedestrian-level comfort metrics (e.g., UTCI/PET/SET) or microclimate modeling where feasible.

For BGI-related parameters, most studies rely on conventional 2D landscape metrics to represent BGI characteristics, while more detailed inputs remain underexplored. Vegetation-related 3D metrics (e.g., tree height [41]) and non-spectral attributes (e.g., species, age [51]) have shown contributions in several studies [41,51], yet they are often treated as supplementary variables in neighborhood- or city-scale analyses dominated by 2D metrics. Future BGI-ML research would benefit from integrating a broader set of BGI descriptors across structural and compositional dimensions.

To reduce the risk of overfitting and improve model efficiency, it is essential to identify and retain only the most influential input variables. This is especially relevant in BGI-ML tasks, where many spatial variables are involved, but only a few significantly influence outcomes, such as LST or AT. Common approaches include correlation analysis, feature selection algorithms such as recursive feature elimination (RFE), dimensionality reduction techniques like principal component analysis (PCA), and model-specific interpretability tools such as SHAP or sensitivity analysis. These methods help streamline the input space while preserving predictive performance. For example, Ling et al. [82] applied SHAP analysis to reduce input variables from 24 to 5–18 without compromising model accuracy. Similarly, Feng et al. [75] identified just five key predictors that explained over 86% of the variance, highlighting the effectiveness of targeted variable selection.

4.5.3. Selecting and Evaluating ML Algorithms in BGI-ML

Selecting an appropriate ML algorithm is a critical step in BGI-ML modeling. Given the diversity of tasks involved—such as regression, classification, or mixed workflows—algorithms can be broadly grouped according to their typical applications. However, relying on a single model is seldom sufficient, as performance can vary substantially depending on data structure, spatial resolution, and indicator type. For example, two studies [85,86] using similar datasets from the same urban area evaluated SVM, RF, and GBDT, yet identified different best-performing models due to minor differences in input variables. A practical strategy is to identify a set of candidate algorithms based on prior research and known task requirements, then train and compare them within the specific study context.

A standard ML workflow in BGI-related studies generally involves training, validation, and testing stages to ensure robust model performance. For spatial regression tasks with limited data, cross-validation is often used to compare models and adjust hyperparameters, though an independent test set is still recommended for unbiased evaluation [66]. In temporal regression, datasets should be split chronologically to avoid information leakage caused by temporal autocorrelation. Careful design of the training and evaluation process is essential, as model performance can vary significantly depending on data structure and split strategy.

Best-practice evaluation should align the validation setup with the data structure and study purpose. To avoid overly optimistic results, models should be tested on data that are kept separate from the training set in space or time, such as using different locations for spatial prediction or chronological splits for temporal prediction. It is also important to report complementary metrics (e.g., R² with RMSE/MAE) and to benchmark against a simple baseline to demonstrate the added value of ML. If transferability is a goal, testing in other cities or climate contexts, together with clear reporting of sample size and tuning, provides stronger evidence. Finally, interpretation can help check whether the learned patterns are consistent with known cooling processes rather than incidental correlations.

4.5.4. Interpreting ML Results for BGI Applications

Once the ML model has been developed, interpreting the outputs is essential for drawing meaningful and actionable conclusions that can inform BGI planning and design. Model interpretability strengthens the reliability of the findings by clarifying whether key variables genuinely contribute to predicted thermal outcomes [93]. To avoid over-claiming transparency, it is important to distinguish local explanations (for specific cases) from global attribution (overall driver importance), and to separate these from physically plausible inference based on established cooling processes. Interpretability can be achieved either through the use of transparent models, which are inherently explainable, or through post hoc techniques applied to black-box models [93]. Post hoc interpretation improves understanding of model behavior but does not make a black-box model fully transparent. Among the reviewed studies, only models like Bayesian Networks and Linear Discriminant Analysis are considered transparent, while most tree ensemble and deep learning models require post hoc interpretation. SHAP is a widely used technique for interpreting feature contributions across ML models [94], but its results should be reported with a clear scope (local vs. global) and treated as associative evidence rather than mechanistic proof. To support evidence-based and transferable BGI-ML applications, researchers should prioritize transparent models when possible or apply appropriate post hoc interpretation tools for black-box models, and report the explanation target and key settings needed for reproducibility.

5. Conclusions and Limitations

BGI represents an important nature-based solution for mitigating urban overheating, while ML techniques provide an effective means of integrating multi-source data to capture the complex and nonlinear nature of BGI cooling effects. Building on the growing application of ML in urban climate research, this review synthesizes recent advances in ML-based assessments of BGI cooling performance.

Using a scale-based classification framework, we integrate evidence from 54 empirical studies to characterize seven BGI types across three spatial scales (micro, local, and city), together with their typical area ranges, cooling extents, and analytical approaches. We further synthesize six categories of influencing metrics used as model inputs and three categories of evaluation indicators used as outputs, highlighting their patterns of application across scales. From a methodological perspective, ML algorithms are grouped into artificial neural networks for temporal regression, tree ensemble algorithms for spatial regression, and other approaches for classification, and their respective application characteristics are examined.

To advance ML applications in BGI research, we emphasize the need to develop comprehensive multi-source and multi-scale BGI databases, improve transparency in dataset construction and model training to enhance generalizability, and strengthen the integration of physical mechanisms, particularly energy-balance processes, into data-driven models. Finally, we propose a systematic BGI-ML modeling framework to support more robust, interpretable, and transferable research on its cooling effects.

Despite its contributions, this review has some limitations. It mainly relies on English-language journal articles from Web of Science, which may overlook important information from other databases or types of publications. This constitutes a systematic limitation. Additionally, we found that since most studies focus on temperate regions, this may limit the generalizability of the conclusions in this review. These issues should receive more attention in future research.