A Spatial Mismatch Analysis of Blue–Green–Gray Infrastructure for Urban Cooling: Linking Supply, Thermal Pressure, and Heat-Sensitive Demand

Abstract

Urban green spaces are increasingly recognized as cooling infrastructure for climate resilience, environmental health, and spatial equity. However, their planning value depends not only on where green spaces are located but also on whether cooling supply is translated into lower thermal pressure and aligned with heat-sensitive demand. This study examines blue–green–gray infrastructure cooling supply, thermal regulation, human-relevant thermal pressure, and supply–demand mismatch across 1385 census tracts in Maryland, USA. An Infrastructure Cooling Supply Index (ICSI) was constructed from green, blue, and gray infrastructure components and evaluated against maximum land surface temperature (LSTmax), Heat Index, extreme heat days, and a Human-Relevant Thermal Pressure Index (HTPI) integrating ambient heat intensity and recurrent extreme heat exposure. Explainable machine learning was used to identify the relative contributions of individual infrastructure components, and cooling supply was compared with socially differentiated heat-sensitive demand. The results show that higher ICSI is significantly associated with lower LSTmax, Heat Index, and HTPI. The association was strongest for surface thermal conditions, remained significant for Heat Index and integrated human-relevant thermal pressure, and was comparatively weaker for recurrent extreme heat days, indicating that infrastructure cooling supply affects thermal outcomes through distinct pathways. Cooling supply was also spatially uneven, with stronger provision generally associated with greater vegetation and tree-canopy coverage, higher park provision, larger water-area proportions, and lower gray-infrastructure pressure. Explainable GeoAI results identify impervious cover, tree canopy, water-area proportion, road density, and park provision as the principal contributors to tract-level thermal pressure. The mismatch analysis identifies 433 census tracts, or 31.3% of all Maryland tracts, as low-supply–high-demand priority areas. These tracts reveal substantial spatial convergence among elevated heat-sensitive demand, insufficient cooling infrastructure, and persistent thermal pressure. This study provides a tract-level approach for targeting cooling interventions where thermal pressure, vulnerability, and infrastructure deficits converge.

1. Introduction

As extreme heat becomes a more persistent feature of urban and regional climate risk, urban green spaces are increasingly being repositioned as infrastructure for adaptation rather than as residual open land or recreational amenities alone [1,2,3,4]. This shift is important for land-use planning because the capacity to cope with heat depends not only on meteorological conditions, but also on how vegetated areas, tree canopies, parks, water bodies, impervious surfaces, road networks, and residential populations are arranged across space. Green spaces can moderate heat through shading, evapotranspiration, reduced surface heat storage, and the creation of cooler outdoor environments [5,6]. At the same time, their cooling functions are conditioned by adjacent blue spaces and by heat-amplifying gray infrastructure, including impervious surfaces and road networks [7,8,9]. Treating green spaces as multifunctional cooling infrastructure therefore requires an analytical framework that connects ecological cooling capacity with thermal outcomes, human exposure, and uneven social needs.

This infrastructural perspective is further clarified through blue–green infrastructure, a concept that has been widely used to describe interconnected networks of natural, semi-natural, and water-related infrastructure components that provide multiple ecosystem, climate, public-health, and social benefits across urban systems [10,11,12]. A census tract may contain designated parkland but still have limited cooling capacity if park areas are dominated by impervious surfaces, sparse canopy, or fragmented vegetation. Conversely, cooling benefits may be generated by tree canopy, vegetated land, wetlands, riparian corridors, or other non-park landscapes that are not always captured by formal park indicators. Water bodies may further contribute to localized cooling, while roads and impervious surfaces can intensify heat storage and disrupt the continuity of cooling landscapes. For this reason, green-space cooling infrastructure should be evaluated as part of a broader blue–green–gray system.

Evaluating green spaces as cooling infrastructure also requires moving from potential supply to realized thermal regulation. Ecosystem service and land-use planning research has long emphasized that ecological structures become socially meaningful when their functions generate benefits and respond to demand [13,14]. In the context of urban heat adaptation, this means that vegetation coverage, park area, and green-space accessibility should not be assumed to produce equivalent cooling benefits across all neighborhoods. These indicators are valuable, but they cannot by themselves demonstrate whether stronger cooling supply corresponds to lower thermal pressure. The relationship between green-space structure and urban heat is mediated by land-cover composition, surface properties, humidity, urban morphology, background climate, and scale of analysis [15,16,17]. Therefore, a planning-relevant assessment needs to test whether areas with stronger cooling-supportive composition experience lower heat outcomes after accounting for a broader spatial context.

A further complication is that urban heat has multiple causes [18,19]. Land surface temperature (LST) is indispensable for assessing surface energy balance and the response of built and vegetated surfaces to solar radiation [8,9]. However, LST is not equivalent to the heat stress experienced by residents. Human-relevant thermal pressure depends on apparent temperature, humidity, repeated heat days, housing conditions, behavioral exposure, and adaptive resources [20,21]. For example, a tract with moderate surface temperature may still experience substantial heat-health pressure if high humidity and frequent extreme heat days coincide with vulnerable populations. Conversely, a high-LST tract may not always represent the highest human exposure burden. Green-space cooling research therefore needs to distinguish surface-oriented thermal regulation from indicators that more directly capture environmental health relevance.

Building on this perspective, it becomes important to move beyond simply identifying where cooling potential exists and to consider how such potential is actually experienced by people. In this study, we use the term cooling service to capture this transition from biophysical capacity to socially relevant benefit. Specifically, cooling service refers to the benefit that emerges when infrastructure cooling supply is effectively translated into reduced thermal pressure for populations in need of heat relief. Cooling services become socially consequential when they reach the places and populations that need them most. Prior studies have shown that heat exposure, urban heat-island intensity, and heat-risk-related land-cover conditions are unevenly distributed across racial, socioeconomic, and neighborhood groups [22,23,24,25]. At the same time, access to green space and its associated health benefits are shaped by environmental justice, social determinants of health, and the uneven distribution of urban greening [26,27,28]. These findings suggest that evaluating the public value of green-space cooling requires more than the size, type and number of green spaces. Heat-related demand in terms of how much people need relief from heat is determined by a mix of factors, including the overlap of thermal pressure, age-related susceptibility, social isolation, socioeconomic disadvantage, limited healthcare access, and constrained household adaptation such as lack of air conditioning [29]. If green-space cooling supply is concentrated in areas with lower heat-health demand, while vulnerable communities face weaker cooling supply, green infrastructure may reproduce rather than reduce spatial inequity.

This concern is especially relevant because urban greening is not automatically equitable. New parks, tree-planting programs, greenways, and waterfront improvements may improve environmental quality, but they can also generate uneven benefits if they are not aligned with the needs of historically underserved or heat-sensitive communities [28,30]. Therefore, the key planning question is not simply whether green spaces cool urban environments but whether cooling benefits are spatially aligned with heat-health demand. A cooling-service equity perspective requires identifying where low cooling supply coincides with high heat-sensitive demand and where interventions such as tree-canopy expansion, vegetation restoration, blue–green corridor enhancement, park redesign, and gray-surface retrofitting may generate the greatest adaptation value.

Research in Maryland has established vegetation and tree canopies as important urban cooling assets, particularly in Baltimore. Earlier neighborhood-scale research linked elevated LST to limited vegetation and socioeconomic disadvantage, while a recent Baltimore city–county study identified distinct tree-canopy thresholds associated with reduced cooling-energy demand [31,32]. However, the evidence remains largely Baltimore-centered and component-specific, leaving a statewide gap in integrating blue–green–gray cooling supply, multiple thermal outcomes, and heat-sensitive demand.

Building on this Maryland-specific limitation and the broader literature, three gaps motivate the present study. First, existing work often evaluates green-space amount, vegetation cover, or landscape composition as proxies for cooling capacity, without sufficiently validating whether such supply corresponds to realized thermal outcomes. Second, studies frequently privilege land surface temperature while giving less attention to apparent heat and recurrent extreme heat exposure, even though these indicators are more directly connected to human thermal burden. Third, research on green-space equity often documents unequal access or unequal exposure but less often integrates cooling supply, realized thermal regulation, environmental-health demand, and spatial mismatch within one tract-level framework. These gaps limit the ability of planners to identify not only where green infrastructure exists but where additional cooling investment is most needed.

Responding to these limitations, this study examines urban green spaces as cooling infrastructure across 1385 census tracts in Maryland, USA. The statewide design extends the predominantly Baltimore-centered literature by examining whether the cooling functions identified in metropolitan studies also hold across heterogeneous urban, suburban, coastal, rural, and upland environments. It further expands the existing emphasis on tree canopies or individual heat outcomes by evaluating the combined structure of blue, green, and gray infrastructure against multiple indicators of realized thermal regulation and heat-sensitive demand. The Baltimore–Washington corridor is characterized by dense development, extensive road networks, high impervious-surface intensity, and fragmented green spaces, making it an important context for examining gray-infrastructure pressure on thermal regulation. In contrast, suburban counties, agricultural areas, the Chesapeake Bay region, the Eastern Shore, and the mountainous areas of western Maryland provide contrasting green and blue landscape conditions. This spatial diversity makes it possible to evaluate whether infrastructure cooling supply operates consistently across heterogeneous urban, suburban, coastal, rural, and upland settings and whether cooling supply is spatially aligned with heat-sensitive demand.

The analysis of green-space cooling infrastructure is organized around four questions. First, how can the cooling-infrastructure function of urban green spaces be quantified for census tracts while accounting for green, blue, and gray components? Second, are multidimensional infrastructure cooling effects linked to different realized thermal regulations? Third, which green, blue, and gray components make the strongest contributions to human-relevant thermal pressure? Fourth, where do low cooling supply and high heat-sensitive demand coincide, and which tracts should be prioritized from a cooling-service equity perspective?

To answer these questions, this study constructs an Infrastructure Cooling Supply Index (ICSI) to represent green-space cooling supply within a blue–green–gray infrastructure system; it captures the relative balance between cooling-supportive green/blue components and heat-amplifying gray infrastructure. Then, we test its association with multiple thermal outcomes using fixed-effects regression, where a Human-Relevant Thermal Pressure Index (HTPI) was derived. This study couples eXtreme Gradient Boosting (XGBoost) and Shapley additive explanations (SHAPs) to identify nonlinear component-level contributions to HTPI and spatializes a Heat-sensitive Cooling Demand Index (HCDI) derived from HTPI accounting for social vulnerability. By comparing the spatial pattern of ICSI and HCDI, this study identifies the low-supply–high-demand priority areas in cooling services.

By linking thermal regulation, environmental health, and cooling-service equity, this study advances three contributions. Conceptually, it reframes urban green spaces as cooling infrastructure for heat adaptation. Methodologically, it connects cooling supply, realized thermal outcomes, human-relevant heat pressure, and demand-side vulnerability in a single spatial framework. Practically, it offers a tract-level diagnostic approach for targeting green-space cooling interventions where climate exposure, health sensitivity, and infrastructure deficits converge.

2. Literature Review

2.1. Urban Green Spaces as Cooling Infrastructure

Urban green spaces are increasingly recognized as multifunctional infrastructure for climate adaptation, local climate regulation, and human well-being [33,34,35]. Their cooling function is produced through multiple biophysical pathways, including shading, evapotranspiration, reduced heat storage, and modification of surface–atmosphere exchange. However, the cooling capacity of green spaces depends not only on the amount of vegetation or parkland but also on how green, blue, and gray elements are combined within urban landscapes. From a landscape ecological perspective, this reflects the broader pattern–process relationship: spatial composition and configuration shape ecological processes, including urban thermal regulation [9,36].

Landscape ecology research has demonstrated that landscape composition and configuration jointly shape thermal processes, although their effects vary across spatial scales and urban contexts [18,37]. This perspective suggests that cooling capacity should be evaluated through the interaction of multiple landscape components rather than isolated green-space indicators. A recent systematic review of urban green-space configuration similarly concluded that larger, more aggregated, and more complex-shaped green spaces often generate stronger cooling effects, while also noting that empirical findings differ across cities, metrics, and methodological designs [17].

Vegetation is one of the most widely studied green-space cooling components in urban thermal regulation. Its cooling effect is usually explained through shading, evapotranspiration, reduced surface heat storage, and modification of near-surface microclimate. Systematic evidence suggests that urban greening generally lowers local temperatures, although the magnitude of cooling varies by vegetation type, spatial extent, background climate, and measurement method [7]. Tree canopies are especially important because they can directly reduce solar radiation reaching built surfaces and pedestrians. Ziter et al. showed that tree canopies and impervious surfaces interact in scale-dependent ways to influence daytime summer heat, implying that vegetation effects should be assessed together with heat-amplifying urban surfaces rather than in isolation [8]. For this reason, indicators such as vegetation coverage and tree-canopy coverage are often used to represent the green components of infrastructure cooling supply.

Water bodies constitute another important component of urban thermal regulation. Blue spaces can contribute to local cooling through evaporative processes, high heat capacity, and the creation of cooler surface conditions relative to surrounding built-up areas. In landscape-based heat adaptation studies, water coverage is therefore commonly treated as a cooling-supportive element, especially in regions where rivers, bays, wetlands, reservoirs, or coastal landscapes form a substantial part of the local environment. However, the effect of water is also spatially contingent: its cooling contribution depends on size, exposure, surrounding land cover, and the degree to which nearby populations or built environments are connected to blue-space cooling effects. This makes water-area proportion a meaningful landscape indicator for evaluating cooling supply at neighborhood or census-tract scales.

In contrast, impervious surfaces and road characteristics represent gray infrastructure that tends to amplify thermal pressure. Impervious surfaces store and re-radiate heat, reduce evapotranspiration, and are often associated with dense built environments, while roads contribute to heat absorption, traffic-related anthropogenic heat, and the fragmentation of cooling landscape elements. Prior research has repeatedly shown that impervious cover is positively associated with higher urban temperatures and can weaken or offset the cooling contribution of tree canopies and other vegetation [8]. Road length and road density can therefore be interpreted not merely as transportation indicators but as proxies for linear gray infrastructure that may intensify or redistribute urban thermal pressure. Including these indicators helps distinguish cooling-supportive landscape elements from heat-amplifying infrastructure.

Parks occupy a more ambiguous position in urban thermal regulation. On the one hand, parks are important providers of recreational, cultural, and potentially cooling ecosystem services. Larger parks or more spatially aggregated green spaces may generate stronger cooling islands and influence surrounding thermal environments [17]. On the other hand, formal park provision does not necessarily equal vegetation-based cooling function. Park boundaries may include lawns, paved surfaces, water bodies, tree canopy, sports facilities, buildings, and other mixed land-cover types. As a result, park area, park proportion, and park area per capita are better interpreted as indicators of formal green-space provision, while vegetation coverage and tree canopy more directly represent the biophysical components that drive cooling processes. This distinction is important for assessing whether planned green-space supply translates into realized thermal regulation. Therefore, existing evidence suggests that cooling infrastructure assessment should distinguish formal green-space provision from the broader landscape composition that generates thermal regulation. This motivates the construction of ICSI in this study.

2.2. Green-Space Cooling and Human-Relevant Thermal Pressure

A body of research has linked green infrastructure, urban heat mitigation, and health-relevant exposure. For example, studies of urban green infrastructure have shown that tree canopy and other green elements can reduce extreme surface temperatures and may contribute to heat-risk mitigation for vulnerable populations [38,39]. However, much of the inequality literature focuses on who is exposed to more heat, who receives fewer cooling benefits, or how heat burdens differ across social groups. These are essential questions for environmental justice, but they do not always explain the landscape ecological process through which cooling supply is produced, how it translates into realized thermal regulation, or why it may fail to match demand. In other words, cooling inequality has been well established as a distributional problem, but it remains less fully connected to the pattern–process logic of landscape ecology and the supply–demand logic of ecosystem services.

This gap is partly related to how urban heat is measured. Land surface temperature is widely used because it is spatially continuous, remotely sensed, and closely linked to surface energy balance. It is therefore highly useful for assessing surface thermal conditions and landscape-driven surface thermal regulation. Prior work has cautioned that surface temperature alone may not adequately represent the thermal conditions experienced by people at the street level, where humidity, shade, wind, radiation, and activity patterns also matter [40]. Therefore, a landscape that reduces LST may not necessarily reduce human-perceived heat pressure to the same extent, and an area with high surface temperature may not always correspond to the highest human-relevant thermal burden.

Existing studies have extensively documented unequal heat exposure and unequal access to cooling benefits. Recent evidence shows that urban green spaces provide substantial cooling, but the magnitude of cooling capacity and resident-level cooling benefit varies strongly across world regions and cities [41]. Studies in the United States and other urban contexts have shown that low-income populations, people of color, and racially segregated communities are often exposed to higher levels of urban heat or heat-risk-related land-cover conditions [25,26,27,42]. These studies establish that cooling inequality is not merely a matter of whether green space exists but also whether cooling benefits are spatially aligned with where people live and experience heat. However, empirical analysis about whether cooling infrastructure itself is spatially aligned with populations experiencing greater thermal pressure is still limited.

2.3. Cooling-Service Equity and Supply–Demand Mismatch

A cooling-service equity perspective requires moving beyond the question of whether green spaces provide potential cooling supply toward the question of whether such supply is translated into realized thermal regulation and spatially aligned with heat-sensitive demand. In this study, cooling service is used in an ecosystem-service sense: it refers to the socially relevant cooling benefit that may be generated by blue–green infrastructure components and delivered to residents through reduced thermal pressure. This concept is analytically separated into four elements: infrastructure cooling supply, realized thermal regulation, heat-sensitive cooling demand, and supply–demand mismatch.

Burkhard et al. established an influential spatial framework for mapping ecosystem service supply, demand, and budgets, showing that service provision cannot be assessed only from the supply side [15]. Later studies further emphasized that mismatches may arise when ecosystem service capacity is spatially separated from demand, when demand exceeds local supply, or when potential services are not effectively delivered to populations who need them [43,44,45]. This supply–demand perspective is particularly relevant for urban cooling services because heat mitigation is both spatially produced by blue–green–gray landscape composition and socially needed by exposed or heat-sensitive populations.

In the context of urban thermal regulation, the supply–demand framework challenges a simple assumption: more green space does not necessarily mean sufficient cooling supply or realized cooling benefit, and stronger cooling capacity does not necessarily mean better alignment with heat-sensitive demand. From an equity perspective, low-supply–high-demand areas are not only service-deficit areas but also priority areas for green-space cooling intervention. Syrbe et al. proposed a national indicator of local climate regulation in German cities that explicitly relates green infrastructure cooling capacity to residential demand, thereby moving beyond the evaluation of green-space quantity alone [46]. This type of approach is important because cooling service should be assessed not only as potential ecological supply but also as a spatially distributed benefit that may or may not reach populations experiencing thermal pressure.

Recent studies have begun to apply this supply–demand logic more directly to urban cooling assessment. Wang et al. examined the spatial relationship between cooling supply and demand provided by urban green and blue spaces and demonstrated the importance of identifying areas where cooling deficits and thermal demand overlap [47]. Related studies on park cooling and accessibility have also shown that cooling effects may be unevenly distributed and may not be accessible to all neighborhoods during extreme heat events [48]. These findings suggest that cooling-service mismatch assessment therefore needs to move beyond measuring total green or blue space area and instead evaluate whether cooling resources are spatially matched with thermal exposure and vulnerable demand. Therefore, a remaining challenge is to connect infrastructure cooling supply, realized thermal regulation, and heat-sensitive demand within a unified spatial framework.

3. Study Area and Methods

This study applies a tract-level analytical framework to examine the relationship among blue–green–gray infrastructure composition, infrastructure cooling supply, human-relevant thermal pressure, and cooling supply–demand mismatch. The analysis proceeds in four steps. First, an Infrastructure Cooling Supply Index (ICSI) is constructed from green, blue, and gray infrastructure indicators. Second, multiple thermal indicators are used to evaluate realized thermal regulation, and a Human-Relevant Thermal Pressure Index (HTPI) is developed from apparent heat and recurrent extreme heat exposure. Third, a Heat-sensitive Cooling Demand Index (HCDI) is constructed by combining HTPI with vulnerability-related demand conditions. Fourth, ICSI and HCDI are cross-classified to identify low-supply–high-demand priority mismatch areas. The following sections describe the study area, variable construction, and analytical methods.

3.1. Study Area and Spatial Unit

The empirical analysis focuses on Maryland, a state-scale urban–regional system in the Mid-Atlantic United States where pronounced variation in urbanization, vegetation, water coverage, transportation infrastructure, and settlement density occurs within a relatively compact geographic area, as shown in Figure 1. This regional diversity makes Maryland suitable for examining urban green spaces as multifunctional cooling infrastructure within a blue–green–gray system. Rather than representing a single urban environment, Maryland includes multiple planning contexts within one institutional boundary, including the densely developed Baltimore–Washington corridor, low-density suburban areas, Chesapeake Bay and Eastern Shore coastal landscapes, agricultural regions, and forested uplands in the western part of the state. These contrasting settings create substantial differences in vegetation cover, tree canopy, park provision, impervious surface, road infrastructure, water-area proportion, and population distribution, all of which are central to the cooling-service cascade examined in this study.

The Baltimore–Washington corridor is characterized by dense development, extensive road networks, and high concentrations of impervious surfaces, while many coastal and rural areas contain larger shares of water, wetlands, forests, and agricultural land. The Chesapeake Bay and its tributaries further shape the state’s blue-space structure and produce strong contrasts between inland, coastal, and estuarine environments. This combination of urbanized corridors, suburban expansion zones, coastal landscapes, and rural uplands allows this study to evaluate green-space cooling supply across a broad range of blue–green–gray conditions without leaving a shared state-level planning and governance context.

The census tract was used as the primary spatial unit of analysis. Census tracts provide an intermediate scale that is more spatially detailed than counties but more stable and policy-relevant than parcels, blocks, or individual observation points. This scale is appropriate for integrating remotely sensed landscape indicators, thermal outcomes, and sociodemographic variables, because it captures neighborhood-level differences while remaining compatible with census-based population and vulnerability data. The final analytic dataset includes 1385 Maryland census tracts after harmonizing landscape, thermal, demographic, and spatial boundary data. This sample was derived from the 2010 Maryland census-tract geography, which contains 1406 tracts statewide. The analytic sample therefore retains approximately 98.5% of Maryland census tracts. The small number of excluded tracts primarily consisted of uninhabited or zero-population tract records and was not the result of failed joins to the landscape, thermal, or census-geometry variables. For the 1385 included tracts, all key spatial fields used in the analysis, including land area, water area, internal-point latitude and longitude, and population-density information, were complete. Therefore, the final sample provides near-complete statewide coverage while excluding only a small number of tracts that do not meaningfully represent residential cooling-service demand.

3.2. Data Sources and Variable Construction

3.2.1. Infrastructure Cooling Supply Index

The Infrastructure Cooling Supply Index (ICSI) serves as an operational net cooling-supply index that summarizes the balance between cooling-supportive infrastructure components and heat-amplifying gray infrastructure. The blue–green–gray infrastructure framework provides the basis for selecting the indicators, while ICSI translates these infrastructure characteristics into a tract-level cooling-supply measure. Specifically, the green infrastructure component was represented by both biophysical vegetation indicators and formal green-space provision indicators, including vegetation coverage, tree canopy, park area, park-area proportion, and park area per capita. The blue infrastructure component was represented by percent water area, while gray infrastructure pressure was represented by percent impervious surface, total road length, and road density. All variables were harmonized to the census-tract scale through area-weighted aggregation, spatial intersection, or tract-level normalization, depending on the original data structure. The variables, data sources, data years or versions, original spatial resolutions or units, and tract-level processing methods used to construct ICSI are summarized in

Table 1. Indicators used to construct the Infrastructure Cooling Supply Index (ICSI).

Dimension	Variable	Data Source (Year)
Green infrastructure components	Percent vegetated area (PctVegeArea)	NLCD Land Cover (2019)
	Percent tree canopy (PctTreeCanopy)	NLCD Tree Canopy Cover (2019)
	Total park area (ParkArea)	Protected Areas Database (PAD) 3.0 (2020)
	Percent park area (PctParkArea)	PAD-US and Census Bureau
	Park area per capita (ParkAreaPerCapita)	PAD-US and Census Bureau
Blue infrastructure coverage	Percent water area (PctWater)	Census Bureau TIGER/Line water features (2020)
Gray infrastructure pressure	Percent impervious surface (PctImpSurface)	NLCD Impervious Surface (2019)
	Total road length (RoadLength)	Census Bureau TIGER/Line Roads (2020)
	Road density (RoadDensity)	Census Bureau TIGER/Line Roads (2021)

.

The green infrastructure component combines vegetation-based cooling indicators and formal green-space provision indicators, while recognizing that park provision is not equivalent to vegetation structure. The blue infrastructure component was represented by percent water area (PctWater), which captures tract-level blue-space presence. Gray infrastructure pressure was represented by impervious surface and road-related indicators, which capture built and transportation infrastructure likely to amplify thermal pressure.

ICSI was constructed in three steps. First, all selected indicators were standardized using z-scores to place percentage, area-based, per-capita, and density variables on a comparable scale. Second, cooling-supportive and heat-amplifying components were calculated separately. Third, ICSI was calculated by subtracting the heat-amplifying gray infrastructure component from the cooling-supportive green and blue infrastructure component. Higher ICSI values indicate stronger infrastructure cooling supply relative to gray infrastructure pressure. For mapping and visualization, the raw ICSI was rescaled to a 0–1 range.

For each variable (x), the standardized value was calculated as follows:

$$Z\left(X_i\right)={\displaystyle \frac{X_i-\bar{X}}{s_X}}$$

Second, the cooling-supportive and heat-amplifying components were calculated separately. The positive cooling component of ICSI was defined as the mean standardized value of vegetation, tree canopy, and park-related indicators:

$$ICS{I}_i^{+}=\left({\displaystyle \frac{1}{5}}\right)\left[Z\left(C_i\right)+Z\left(V_i\right)+Z\left(A_i\right)+Z\left(P_i\right)+Z\left(Q_i\right)\right]$$

The negative heat-amplifying component was defined as the mean standardized value of impervious surface and road-related indicators:

$$ICS{I}_i^{-}=\left({\displaystyle \frac{1}{3}}\right)\left[Z\left(I_i\right)+Z\left(R_i\right)+Z\left(D_i\right)\right]$$

Third, the final ICSI was calculated by subtracting the heat-amplifying component from the cooling-supportive component:

$$ICS{I}_i=ICS{I}_i^{+}-ICS{I}_i^{-}$$

A higher value of ($ICS{I}_i$) indicates stronger green-space cooling supply relative to heat-amplifying gray infrastructure, whereas a lower value indicates weaker cooling-supportive blue–green composition and/or stronger gray-infrastructure pressure. For mapping and visualization, the raw index was also rescaled to a 0–1 range:

$$ICS{I}_{i,01}={\displaystyle \frac{ICS{I}_i-min\left(ICSI\right)}{max\left(ICSI\right)-min\left(ICSI\right)}}$$

where $C_i$, $V_i$, $A_i$, $P_i$, and $Q_i$ denote percent tree canopy, percent vegetated area, log-transformed park area, percent park area, and log-transformed park area per capita, respectively. $I_i$, $R_i$, and $D_i$ denote percent impervious surface, log-transformed road length, and road density, respectively.

3.2.2. Human-Relevant Thermal Pressure Index (HTPI)

To examine whether infrastructure cooling supply is associated with thermal conditions relevant to human exposure in a robust manner, this study distinguishes among three types of thermal outcomes: maximum land surface temperature (LSTmax), Heat Index (HeatIndex), and extreme heat days (EHD). LSTmax was derived from satellite-based land surface temperature in 2019 and is used to represent surface thermal condition and surface-oriented thermal regulation. HeatIndex and EHD were derived from ERA5-Land near-surface meteorological variables in 2020. Specifically, HeatIndex was calculated from 2 m air temperature and 2 m dewpoint temperature in 2020 using the NOAA Heat Index formulation, representing humidity-adjusted apparent heat pressure [22,49]. EHD measured the number of days exceeding the 90 °F heat-index criterion, representing recurrent extreme heat exposure rather than a single thermal snapshot [50].

Considering the possible variability of these three indicators, we tested the statistical significance of the pairwise Spearman correlations, and the results are reported in

Table 2. Correlations among thermal indicators with significance level.

Variable	LSTmax	HeatIndex	EHD
LSTmax	1	−0.003 (0.914)	−0.002 (0.938)
HeatIndex	−0.003 (0.914)	1	0.907 (0.000) ***
EHD	−0.002 (0.938)	0.907 (0.000) ***	1

Significance levels: *** p < 0.001.

. The correlations between LSTmax and the two human-relevant meteorological indicators were close to zero and not statistically significant (LSTmax–HeatIndex: ρ = −0.003, p = 0.914; LSTmax–EHD: ρ = −0.002, p = 0.938). They indicate that LSTmax was not meaningfully correlated with either apparent heat pressure or recurrent extreme heat exposure in this context. By contrast, HeatIndex and EHD were strongly and significantly correlated, suggesting that they captured related dimensions of human-relevant heat exposure. In addition, we believe that this weak correspondence is also consistent with the different data sources and measurement scales of the indicators. Therefore, LSTmax was retained as a separate surface-oriented thermal outcome, while HTPI was constructed only from HeatIndex and EHD to represent human-relevant apparent and recurrent heat pressure.

The HTPI for census tract (i) was calculated as follows:

$$HTP{I}_i=\left({\displaystyle \frac{1}{2}}\right)\left[Z\left(HeatInde{x}_i\right)+Z\left(EH{D}_{HI90F_i}\right)\right]$$

where $Z\left(HeatInde{x}_i\right)$ and $EH{D}_{HI90F_i}$ denote the standardized values of Heat Index and extreme heat days for census tract (i), respectively. The HTPI will be used as an integrated thermal pressure outcome in the regression analysis and as the thermal pressure component of HCDI, which will be introduced in Section 3.2.3.

3.2.3. Heat-Sensitive Cooling Demand Index (HCDI)

To evaluate whether green-space cooling supply is spatially aligned with demand-side vulnerability, this study constructed a Heat-sensitive Cooling Demand Index (HCDI) at the census-tract scale. HCDI represents planning-oriented cooling demand generated by the overlap between human-relevant thermal pressure and vulnerability-related demand conditions. It is not a direct measure of expressed behavioral demand, such as park visitation, cooling-center use, or self-reported heat-relief needs.

HCDI includes one thermal pressure component and six vulnerability-related components. The thermal pressure component was represented by HTPI. Population sensitivity was represented by the percentage of residents aged 65 years and above and the percentage of householders aged 65 years and above living alone. Adaptive-capacity constraints were represented by the percentage of households without air conditioning, the percentage of residents below poverty, the percentage of residents without health insurance, and the percentage of adults aged 25 years and above without a high school diploma [35,51].

All seven variables were standardized using z-scores. Since higher values indicate stronger heat-sensitive cooling demand, no reverse coding was required. HCDI was calculated as the unweighted mean of the seven standardized components and was also rescaled to a 0–1 range for mapping. The z-score transformation was defined as follows:

$$Z\left(X_i\right)={\displaystyle \frac{X_i-mean\left(X\right)}{sd\left(X\right)}}$$

The HCDI for census tract i was calculated as the unweighted mean of the seven standardized components:

$$\begin{array}{c}HCD{I}_i=\left({\displaystyle \frac{1}{7}}\right)[Z\left(HTP{I}_i\right)+Z\left(PctPop{65}_i\right)+Z\left(PctHouseholderOver65LivingAlon{e}_i\right)+Z\left(PctNoA{C}_i\right)\\ +Z\left(PctPopUnderPovert{y}_i\right)+Z\left(PctNoHealthInsuranc{e}_i\right)+Z\left(PctNoHSOver{25}_i\right)]\end{array}$$

where $HTP{I}_i$ denotes human-relevant thermal pressure, $PctPop{65}_i$ denotes the percentage of residents aged 65 years and above, $PctHouseholderOver65LivingAlon{e}_i$ denotes the percentage of older householders living alone, PctNoAC_i denotes the percentage of households without air conditioning, $PctPopUnderPovert{y}_i$ denotes the percentage of residents below poverty, $PctNoHealthInsuranc{e}_i$ denotes the percentage of residents without health insurance, and $PctNoHSOver{25}_i$ denotes the percentage of adults without a high school diploma.

For mapping and comparison, a min–max normalized version was also calculated:

$${\mathrm{H}\mathrm{C}\mathrm{D}\mathrm{I}}_{\mathrm{i},01}={\displaystyle \frac{{\mathrm{H}\mathrm{C}\mathrm{D}\mathrm{I}}_{\mathrm{i}}-\mathrm{m}\mathrm{i}\mathrm{n}\left(\mathrm{H}\mathrm{C}\mathrm{D}\mathrm{I}\right)}{\mathrm{m}\mathrm{a}\mathrm{x}\left(\mathrm{H}\mathrm{C}\mathrm{D}\mathrm{I}\right)-\mathrm{m}\mathrm{i}\mathrm{n}\left(\mathrm{H}\mathrm{C}\mathrm{D}\mathrm{I}\right)}}$$

The resulting HCDI was used as the demand-side indicator in the cooling supply–demand mismatch analysis. Census tracts with above-median HCDI were classified as high-demand areas, while those with below-median HCDI were classified as low-demand areas. This demand classification was then cross-tabulated with ICSI-based cooling supply categories to identify low supply–high demand priority mismatch areas.

3.3. Analytical Strategies

This study applies a multidimensional infrastructure cooling-service cascade framework to organize the empirical analysis. The framework links cooling supply, realized thermal regulation, human-relevant thermal pressure, and cooling-service mismatch, combining fixed-effects regression, explainable GeoAI, and supply–demand typology to evaluate both thermal performance and equity-oriented planning relevance. Data processing, statistical analyses, and machine-learning modeling were conducted using Python 3.14, while spatial data processing and visualization were performed using ArcMap 10.8.

As shown in Figure 2, this study constructed an ICSI from blue–green–gray infrastructure dimensions. The second step used fixed-effects regression models to test associations between ICSI and thermal outcomes, i.e., LSTmax, HeatIndex, EHD, and the composite HTPI. County fixed effects were included to account for unobserved county-level conditions, including regional climate background, land-development history, planning context, and physiographic setting. The models also controlled for population density, water proportion, geographic location, and socioeconomic characteristics.

The fixed-effects regression model was specified as follows:

$$Y_{i,k}={\beta }_0+{\beta }_{1ICS{I}_i}+{\beta }_{2X_i}+{\mu }_c+{\epsilon }_i, k\in \left\{LSTmax,HeatIndex,EH{D}_{HI90F},HTPI\right\}$$

where i denotes census tract, c denotes county, ${\mu }_c$ represents county fixed effects, and ${\epsilon }_i$ is the error term. The coefficient of interest is ${\beta }_1$, which estimates the association between green-space cooling supply and each thermal outcome after controlling all control variables. $X_i$ includes population density, water proportion, geographic location, and socioeconomic controls.

In the third step, XGBoost and SHAP were applied to interpret nonlinear contributions of infrastructure components to HTPI [52,53]. The fourth step cross-classified ICSI and HCDI to identify cooling supply–demand mismatch types. XGBoost was used to model HTPI as a function of green, blue, and gray infrastructure indicators. SHAP was then used to interpret the relative importance and direction of each predictor’s contribution to model-predicted HTPI. Positive SHAP values indicate higher predicted HTPI, while negative SHAP values indicate lower predicted HTPI. Global importance was calculated as the mean absolute SHAP value across census tracts.

In the XGBoost model, the predicted thermal pressure for census tract i is represented as an additive ensemble of regression trees:

$${\hat{\mathrm{y}}}_i=F\left(x_i\right)={\Sigma }_{m=1}^M{f}_{m\left(x_i\right)}, f_m\in F$$

where $x_i$ is the vector of landscape predictors for census tract i, $x_i$ is the number of trees, and $f_m$ denotes the $m$th regression tree. The model is estimated by minimizing a regularized objective function:

$$Obj={\Sigma }_{i=1}^{n}l\left(y_i,{\hat{\mathrm{y}}}_i\right)+{\Sigma }_{m=1}^M\Omega \left(f_m\right)$$

where $l\left(y_i,{\hat{\mathrm{y}}}_i\right)$ is the prediction loss, and $\Omega \left(f_m\right)$ is the regularization term that penalizes model complexity. In this study, $y_i$ denotes HTPI, and the predictor set includes green-space indicators, blue-space indicators, and gray-infrastructure indicators: vegetation coverage, tree canopy, park provision, water area, impervious surface, road length, and road density.

SHAP was then used to decompose the XGBoost prediction into the additive contribution of each landscape predictor. For each census tract i, the model output can be expressed as follows:

$$F\left(x_i\right)={\phi }_0+{\Sigma }_{j=1}^p{\phi }_{ij}$$

where ${\phi }_0$ is the baseline model output, $p$ is the number of predictors, and ${\phi }_{ij}$ is the SHAP value of predictor j for census tract i. A positive ${\phi }_{ij}$ indicates that predictor j increases predicted HTPI, while a negative ${\phi }_{ij}$ indicates that it decreases predicted HTPI. The global importance of each predictor was calculated as the mean absolute SHAP value across all census tracts:

$$I_j=\left({\displaystyle \frac{1}{n}}\right){\Sigma }_{i=1}^n\left|{\phi }_{ij}\right|$$

where $I_j$ represents the average contribution magnitude of predictor j to predicted HTPI. Higher $I_j$ values indicate that a landscape component has a stronger average influence on model output.

Finally, ICSI and HCDI were cross-classified to construct the cooling supply–demand mismatch typology [47]. Census tracts were classified into four groups: high-supply–low-demand, high-supply–high-demand, low-supply–low-demand, and low-supply–high-demand. The low-supply–high-demand group was interpreted as the priority mismatch category. To evaluate the stability of the mismatch classification, priority status was coded as a binary variable, with low-supply–high-demand tracts coded as 1 and all other tracts coded as 0. Agreement between the original HCDI classification and the robustness classification excluding HTPI was assessed using the raw overlap, the retention rate of the originally identified priority tracts, overall classification agreement, and Cohen’s kappa, which adjusts the observed agreement for agreement expected by chance.

4. Results

4.1. Spatial Heterogeneity of Green-Space Cooling Supply

Before examining ICSI’s spatial pattern, its distribution was first summarized to assess whether the index provides sufficient variation for subsequent spatial and regression analyses. As shown in Figure 3, the histogram and density plot indicate that ICSI is approximately centered around zero, with most census tracts clustered near the middle of the distribution and fewer tracts located at the two extremes. This pattern is consistent with the standardized construction of the index. The descriptive statistics further show that the mean value of ICSI is 0.000, the median is 0.012, and the standard deviation is 0.915, with values ranging from −3.086 to 3.027 across 1385 census tracts. These results suggest that ICSI captures substantial tract-level variation in green-space cooling supply within Maryland’s blue–green–gray system, while maintaining a relatively balanced distribution suitable for comparing census tracts.

As shown in Figure 4, ICSI exhibits pronounced spatial heterogeneity across Maryland census tracts, indicating that green-space cooling supply is unevenly distributed across the state. Areas with higher ICSI values are not randomly distributed but form clear regional and intra-metropolitan patterns. Higher ICSI values are mainly observed in tracts with more extensive vegetation, stronger tree canopy, greater water-related landscape presence, and lower gray-infrastructure intensity. These areas are distributed across parts of Howard County, Anne Arundel County, Southern Maryland, Queen Anne’s County, Harford County, and several less densely developed suburban or rural tracts. In contrast, lower ICSI values are concentrated in more infrastructure-intensive or less vegetated areas, especially in Baltimore City and surrounding urbanized tracts, parts of Prince George’s and Montgomery Counties, and several Eastern Shore and western Maryland tracts. This pattern indicates that green-space cooling infrastructure is spatially uneven even within the same state-level planning context.

The Baltimore–Washington corridor displays substantial internal variation in cooling supply. Although this region is generally more urbanized and infrastructure-intensive, the spatial pattern is not uniformly low. Some suburban tracts with relatively high vegetation or park coverage show stronger cooling supply, while more densely developed tracts with higher impervious surface and road intensity show weaker cooling supply. This internal variation indicates that green-space cooling supply cannot be inferred from metropolitan location alone; rather, it reflects tract-level differences in the balance between cooling-supportive blue–green components and heat-amplifying gray infrastructure.

The Eastern Shore and coastal areas also show a mixed pattern. Some tracts near the Chesapeake Bay and its tributaries exhibit relatively high ICSI values, likely reflecting the role of water and vegetated landscapes. However, several tracts in Dorchester, Somerset, Wicomico, and Worcester Counties show lower cooling supply. This indicates that coastal or rural location alone does not guarantee strong infrastructure cooling supply; the balance between vegetation, water, impervious surface, roads, and park provision remains important. Similarly, western Maryland shows spatial contrasts between more forested or vegetated tracts and lower-supply areas associated with settlement corridors or infrastructure concentration.

4.2. Realized Thermal-Regulation Performance of Infrastructure Cooling Supply

This section examines whether ICSI corresponds to realized thermal-regulation performance. The analysis proceeds in two steps: first, models for single thermal outcomes are used to test whether green-space cooling supply is associated with surface thermal condition, apparent heat pressure, and extreme heat exposure; second, the HTPI model is used to evaluate whether green-space cooling supply is associated with integrated human-relevant thermal pressure. This step is essential for assessing whether urban green spaces function as cooling infrastructure rather than merely as potential land-cover resources.

4.2.1. Associations with Single Thermal Outcomes

The first set of models examined whether ICSI was associated with three single thermal outcomes: LSTmax, HeatIndex, and EHD. As shown in

Table 3. Fixed-effects regression results for single thermal outcomes.

Outcome	Variables	Estimate	Std. Error	Statistic	p-Value		R2
LSTMax	ICSI	−0.457	0.077	−5.897	<0.001	***	0.057
	log_PopDensity	−0.002	0.065	−0.028	0.977
	PctWater	0.001	0.007	0.181	0.856
	Latitude	0.322	0.73	0.441	0.659
	Longitude	0.27	0.591	0.457	0.648
	PctPopUnderPoverty	−1.779	0.878	−2.026	0.043	*
	PctNonWhite	0.144	0.314	0.457	0.648
	GiniIndex	1.808	1.045	1.731	0.084	*
	PctNoHSOver25	0.578	0.903	0.64	0.522
HeatIndex	ICSI	−0.057	0.018	−3.13	0.002	**	0.9
	log_PopDensity	0.044	0.019	2.342	0.019	*
	PctWater	−0.002	0.001	−1.809	0.071	*
	Latitude	−3.862	0.204	−18.902	<0.001	***
	Longitude	0.373	0.177	2.11	0.035	*
	PctPopUnderPoverty	0.166	0.19	0.874	0.382
	PctNonWhite	−0.166	0.058	−2.885	0.004	**
	GiniIndex	0.141	0.229	0.615	0.539
	PctNoHSOver25	0.817	0.16	5.117	<0.001	***
EHD	ICSI	−0.389	0.23	−1.691	0.091	*	0.582
	log_PopDensity	−0.253	0.197	−1.283	0.2
	PctWater	−0.271	0.035	−7.743	<0.001	***
	Latitude	−14.486	2.449	−5.914	<0.001	***
	Longitude	−10.233	2.118	−4.832	<0.001	***
	PctPopUnderPoverty	−0.936	2.072	−0.452	0.652
	PctNonWhite	−1.658	0.747	−2.221	0.027	*
	GiniIndex	−1.674	3.461	−0.484	0.629
	PctNoHSOver25	2.449	2.019	1.213	0.225

Significance levels: *** p < 0.001, ** p < 0.01, * p < 0.1.

, ICSI was negatively associated with all three indicators, although the strength and significance of the association varied across thermal dimensions. This result indicates that green-space cooling supply has measurable thermal-regulation relevance, but its performance differs depending on how urban heat is defined.

The strongest association was observed for LSTmax, although the model explained only a limited proportion of variance (R² = 0.057). The coefficient of ICSI was negative and highly significant (β = −0.457, p < 0.001), indicating that census tracts with higher infrastructure cooling supply tended to have lower LSTmax after controlling for covariates and county fixed effects. This result suggests that ICSI_raw is associated with surface-oriented thermal conditions related to land-cover composition and surface energy balance.

ICSI was also significantly and negatively associated with HeatIndex (β = −0.057, p = 0.002). This finding suggests that infrastructure cooling supply shows a detectable association with apparent heat pressure beyond surface temperature conditions. By contrast, the association with EHD was negative but only marginally significant (β = −0.389, p = 0.091). This weaker result indicates that recurrent extreme heat exposure may be less sensitive to tract-level landscape composition than surface thermal condition and apparent heat pressure, possibly because extreme hot-day frequency is more strongly shaped by broader meteorological and regional factors.

Overall, the single-outcome models indicate that ICSI was negatively associated with multiple dimensions of thermal pressure, but the statistical strength of this association varied across indicators. The association was strongest and most robust for LSTmax, remained statistically significant for HeatIndex, and was negative but only suggestive for EHD. These differences support the need to distinguish among surface, apparent, and recurrent extreme heat indicators when evaluating blue–green–gray infrastructure as multifunctional cooling infrastructure, rather than treating urban heat as a single interchangeable construct.

To evaluate whether the associations between ICSI and thermal outcomes were primarily driven by broad geographic gradients, we conducted a nested-model comparison using partial R². The results show that geographic and contextual variables explained most of the variation in Heat Index and HTPI, whereas the additional contribution of ICSI was relatively modest but statistically significant. In contrast, ICSI provided the largest incremental contribution for LSTmax, accounting for approximately 38.6% of the variance explained by the full model. These findings indicate that landscape-based cooling supply is most directly reflected in surface thermal regulation, while human-relevant thermal indicators are additionally shaped by broader atmospheric conditions.

4.2.2. Association with Integrated Human-Relevant Thermal Pressure

The second set of analysis focused on the Human-Relevant Thermal Pressure Index (HTPI), which integrates HeatIndex and EHD to represent combined apparent heat pressure and recurrent extreme heat exposure. Compared with single thermal indicators, HTPI provides a more integrated measure of heat pressure relevant to human exposure and later cooling-demand analysis, and the fixed-effects regression results are shown in

Table 4. Fixed-effects regression results for integrated human-relevant thermal pressure.

Outcome	Variables	Estimate	Std. Error	Statistic	p-Value		R2
HTPI	ICSI	−0.041	0.016	−2.523	0.012	*	0.8
	log_PopDensity	0.001	0.015	0.091	0.927
	PctWater	−0.016	0.002	−7.574	<0.001	***
	Latitude	−2.135	0.172	−12.439	<0.001	***
	Longitude	−0.43	0.15	−2.859	0.004	**
	PctPopUnderPoverty	0.007	0.15	0.044	0.965
	PctNonWhite	−0.148	0.052	−2.829	0.005	**
	GiniIndex	−0.043	0.23	−0.185	0.853
	PctNoHSOver25	0.418	0.143	2.93	0.003	**

Significance levels: *** p < 0.001, ** p < 0.01, * p < 0.1.

.

The results show that ICSI was significantly and negatively associated with HTPI (β = −0.041, p = 0.012). This indicates that census tracts with stronger infrastructure cooling supply tended to experience lower integrated human-relevant thermal pressure after controlling for population density, water proportion, geographic location, socioeconomic characteristics, and county fixed effects. The HTPI model also showed strong overall explanatory performance, with an R² of 0.800.

This finding is important because it extends the evidence from surface-oriented thermal regulation to human-relevant heat-pressure mitigation. While the LSTmax model confirms that ICSI is strongly related to surface thermal condition, the HTPI model shows that infrastructure cooling supply also has empirical relevance for integrated heat pressure constructed from apparent heat and extreme heat frequency. Therefore, HTPI provides a bridge between the thermal-regulation analysis and the subsequent cooling supply–demand mismatch assessment.

4.3. Explainable Machine Learning Analysis of Landscape Effects on Thermal Pressure

The performance of the XGBoost model was evaluated before conducting the SHAP analysis. The dataset was randomly divided into training and testing subsets using an 80:20 split. Model performance was assessed using the coefficient of determination (R²), root-mean-square error (RMSE), and mean absolute error (MAE). In addition, 5-fold cross-validation was conducted to evaluate model stability across different data partitions. As shown in

Table 5. Performance metrics of the XGBoost regression model used for SHAP analysis.

Dataset/Validation Scheme	R2	RMSE	MAE
Training set	0.88	0.526	0.398
Testing set	0.797	0.432	0.353
5-fold cross-validation	0.793 ± 0.066	0.458 ± 0.036	0.357 ± 0.012

Note: The data were randomly split into training and testing subsets at an 80:20 ratio. Cross-validation values are reported as mean ± standard deviation.

, the XGBoost model achieved strong predictive performance, with an R² of 0.797 on the testing dataset and a cross-validation R² of 0.793 ± 0.066. The relatively consistent performance between training, testing, and cross-validation results indicates that the model achieved satisfactory generalizability and avoided substantial overfitting. Therefore, the subsequent SHAP analysis was conducted to interpret the contribution of blue–green–gray infrastructure components to predicted HTPI.

4.3.1. Relative Importance of Infrastructure Components for HTPI

The XGBoost feature-importance results identify overall vegetation coverage as the dominant predictor of HTPI (Figure 5). Among all predictors, PctVegeArea showed the highest relative importance, indicating that overall vegetation coverage plays a stronger role in shaping human-relevant thermal pressure than formal park provision alone.

PctWater was the second most important predictor, highlighting the relevance of blue infrastructure in the cooling process. RoadDensity, PctImpSurface, and RoadLength_log also showed substantial importance, suggesting that gray infrastructure contributes meaningfully to the differentiation of thermal pressure across census tracts. By contrast, park-related indicators ranked lower than vegetation, water, and road-related variables. Overall, the feature-importance results show that HTPI is shaped more strongly by broad blue–green–gray infrastructure composition than by park provision alone.

4.3.2. SHAP-Based Importance and Directional Effects of Infrastructure Components

The SHAP results confirm the XGBoost feature-importance ranking and clarify the direction of each variable’s contribution (Figure 6). PctVegeArea had the largest mean absolute SHAP value, and higher values of PctVegeArea were mainly associated with lower predicted HTPI. This indicates that vegetation coverage is the strongest cooling-supportive infrastructure component in the model.

PctWater also showed a cooling-related contribution, with higher water-area proportions generally associated with lower predicted HTPI. RoadLength_log and RoadDensity showed more heat-amplifying or context-dependent patterns, indicating that road-related gray infrastructure contributes to elevated or spatially differentiated thermal pressure.

Tree canopy and park-related variables showed weaker SHAP contributions than vegetation coverage, water area, and road-related infrastructure. This result suggests that formal park provision alone does not fully capture the surface composition and vegetation structure needed for thermal pressure mitigation.

In sum, the XGBoost–SHAP results highlight three key findings: vegetation coverage is the dominant cooling-supportive predictor of HTPI, water-area proportion provides additional cooling-related contribution, and road-related gray infrastructure contributes to higher or more spatially heterogeneous thermal pressure.

4.4. Cooling Supply–Demand Mismatch

4.4.1. Supply–Demand Mismatch Continuous Map and Quadrant Typology

The normalized ICSI and HCDI were overlayed to show the spatial relationship between infrastructure cooling supply and heat-sensitive cooling demand, as shown in Figure 7. Higher HCDI values are concentrated mainly in the Baltimore metropolitan area and the Washington metropolitan corridor, while additional high HCDI tracts appear in more dispersed suburban and Eastern Shore locations.

The overlay indicates that high cooling demand does not always coincide with high infrastructure cooling supply. Several high HCDI tracts overlap with relatively low ICSI values, especially in dense urbanized portions of the Baltimore–Washington corridor. These areas represent potential cooling supply–demand stress.

This continuous overlay provides the spatial basis for the subsequent four-quadrant mismatch classification. By cross-classifying ICSI and HCDI into high and low groups using median thresholds, the analysis further identifies census tracts that fall into the low supply–high demand category (Figure 8). High supply–low demand tracts represent relatively buffered areas where infrastructure cooling supply is comparatively strong and heat-sensitive demand is comparatively low. High supply–high demand tracts represent buffered high-demand areas, where elevated heat-sensitive demand overlaps with relatively strong infrastructure cooling supply. Low supply–low demand tracts represent areas with weaker cooling supply but lower immediate demand pressure. Low supply–high demand tracts were defined as priority mismatch areas because they combine insufficient infrastructure cooling supply with elevated heat-sensitive cooling demand.

The four-quadrant classification identified 433 census tracts, accounting for 31.3% of all Maryland census tracts, as low-supply–high-demand priority mismatch areas. An equal number of tracts were classified as high-supply–low-demand areas, while high-supply–high-demand and low-supply–low-demand areas accounted for 18.8% and 18.7%, respectively. This distribution indicates that cooling supply and heat-sensitive demand are not evenly aligned across Maryland.

To assess whether the priority classification was driven by the inclusion of the HTPI, the HCDI was reconstructed using only sociodemographic components. This robustness analysis identified 412 low-supply–high-demand tracts, compared with 433 in the main analysis. Among the original priority tracts, 398 retained their classification, representing a retention rate of 91.9%; these tracts also accounted for 96.6% of the priority areas identified in the robustness analysis. Across all 1385 tracts, the two binary classifications showed an overall agreement of 96.5% and a Cohen’s kappa of 0.917, indicating very strong chance-corrected agreement. Only 35 original priority tracts were reclassified as non-priority, while 14 previously non-priority tracts entered the priority category. These results indicate that excluding HTPI produced only a modest reduction in the number of priority tracts and did not materially alter the spatial mismatch pattern.

4.4.2. Spatial Concentration of Low Supply–High Demand Priority Mismatch Tracts

To distinguish high-priority urban implementation areas from lower-density vulnerability pockets, Figure 9 overlays priority mismatch tracts with population density. The most concentrated priority mismatch areas are located in the Baltimore metropolitan area and the Washington metropolitan corridor, where low cooling supply, high heat-sensitive demand, and high population density overlap.

Priority mismatch tracts also appear outside the two largest metropolitan cores, including suburban, peri-urban, and Eastern Shore locations. These areas should be interpreted as localized vulnerability pockets rather than necessarily the largest population-exposure priorities.

The population-density overlay refines the interpretation of the mismatch typology. The low-supply–high-demand category identifies relative cooling deficits, while population density helps identify where interventions may benefit larger numbers of residents. From this perspective, the Baltimore–Washington corridor represents the most practically significant concentration of priority mismatch areas.

5. Discussion

5.1. Advancing Cooling-Service Research Through a Multidimensional Infrastructure Perspective

Existing studies have substantially advanced the understanding of how urban vegetation, landscape configuration, and green infrastructure influence thermal environments. Landscape ecology has demonstrated that spatial composition and configuration shape ecological processes, including urban thermal regulation [36,37]. However, previous cooling studies have often focused on individual landscape components, such as vegetation coverage or park availability, or relied primarily on surface temperature as an indicator of cooling effectiveness [7,17]. These approaches provide valuable evidence of cooling mechanisms but offer limited understanding of whether cooling-supportive infrastructure is translated into human-relevant thermal benefits and whether such benefits are spatially aligned with heat-sensitive populations.

This study extends previous research by integrating landscape ecological theory with an infrastructure-oriented cooling-service perspective. The blue–green–gray infrastructure framing translates the pattern–process relationship into measurable planning components, while the sequential construction of ICSI, HTPI, and HCDI distinguishes infrastructure cooling supply, realized thermal regulation, human-relevant thermal pressure, and heat-sensitive demand. This layered approach advances cooling-service assessment beyond identifying where green spaces exist toward evaluating whether cooling capacity is associated with lower thermal pressure and whether cooling-supportive infrastructure is spatially aligned with populations experiencing greater heat-sensitive demand.

Empirically, this study provides tract-level evidence that cooling supply is not simply determined by urban–rural differences but reflects continuous spatial heterogeneity across landscape contexts. The ICSI results demonstrate substantial variation among census tracts, suggesting that cooling infrastructure assessment requires finer spatial units rather than broad administrative classifications. Moreover, the SHAP analysis identifies vegetation coverage as the dominant cooling-supportive component, while water-area proportion and road-related gray infrastructure show additional predictive relationships with human-relevant thermal pressure. These findings suggest that the infrastructure-related portion of thermal pressure variation reflects multiple blue–green–gray components rather than green-space quantity alone [8,17]. However, the SHAP rankings indicate the relative importance of infrastructure variables within the predictive model; they should not be interpreted as evidence that these components explain a large share of the total statewide variation in human-relevant thermal pressure.

5.2. Interpreting Spatial Patterns of Cooling Supply and Human-Relevant Thermal Pressure

The empirical results reveal substantial differences in the explanatory importance of cooling infrastructure across thermal outcomes. The nested-model analysis shows that geographic controls alone explain most of the tract-level variation in Heat Index and HTPI, with baseline (R²) values of 0.899 and 0.799, respectively. Adding ICSI produces statistically significant negative associations but only modest incremental explanatory contributions, with partial (R²) values of 0.007 for Heat Index and 0.005 for HTPI. Statistical significance in these models should therefore not be interpreted as equivalent to substantial practical or explanatory importance. Instead, ICSI appears to operate as a secondary local correlate within a thermal structure dominated by broader geographic and climatic gradients.

A different pattern is observed for LSTmax. Although the overall explanatory power of the LSTmax model is lower, the independent contribution of ICSI is comparatively more substantial: its partial (R²) value is 0.022, accounting for approximately 38.6% of the model’s explained variance. This contrast indicates that blue–green–gray infrastructure composition is more directly related to tract-level surface thermal regulation than to statewide variation in Heat Index or integrated human-relevant thermal pressure. The lower overall explanatory power of LSTmax likely reflects its sensitivity to fine-scale surface processes and transient meteorological conditions not fully represented by tract-level indicators. At the same time, this sensitivity makes LSTmax more responsive to local land-cover and infrastructure conditions, including vegetation, canopy, water, impervious surfaces, and roads.

By comparison, Heat Index is jointly determined by ambient temperature and humidity, both of which are strongly structured by latitude, coastal influence, elevation, and regional atmospheric conditions. HTPI further incorporates recurrent extreme heat exposure and therefore retains much of this broader geographic and climatic structure. The high baseline (R²) values for these outcomes are consequently not evidence of stronger infrastructure effects; rather, they show that geography already accounts for most of their spatial variation before ICSI is introduced. This interpretation is consistent with previous findings that human thermal experience is co-determined by humidity, radiation, shade availability, background climate, and adaptive condition [40].

By distinguishing LST-based surface conditions from human-relevant thermal pressure, this study shows that evidence of surface cooling should not automatically be generalized to broader human-relevant thermal conditions. Landscape infrastructure may meaningfully modify local surface temperatures, while its capacity to alter Heat Index or recurrent heat exposure at the state scale is comparatively constrained. Although integrating HTPI into HCDI follows the ecosystem service cascade logic by connecting realized thermal pressure with heat-sensitive demand, this design may introduce partial conceptual overlap because HTPI is also associated with ICSI. The robustness analysis excluding HTPI demonstrates that the identified priority mismatch areas are largely stable, suggesting that the results are not solely an artifact of index construction.

The SHAP results provide additional evidence that different infrastructure components have unequal relative predictive importance within the infrastructure-related component of the HTPI model. Vegetation coverage shows the relative association, consistent with previous findings regarding tree canopies, evapotranspiration, and shading effects [8,38]. Meanwhile, water-area proportion also shows a relationship with predicted thermal pressure, although its effect depends on spatial context and surrounding landscape conditions. In contrast, road-related gray infrastructure is associated with higher predicted thermal pressure, potentially reflecting heat storage, anthropogenic heat, and landscape fragmentation. These component-level rankings complement, but do not override, the nested-model results: they identify which infrastructure variables matter most relative to one another, not how much of the total variation in HTPI is explained by infrastructure as a whole. These findings suggest that cooling interventions should consider the composition and interaction of multiple infrastructure elements rather than treating green-space expansion as a universal solution.

5.3. Policy Recommendations for Cooling Supply–Demand Mismatch

The mismatch analysis provides direct implications for targeted cooling-infrastructure planning. The HCDI developed in this study does not represent observed behavioral demand, such as park visitation, cooling-center use, or self-reported heat-relief needs. Instead, it identifies planning-oriented cooling demand generated by the overlap between human-relevant thermal pressure and vulnerability-related constraints. Therefore, low-supply–high-demand tracts should be interpreted as locations where insufficient cooling-supportive infrastructure coincides with elevated heat-sensitive demand. They should not be interpreted as areas in which infrastructure intervention alone would necessarily produce large reductions in Heat Index or HTPI.

The spatial distribution of mismatch areas highlights the importance of context-specific adaptation strategies. In densely developed areas such as the Baltimore–Washington corridor, cooling deficits overlap with high population concentration and intensive gray infrastructure. These areas may benefit from locally targeted measures such as tree-canopy expansion, neighborhood greening, and impervious-surface retrofitting. The empirical evidence provides the strongest support for interpreting these interventions as measures for improving surface thermal conditions. Their effects on broader human-relevant thermal pressure are likely to be more modest and should therefore be complemented by measures addressing indoor exposure, household cooling access, and population vulnerability.

In contrast, low-supply–high-demand tracts in rural or coastal areas represent localized vulnerability pockets, where older populations, social isolation, limited household cooling access, or socioeconomic constraints may increase heat sensitivity despite lower population density. In these locations, landscape-based cooling interventions may remain useful, but they should be integrated with cooling-center accessibility, household air-conditioning support, public-health outreach, transportation assistance, and heat-warning services.

More broadly, the mismatch map should be understood as a spatial prioritization tool rather than as a prediction of intervention effectiveness. It identifies where infrastructure deficits and heat-sensitive demand converge, but the appropriate response may differ according to the dominant source of risk in each tract. Where elevated surface temperature and imperviousness are central concerns, landscape retrofit may be prioritized. Where regional heat, humidity, social isolation, or inadequate household cooling dominate, infrastructure investment should form part of a wider heat-adaptation portfolio rather than serve as a stand-alone response.

5.4. Limitations and Future Research

Several limitations should be acknowledged. First, although fixed-effects regression helps control for time-invariant tract characteristics, the observational design does not establish causal effects of specific cooling interventions. Moreover, the statistically significant associations between ICSI and Heat Index or HTPI have limited incremental explanatory power after geographic controls are included. These relationships should therefore be interpreted as modest associations rather than evidence that infrastructure cooling supply is a major determinant of statewide human-relevant thermal pressure. Future research could incorporate longitudinal intervention data or quasi-experimental approaches to better evaluate the effectiveness of blue–green–gray infrastructure investments.

Second, HCDI represents a planning-oriented estimate of heat-sensitive cooling demand rather than direct behavioral demand. Future studies could integrate mobility data, cooling-center utilization, household surveys, or health outcomes to better capture how residents experience and respond to extreme heat.

Third, this study uses census tracts as the primary spatial unit. Although this scale enables integration of demographic and infrastructure datasets, future research could examine finer neighborhood scales and temporal dynamics to understand how cooling-service mismatch changes during extreme heat events.

Fourth, SHAP values quantify relative predictive contributions within the fitted machine learning model and do not establish causal effects or the absolute share of statewide thermal pressure variation attributable to each infrastructure component. Future studies should combine explainable machine learning with intervention-based, longitudinal, or multiscale analyses to distinguish relative predictive importance from practical cooling effectiveness.

6. Conclusions

This study developed a cooling-service cascade framework to examine how blue–green–gray infrastructure composition, grounded in landscape ecological pattern–process logic, shapes infrastructure cooling supply, realized thermal regulation, human-relevant thermal pressure, and supply–demand mismatch across 1385 census tracts in Maryland. The results show that infrastructure cooling supply is spatially uneven rather than uniformly distributed across urban, suburban, coastal, and rural contexts. The constructed Infrastructure Cooling Supply Index reveals a clear tract-level gradient, with stronger cooling supply generally associated with higher vegetation coverage, tree canopy, park provision, water-area proportion, and weaker gray-infrastructure pressure. This pattern confirms that infrastructure cooling capacity cannot be reduced to the presence of green space alone; it is produced by the combined structure of blue, green, and gray infrastructure components. This finding shifts the analytical focus from green-space quantity to infrastructure composition and from static land-cover description to planning-relevant cooling supply.

The regression results further demonstrate that higher infrastructure cooling supply is associated with lower thermal pressure, although this relationship varies across thermal indicators. The association is strongest for surface thermal conditions, remains significant for apparent heat pressure, and is also evident for the integrated Human-Relevant Thermal Pressure Index. This finding is important because it suggests that infrastructure cooling supply is not only linked to land surface temperature reduction but also to heat conditions more directly relevant to human exposure. At the same time, the weak correspondence between surface temperature and human-relevant thermal indicators indicates that surface cooling should not be treated as a complete proxy for heat adaptation performance. In other words, surface thermal regulation and human-relevant heat-pressure mitigation should be evaluated as related but distinct dimensions of urban cooling performance.

The explainable GeoAI analysis provides a more detailed interpretation of the infrastructure components underlying this relationship. Both XGBoost feature importance and SHAP-based interpretation consistently identify overall vegetation coverage as the most influential landscape component associated with human-relevant thermal pressure. Water-area proportion and road-related gray infrastructure also play important roles, while formal park provision shows weaker direct contributions. The SHAP results further reveal that high vegetation coverage and water-area proportion tend to reduce predicted thermal pressure, whereas road infrastructure tends to increase or nonlinearly shape it. These findings suggest that effective cooling-service planning should move beyond park quantity or green-space designation and pay closer attention to vegetation structure, blue-space presence, and the thermal burden created by gray infrastructure. This result is especially important for planning practice because it shows that designated green space, vegetation-based cooling function, blue-space contribution, and gray-infrastructure pressure are not interchangeable.

The supply–demand mismatch analysis translates these findings into a planning-relevant spatial diagnosis. A total of 433 census tracts, accounting for 31.3% of all Maryland census tracts, were classified as low supply–high demand priority mismatch areas. These tracts include both densely populated areas in the Baltimore–Washington corridor and more dispersed rural or coastal vulnerability pockets. This distinction shows that cooling mismatch has two meanings: in dense urban corridors, it indicates places where interventions may benefit larger exposed populations; in low-density areas, it identifies localized vulnerability where small populations may still face substantial heat-related constraints. By connecting infrastructure cooling supply, realized thermal regulation, human-relevant pressure, and heat-sensitive demand, this study provides a tract-level diagnostic framework for identifying where cooling-infrastructure interventions are most needed. More broadly, the findings suggest that urban heat adaptation should move from generalized greening strategies toward spatially targeted infrastructure planning that jointly considers cooling supply, thermal pressure, social vulnerability, and implementation priority.