Integrating Multi-Source Remote Sensing and Spatial Metrics to Quantify Urban Park Design Effects on Surface Cool Islands in Mexicali, Mexico

Highlights

What are the main findings?

• Urban parks in Mexicali produced measurable daytime cooling in both intensity and extent (avg. ΔTmax = 0.81 °C; Lmax = 120.15 m), whereas nighttime effects were weaker and more variable (avg. ΔTmax = 0.43 °C; Lmax = 47.85 m).
• A 10% increase in vegetation raised SCI intensity by up to +0.11 °C and extended cooling reach by +3–6 m, whereas greater fragmentation (FD + 0.1) reduced Lmax by up to −34 m; null SCI occurred in 35% of parks, mainly those with <30% vegetation.

What is the implication of the main finding?

• The limited and variable nighttime SCI highlights the importance of vegetation type, irrigation, and surface thermal inertia in arid cooling strategies.
• Identifying structural thresholds for SCI failure enables planners to remotely detect underperforming parks and prioritize vegetation-based interventions.

Abstract

The Surface Cool Island (SCI) refers to localized reductions in land surface temperature (LST) produced by features that enhance evapotranspiration, shading, and energy flux regulation. In arid urban areas, vegetated parks play a key role in mitigating heat through these mechanisms. This study evaluates how park vegetation structure and spatial configuration influence SCI intensity (ΔTmax) and extent (Lmax) using multi-seasonal, day–night satellite observations in Mexicali, Mexico. A total of 435 parks were analyzed using Landsat 8/9 TIRS (30 m) for LST and Sentinel-2 MSI (10 m) for vegetation mapping via NDVI thresholding and supervised random forest (RF) classification. On average, parks lowered daytime LST by 0.81 °C (max: 6.41 °C), with a mean Lmax of 120 m; nighttime cooling was weaker (avg. ΔTmax: 0.37 °C; Lmax: 48 m). RF-derived metrics explained SCI variability more effectively (R² up to 0.64 for ΔTmax; 0.48 for Lmax) than NDVI-based metrics (R² < 0.35), highlighting the value of object-based land cover classification in capturing vegetation structure. This remote sensing framework offers a scalable method for assessing urban cooling performance and supports climate-adaptive green space design in hot-arid cities.

1. Introduction

Desert cities face unique thermal and climatic challenges due to their extreme aridity, scarce vegetation, and expansive impervious surfaces, which collectively intensify urban heat stress. These environmental constraints heighten surface energy retention and reduce evapotranspiration, making desert urban areas particularly vulnerable to the Urban Heat Island (UHI) effect, defined as the temperature difference between urban and non-urban areas resulting from land cover modification and anthropogenic heat emissions [1,2]. Over the past two decades, advances in satellite-based thermal remote sensing have transformed UHI research by enabling the spatially continuous monitoring of Land Surface Temperature (LST) and the delineation of the Surface Urban Heat Island (SUHI), a surface-level representation of UHI that reveals spatial thermal disparities across urban landscapes [3,4,5].

While SUHI studies have significantly advanced our understanding of large-scale urban thermal behavior, recent research emphasizes the importance of examining intra-urban thermal heterogeneity, especially in relation to localized mitigation features such as parks, green corridors, and water bodies [6,7]. In this context, urban green infrastructure has emerged as a key strategy for microclimate regulation, offering localized cooling benefits that broader SUHI analyses often overlook.

Among green infrastructure elements, urban parks play a particularly critical role in microclimate regulation. They generate localized cooling effects known as Park Cool Islands (PCIs), which arise from biophysical processes including evapotranspiration, shading, and increased surface albedo. These mechanisms collectively reduce temperatures relative to surrounding impervious surfaces, mitigating UHI intensity and enhancing outdoor thermal comfort [8,9]. Within this framework, the term Surface Cool Island (SCI) refers more specifically to the measurable reduction in LST produced by urban surfaces with favorable physical or biophysical characteristics, such as vegetation cover, permeable substrates, or water bodies, which enhance evaporative cooling and modulate surface energy fluxes.

SCI effects are typically assessed using thermal infrared remote sensing, which enables the quantification of two key indicators: cooling intensity (ΔTmax), defined as the maximum temperature reduction within the park compared to its surrounding urban environment, and spatial extent (Lmax), defined as the maximum distance beyond the park boundary where the cooling influence remains detectable. These indicators are derived through thermal gradient profiling, which captures how surface temperatures transition from green to built-up areas [10,11].

Multiple studies have quantified SCI effects, consistently reporting LST reductions ranging from 1 °C to over 5 °C, depending on climatic conditions, park design, and vegetation characteristics. For example, Zhou et al. [11] observed cooling effects of up to 4.5 °C in Beijing during summer, while Chakraborty and Lee [10] reported SCI magnitudes reaching 5 °C in large, densely vegetated parks. In arid and semi-arid contexts, He and Yang [12] found SCI intensities ranging from 2 °C to 6 °C, heavily influenced by vegetative structure and spatial configuration. Similarly, Middel et al. [13] documented cooling effects of 1 °C to 4 °C in Phoenix, Arizona, underscoring the influence of vegetation type and land use on thermal performance.

Regarding spatial propagation, SCI effects typically extend between 100 m and 300 m into the surrounding urban matrix, though some studies report effects reaching up to 500 m under favorable conditions [6,11,14]. This distance is influenced by factors such as vegetation density, park shape, surrounding land cover, and local meteorological conditions. Compared to air temperature (Ta) measurements, SCI effects based on LST tend to show greater intensity but shorter persistence, as surface temperatures are more sensitive to solar radiation and surface material properties [10,13]. In contrast, air-based PCI effects are usually smaller in magnitude, typically between 0.5 °C and 2.5 °C, but can extend farther spatially and persist longer into the evening due to thermal inertia and atmospheric mixing [15,16].

Recent studies have increasingly focused on identifying the physical and structural characteristics that shape SCI indicators, with the goal of defining optimal park design parameters based on their urban context to maximize cooling performance [7,12,14]. These efforts have been significantly enhanced by the growing availability of moderate- to high-resolution, multi-sensor satellite platforms, which now enable detailed, multi-temporal analyses of urban thermal behavior and its relationship with vegetation structure and land cover composition [17,18]. Despite this progress, vegetation indices such as the Normalized Difference Vegetation Index (NDVI) remain the most widely used in remote sensing due to their simplicity, global applicability, and sensitivity to photosynthetic activity. NDVI provides a practical means of characterizing urban green infrastructure at large scales [19]. However, in spectrally complex environments, particularly arid cities with sparse and heterogeneous vegetation, NDVI often fails to capture critical structural attributes such as canopy layering, height, and vertical continuity, all of which are key to understanding cooling performance [10,20]. Moreover, overlapping NDVI values from impervious surfaces, dry soils, and stressed vegetation may reduce classification accuracy [21]. As a pixel-based index, NDVI also struggles to resolve detailed spatial configurations of vegetation in complex urban contexts.

To address these limitations, object-based image analysis (OBIA) has gained traction as a more robust alternative. By integrating spectral, spatial, geometric, and contextual information derived through image segmentation, OBIA offers a more detailed representation of urban land cover. When combined with supervised machine learning classifiers, such as Random Forest (RF) or Support Vector Machines (SVMs), OBIA enables more accurate land cover differentiation in complex urban environments [22,23]. This methodological framework is particularly well-suited for SCI assessments, where not only vegetation presence but also structural continuity, spatial coherence, and landscape configuration strongly influence cooling outcomes.

Mexicali, Mexico, exemplifies the extreme vulnerability of desert cities to urban heat. As one of the hottest cities in North America, Mexicali experiences summer air temperatures that frequently exceed 45 °C, with prolonged heat waves exacerbating UHI effects [5,24,25,26]. These extreme conditions increase public exposure to thermal stress, elevate cooling energy demands, and emphasize the need for climate-adaptive urban design strategies [3,27]. Despite its relevance, Mexicali has received limited attention in fine-scale SCI studies, particularly those using remote sensing techniques. In such arid environments, SCI dynamics are shaped by environmental constraints including limited water availability, seasonal vegetation changes, and widespread impervious cover, all of which restrict the cooling potential of urban parks [13,28]. Under these conditions, internal park characteristics, such as vegetation density, canopy connectivity, and spatial configuration, become key factors in determining thermal effectiveness [12,29]. Denser and more cohesive vegetation not only increases evapotranspiration but also facilitates the lateral propagation of cooling into adjacent urban areas.

This study contributes to the literature by conducting a large-scale, multi-seasonal assessment of SCI dynamics across 435 urban parks in Mexicali using a combined day–night remote sensing framework. LST was derived from Landsat 8/9 TIRS imagery, while Sentinel-2 MSI data supported land cover mapping. Two approaches, NDVI thresholding and RF classification, were applied to examine their respective capacities to characterize vegetation structure and predict SCI variation. SCI indicators (ΔTmax and Lmax) were computed using concentric buffer-based thermal profiling.

The objectives of this study are as follows: (1) Quantify the SCI intensity and spatial extent of urban parks using multi-season remote sensing data; (2) Identify the relative influence of vegetation structural and compositional attributes on SCI performance in arid conditions; (3) Compare the explanatory power of land cover descriptors derived from NDVI and RF classification in relation to SCI variability; and (4) Develop a replicable, satellite-based methodology to support climate-adaptive green infrastructure design in desert cities.

2. Materials and Methods

This study followed a four-step methodological framework to evaluate the influence of park design and location on the SCI effect in Mexicali, Mexico: (1) Satellite data acquisition; (2) SCI indicator quantification using a multiple-buffer LST profiling approach; (3) Computation of physical and spatial metrics, including park size, shape complexity, land cover composition, vegetation landscape metrics, and locational attributes; and (4) Statistical analysis, including bivariate correlation and multiple linear regression to assess relationships between physical descriptors and SCI indicators.

2.1. Study Area

The city of Mexicali is located in northwestern Mexico, adjacent to the United States border in the state of Baja California [32°39′55″N–32°38′36″N; 115°28′45″W–115°29′52″W] (Figure 1). It serves as the primary urban center of the municipality. As of 2020, the urbanized area covers approximately 22,405.57 ha (224 km²), with a population of 854,186 residents [30,31]. Mexicali exhibits a predominantly low-density, grid-based urban morphology spread across flat desert terrain, situated at an elevation of approximately 10 m above sea level. The surrounding landscape consists primarily of agricultural land and desert, with negligible topographic variation.

Mexicali is one of the hottest cities in North America, characterized by an arid climate and extreme summer temperatures. Historical climate records indicate that daily maximum air temperatures frequently exceed 45 °C during July and August, and heatwaves have become increasingly prevalent in recent decades [26]. Additionally, the city experiences a pronounced UHI effect, particularly at night, with recorded urban–rural air temperature differences reaching up to 5.7 °C in winter [32]. These thermal conditions contribute to elevated energy demands and exacerbate heat-related health risks, making Mexicali a critical case study for urban climate adaptation and mitigation.

A total of 435 urban parks located within the urbanized perimeter were selected for analysis. Parks were identified based on the Parks and Gardens Report published by the Municipal Institute of Research and Planning (IMIP) [33], supplemented with newly developed residential parks identified via high-resolution ESRI Worldview imagery [34]. Park boundaries were delineated using official land use datasets provided by IMIP [35]. The sample includes a range of public spaces, such as gardens, plazas, civic squares, sports complexes, and cultural areas, not limited strictly to vegetated zones (Figure 1).

Together, these parks comprise a total area of 441.55 ha, representing approximately 2% of Mexicali’s urbanized land. To assess their potential microclimatic influence on the surrounding urban environment, particularly regarding cooling effects, a 500 m buffer was applied around each park. This buffer distance is commonly used in urban climate studies to capture the effective spatial range of PCI effects, where green spaces reduce surrounding LST through evapotranspiration and shading [36].

Figure 1. Urban parks of Mexicali: (a) Global location of Mexico; (b) Baja California region within Mexico; (c) Location of Mexicali within the municipality; (d) Spatial distribution of urban parks and 500 m buffer zones in relation to existing water bodies, current urbanized areas, and the designated urban expansion boundary under the 2025 Urban Development Plan [37].

Figure 1. Urban parks of Mexicali: (a) Global location of Mexico; (b) Baja California region within Mexico; (c) Location of Mexicali within the municipality; (d) Spatial distribution of urban parks and 500 m buffer zones in relation to existing water bodies, current urbanized areas, and the designated urban expansion boundary under the 2025 Urban Development Plan [37].

2.2. Remote Sensed Data Retrieval

Seasonal satellite imagery was acquired from Landsat 8/9 OLI/TIRS [38] for daytime and nighttime LST retrieval, and from Sentinel-2 MSI [39] for land cover and vegetation metrics (

Table 1. Summary of satellite data acquisition.

Season	Satellite	Sensor	Preprocessing	Path/Row	Date (dd/mm/aaaa)	Time (UTM)	Period
Spring	Landsat-8	OLI/TIRS	Collection 2 Level 2	039/037	19 April 2021	10:16:01	Day
	Landsat-8	OLI/TIRS	Collection 2 Level 1	137/207	18 April 2021	21:29:16	Night
	Sentinel 2	MSI	Level 2A	11SPS	20 April 2021	10:19:21	Day
Summer	Landsat-8	OLI/TIRS	Collection 2 Level 2	039/037	25 August 2021	10:16:33	Day
	Landsat-8	OLI/TIRS	Collection 2 Level 1	137/207	24 August 2021	21:29:47	Night
	Sentinel 2	MSI	Level 2A	11SPS	18 August 2021	10:19:21	Day
Autumn	Landsat-9	OLI/TIRS	Collection 2 Level 2	039/037	8 November 2021	10:17:38	Day
	Landsat-8	OLI/TIRS	Collection 2 Level 1	138/207	7 November 2021	21:35:46	Night
	Sentinel 2	MSI	Level 2A	11SPS	6 November 2021	10:25:51	Day
Winter	Landsat-9	OLI/TIRS	Collection 2 Level 2	039/037	9 February 2022	10:16:36	Day
	Landsat-8	OLI/TIRS	Collection 2 Level 1	137/207	8 February 2022	21:30:10	Night
	Sentinel 2	MSI	Level 2A	11SPS	4 February 2022	10:25:41	Day

). For each season (spring, summer, autumn, winter), paired Landsat daytime and nighttime acquisitions were selected. Daytime data correspond to Collection 2 Level 2 (surface reflectance corrected), while nighttime LST was derived from Collection 2 Level 1 due to the absence of nighttime surface reflectance correction.

Landsat daytime images were acquired during the morning descending orbit, and nighttime images during the evening ascending orbit. Nighttime scenes corresponded to the evening preceding each daytime acquisition. Sentinel-2 Level 2A imagery was selected to closely match Landsat overpass dates. All images were chosen to exclude post-precipitation periods and were aligned with stable seasonal meteorological conditions to ensure phenological consistency (Figure A1). All satellite image and vector data processing was performed in QGIS 3.40.

2.2.1. NDVI Calculation

The NDVI was derived using atmospherically corrected surface reflectance data from Landsat 8/9 and Sentinel-2 Level 2 products [39,40]. Full formula details and threshold values are presented in Appendix A (Equation (A1)). The derived NDVI layers were then used to characterize vegetation conditions within park interiors, cooling footprint areas, and surrounding urban buffers.

2.2.2. LST Retrieval

LST was retrieved from Landsat 8/9 Thermal Infrared Sensor (TIRS) imagery using two complementary approaches based on the level of preprocessing. For daytime data, surface temperature values were directly obtained from Landsat Collection 2 Level 2 products, which include a preprocessed Surface Temperature (ST) band. These data are radiometrically calibrated, atmospherically corrected, and include surface emissivity adjustments [40]. The final LST values in Celsius were calculated by applying a scaling factor and conversion equation (Equations (A2) and (A3)).

In contrast, nighttime LST required a physically based retrieval from Landsat Collection 1 Level 1 data due to the absence of standard Level 2 products [41]. The retrieval followed a single-channel radiative transfer algorithm calibrated for Band 10, incorporating both atmospheric correction and surface emissivity adjustment [42,43]. The calculation involved deriving at-sensor brightness temperature and applying the emissivity-corrected LST formula (Equation (A4)).

Spectral emissivity (ε) was estimated using the NDVI threshold method [44,45], which classifies pixels into bare soil, mixed, or fully vegetated surfaces based on NDVI values [45,46]. This method uses NDVI-derived vegetation proportion (P_V) to assign pixel-level emissivity (Equations (A5) and (A6)). NDVI values used in nighttime LST retrieval were extracted from the nearest daytime image to maintain consistency under stable meteorological conditions [47].

As a result, eight seasonal LST maps (in °C) were generated at 30 m spatial resolution, representing both daytime and nighttime surface temperatures for spring, summer, autumn, and winter (Figure 2).

2.3. Physical Variables Calculation

The influence of park structural characteristics and positioning variables on the SCI was evaluated through physical descriptors tailored to the regional context and data availability for Mexicali. The variables were organized into four categories: (1) shape descriptors, (2) remote sensing indices, (3) land cover composition indicators based on spectral and object-based classifications, and (4) geospatial positioning variables (

Table 2. Physical variables evaluated in relation to the SCI in Mexicali.

Variable	Units	Data Source	Calculation
1. Shape descriptors 1
1.1. Park area	m2	Polygons of parks	Area of polygons
1.2. Landscape Shape Index (LSI)	LSI ≥ 1	Parks area (A) and perimeter (P)	LSI = P/(2$\sqrt{\pi A})$
1.3. Fractal dimension	FD ≥ 1		FD = (2 × log P)/(log A)
2. Remote sensing indices 2
2.1. Daytime LST	°C	Landsat 8/9 Level 2	Mean LST within park polygon
2.2. Nighttime LST	°C	Landsat 8 Level 1	Mean LST within park polygon
2.3. NDVI	−1 to +1	Sentinel 2. Level 2A	Mean NDVI within park polygon
3. Land cover composition 1
3.1. NDVI range classification	%	Sentinel 2. Level 2A	Proportion and area of each NDVI range
	m2
3.2. Cover diversity—NDVI ranges	H′ > 0		H′ = −∑(pi × log(pi))
3.3. RF classification (RF)	%		Proportion and area of each land cover class
	m2
3.4. Cover diversity—RF classes	H′ > 0		H′ = −∑(pi × log(pi))
4. Green cover structure 3
4.1. NDVI > 0.30 cover	%	Sentinel 2. Level 2A	Proportion and area of NDVI > 0.30 pixels
	m2
4.2. NDVI > 0.30 cover LSI	LSI ≥ 1	NDVI > 0.30 clusters	LSI = P/(2$\sqrt{\pi A})$
4.3. NDVI > 0.30 cover FD	FD ≥ 1		FD = (2 × log P)/(log A)
4.4. RF-derived vegetation cover	%	Sentinel 2. Level 2A	Proportion and area of RF-classified vegetation and grass
	m2
4.5. RF vegetation LSI	LSI ≥ 1	RF vegetation polygons	LSI = P/(2$\sqrt{\pi A})$
4.6. RF vegetation FD	FD ≥ 1		FD = (2 × log P)/(log A)
5. Positioning variables 4
5.1. Latitude	m	Park centroid coordinates	Latitude of centroid
5.2. Longitude	m		Longitude of centroid
5.3. Elevation	masl	Digital Elevation Model	Mean elevation within park polygon

¹ Detailed in Section 2.3.1; ² Section 2.2; ³ Section 2.3.2; ⁴ Section 2.3.3.

).

For the analysis of the influence of physical descriptors on the SCI effect, the study area was partitioned into three spatial zones: (1) the park area, (2) the area within the identified SCI spatial extent (Lmax), and (3) the surrounding buffer area beyond the SCI Lmax, extending to a maximum distance of 500 m from the park boundary. For the computation of physical descriptors, only the raster cells whose centroids were located within the corresponding polygon boundaries were included.

2.3.1. Land Cover Classification Methods

To extract detailed physical descriptors of the park structure and the surrounding urban fabric, land cover classification was performed on Sentinel-2 imagery using two complementary approaches: NDVI thresholding and object-based supervised classification with a RF algorithm.

The NDVI-based classification provided a simplified assessment of vegetation quality. The study area was stratified into seven NDVI categories (Figure 3) [7]: Non-vegetated surfaces (NDVI < 0.10); very low vegetation (NDVI 0.10–0.20); low vegetation (NDVI 0.20–0.30); moderate vegetation (NDVI 0.30–0.40); medium-high vegetation (NDVI 0.40–0.50); high vegetation (NDVI 0.50–0.60); and dense vegetation (NDVI > 0.60). For each analysis unit, both the proportional coverage (%) and total area (m²) of each NDVI class were computed.

Figure 3. Visualization of NDVI ranges at 0.10 intervals for spring (a), summer (b), autumn (c), and winter (d) [39].

Figure 3. Visualization of NDVI ranges at 0.10 intervals for spring (a), summer (b), autumn (c), and winter (d) [39].

The RF classification was applied using all available bands from Sentinel-2 multispectral imagery to capture land cover heterogeneity with enhanced thematic detail (Figure 4). Eleven land cover classes were defined: (1) Agricultural land, (2) Vegetation (trees and shrubs), (3) Grass (lawns and low vegetation), (4) Rustic sandy soil, (5) Rocky bare soil, (6) Urban sandy surfaces, (7) Water bodies, (8) Concrete, (9) Asphalt, (10) Buildings, and (11) Industrial areas. The RF model was trained using 400 manually labeled polygons and configured with 500 decision trees (n_estimators = 500), full tree depth (max_depth = None), and square root feature selection at each node (max_feature = “sqrt”), using Gini impurity as the splitting criterion and bootstrap sampling enabled. A stratified hold-out validation approach was applied, in which the dataset was split into training and testing subsets while preserving the proportional representation of all cases. Finally, the model performance was evaluated using a confusion matrix, from which per-class recall values and Cohen’s kappa coefficients were calculated (Figure A3 and Figure A4).

Classification performance exhibited clear seasonal variation, with overall recall ranging from 71.0% in summer to 79.2% in autumn, and Cohen’s kappa values ranging from 0.610 to 0.723, indicating moderate to substantial agreement (

Table A1. Seasonal per-class recall (%) and overall classification performance (accuracy and Cohen’s kappa) for land cover classification using the RF algorithm.

Class	Spring	Summer	Autumn	Winter
Agricultural	20.9%	15.6%	44.8%	63.2%
Vegetation	54.4%	66.9%	85.5%	69.3%
Grass	89.6%	89.1%	77.4%	85.4%
Rustic sandy soil	91.1%	85.2%	90.9%	82.9%
Rocky bare soil	96.6%	97.1%	91.4%	83.3%
Urban sandy area	61.7%	64.6%	77.6%	82.7%
Water	92.1%	95.5%	89.6%	85.1%
Concrete	79.2%	84.6%	79.3%	78.6%
Industrial	56.7%	64.1%	60.7%	65.1%
Asphalt	78.5%	70.1%	73.0%	82.3%
Buildings	84.8%	69.4%	85.2%	78.5%
Overall Recall	74.0%	71.0%	79.2%	78.2%
Cohen’s Kappa	0.653	0.610	0.723	0.712

). Certain classes, such as Water, Rustic sandy soil, and Rocky bare soil, maintained high recall across all seasons (≥82.9%) due to their distinct and stable spectral responses. In contrast, Agricultural land and Vegetation classes demonstrated higher seasonal sensitivity, with Agricultural recall increasing from 15.6% in summer to 63.2% in winter, and Vegetation recall peaking at 85.5% in autumn. These fluctuations correspond to known phenological cycles, such as vegetation senescence, regrowth, and harvest, which alter spectral characteristics over time and influence class separability. Similar effects of phenology on classification accuracy have been widely reported in the literature [48,49].

2.3.2. Landscape Metrics

Three landscape metrics were calculated to characterize the spatial structure of vegetated areas and compositional heterogeneity of the land cover types: (1) Landscape Shape Index (LSI), (2) Fractal Dimension (FD), and (3) Shannon Diversity Index (H′).

LSI and FD characterize shape complexity based on the relationship between perimeter (P) and area (A). LSI [LSI = P/(2$\sqrt{\pi A})$] measures deviation from a circular shape (LSI = 1 indicates a perfect circle and higher values indicate increasing shape irregularity), while FD [FD = (2 × log P)/(log A)] provides a scale-invariant measure of perimeter roughness [50]. These metrics were applied to vegetated areas identified through NDVI thresholding (NDVI > 0.30) and the RF-derived Agricultural, Vegetation, and Grass classes.

The compositional heterogeneity of land cover was assessed using the H′ index, calculated as H′ = −∑(pi × log pi), where pi represents the proportional area of each land cover type [51]. This index was computed from NDVI ranges and RF-derived land cover classes.

2.3.3. Location Variables

Location variables were incorporated to characterize the spatial positioning and topographic attributes of each park. Geographic coordinates (latitude and longitude) were used to define park centroids within the urban area, allowing for spatial analyses of potential locational effects on microclimatic variability. Elevation was extracted from a high-resolution Digital Elevation Model (DEM) derived from Lidar data, representing the mean altitude above sea level (masl) for each park [52].

2.4. Analytical Methods

2.4.1. SCI Calculation

The SCI was quantified for each park using two indicators: cooling intensity (ΔTmax), representing the maximum reduction in LST within the park relative to its surrounding urban context, and spatial extent (Lmax), defined as the distance over which this cooling influence extends into adjacent urban areas. These metrics were derived from standardized radial LST profiles, following established SCI quantification methods based on concentric buffer zones [14,53].

To generate these profiles, mean LST values were extracted from each park interior and 50 concentric buffer zones at 10 m intervals, extending 500 m from the park boundary (Figure 5a–c). Landsat 8/9 LST data (30 m) were resampled to 10 m using bilinear interpolation to improve spatial continuity, yielding 51 sequential mean LST values per profile, including the park core and each successive buffer zone (Figure 5c–e). To reduce short-range variability and enhance the detection of thermal gradients, a sixth-order polynomial curve was fitted to each LST profile (Figure 5d). The curve fitting quality was evaluated using the coefficient of determination (R²). In most cases, the polynomial provided a strong fit (R² ≥ 0.90) during daytime periods (

Table A2. Seasonal distribution of curve fitting accuracy (R²) for daytime and nighttime LST retrievals: (a) Absolute frequency; and (b) Relative percentage.

			(a)
	Spring		Summer		Autumn		Winter
	Day	Night	Day	Night	Day	Night	Day	Night
<0.30	0	0	0	0	0	0	0	0
0.30–0.40	0	1	0	0	0	1	0	1
0.40–0.50	0	1	0	2	1	4	0	0
0.50–0.60	0	1	0	2	0	9	0	11
0.60–0.70	0	2	0	12	2	25	3	24
0.70–0.80	4	19	2	27	0	39	2	35
0.80–0.90	11	52	9	105	11	99	15	76
0.90–1.00	420	359	424	287	421	258	415	288
				(b)
	Spring		Summer		Autumn		Winter
	Day	Night	Day	Night	Day	Night	Day	Night
<0.30	0.0%	0.0%	0.0%	0.0%	0.0%	0.0%	0.0%	0.0%
0.30–0.40	0.0%	0.2%	0.0%	0.0%	0.0%	0.2%	0.0%	0.2%
0.40–0.50	0.0%	0.2%	0.0%	0.5%	0.2%	0.9%	0.0%	0.0%
0.50–0.60	0.0%	0.2%	0.0%	0.5%	0.0%	2.1%	0.0%	2.5%
0.60–0.70	0.0%	0.5%	0.0%	2.8%	0.5%	5.7%	0.7%	5.5%
0.70–0.80	0.9%	4.4%	0.5%	6.2%	0.0%	9.0%	0.5%	8.0%
0.80–0.90	2.5%	12.0%	2.1%	24.1%	2.5%	22.8%	3.4%	17.5%
0.90–1.00	96.6%	82.5%	97.5%	66.0%	96.8%	59.3%	95.4%	66.2%

) and moderate (R² ≥ 0.60) during nighttime.

Four typical SCI response types emerged: (Figure 6a) Null SCI—no observable cooling effect, with the park exhibiting higher LST than its surroundings; (Figure 6b) Stable cooling gradient—a distinct cooling curve followed by clear LST stabilization; (Figure 6c) Diffuse cooling—a pronounced cooling curve with a gradual but non-stabilized LST trend; and (Figure 6d) Bimodal response—an initial cooling curve followed by a secondary LST drop at greater distances, likely influenced by nearby land uses.

From these profiles, Lmax was identified as the distance where cooling stabilized, determined using three criteria: (1) Null SCI, assigned when the park’s mean LST was higher than the mean LST of at least three consecutive adjacent buffer zones (within the first 40 m), indicating no observable cooling effect; (2) the inflection point, selected when the cooling curve stabilized, began to rise, or showed a clear shift in slope, marking the end of a measurable cooling gradient; and (3) stable trend without a clear curve, applied when no distinct inflection was observed but the profile showed a consistent temperature change of less than 0.1 °C across three or more consecutive buffers or a clear flattening of the gradient within the first 200–300 m, suggesting a diffuse but valid SCI extent.

Profiles not fitting these criteria (e.g., oscillatory profiles, null Lmax despite visible cooling) were revised and corrected through a standardized supervised review, prioritizing spatial coherence and thermal stability over rigid rule hierarchy. This hybrid approach, automated detection with expert oversight, ensured consistent SCI identification while addressing the limitations of curve fitting in noisy profiles. Manual corrections were required in only 2.7% of all Lmax cases (93 of 3480 profiles across all parks, seasons, and times of day), demonstrating the robustness of the protocol (

Table A3. Number and percentage of supervised adjustments of Lmax values in the LST profiles by season and time of day (n = 435).

Season	Time	Adjusted Lmax	% of Total (n = 435)
Spring	Day	29	6.7%
	Night	5	1.1%
Summer	Day	9	2.1%
	Night	4	0.9%
Autumn	Day	19	4.4%
	Night	1	0.2%
Winter	Day	25	5.7%
	Night	1	0.2%

).

Finally, SCI ΔTmax was calculated as the difference between the park’s mean LST and the LST at the identified Lmax distance. All calculations were conducted independently for daytime and nighttime datasets across the four seasons.

2.4.2. Statistics Analysis

Statistical analysis was conducted to evaluate the relationship between physical park characteristics and SCI indicators (ΔTmax and Lmax) across daytime and nighttime conditions for each season. To ensure consistency and reduce size-related bias, parks smaller than 5000 m² were excluded, following World Health Organization (WHO) guidelines recommending access to at least 0.5 ha of green space within 300 m of residences [54]. Additionally, parks exceeding three standard deviations above the mean size were removed as outliers. These extremely large parks, while few, can disproportionately influence correlation and regression outcomes due to their anomalous scale, spatial structure, and distinct thermal dynamics, which are often not comparable to smaller neighborhood parks. By excluding these outliers, the analysis focuses on the more representative distribution of urban parks and ensures that statistical relationships are not driven by scale extremes but rather reflect broader patterns applicable to the greenest spaces in the study area.

The final sample consisted of 148 parks (Figure A2). A two-step analytical approach was applied. First, bivariate correlation analyses were performed to identify the strength and direction of individual relationships between each physical descriptor and the SCI indicators. This preliminary step helped highlight key variables with potential explanatory value. Next, multiple linear regression models were constructed separately for ΔTmax and Lmax under both daytime and nighttime conditions. Model sets were developed for both NDVI-derived and RF-based land cover classifications. In each case, only vegetated land cover types were considered, along with their structural metrics (e.g., area, LSI, FD). Additionally, land cover complexity was incorporated using the H’ index to capture the heterogeneity and spatial fragmentation of vegetated surfaces. All regression models were refined using a stepwise selection method to retain the most significant predictors, resulting in four optimized model sets for each of the eight periods assessed.

3. Results

3.1. SCI Effect of Urban Parks in Mexicali

3.1.1. Seasonal Variation of LST and SCI

The initial results, based on the full set of 435 urban parks within the urban area of Mexicali, confirm the consistent presence of the SCI effect throughout the year. By comparing the mean LST values between parks and their surrounding urban context, a clear and recurring cooling gradient is observed across all four seasons, particularly during the day (

Table 3. Seasonal mean LST during daytime and nighttime for the full set of 435 urban parks in Mexicali, their 500 m buffer zones, the urbanized area, the Plan 2025 boundary, and the surrounding water bodies.

Mean LST (°C)	Spring		Summer		Autumn		Winter
	Day	Night	Day	Night	Day	Night	Day	Night
Parks	46.46	22.56	55.52	31.99	32.12	20.27	28.81	4.43
500 m area	48.19	22.29	57.41	32.21	33.23	20.21	30.13	3.84
Urbanization	48.45	22.25	57.52	32.18	33.38	20.17	30.26	3.79
Plan 2025 limit	47.85	21.25	57.48	31.61	33.70	19.39	30.57	2.72
Water bodies	34.98	23.54	43.81	31.81	27.26	22.72	21.63	7.27

). Parks consistently exhibit lower LST values than both their immediate 500 m buffer zones and the broader urban area.

This gradient is evident in the progressive temperature increase moving outward from park interiors (Figure 7). In spring, the average daytime LST in parks is 46.46 °C, increasing to 48.19 °C at 500 m (+1.73 °C) and reaching 48.45 °C across the urban area (+1.99 °C relative to parks). In summer, the difference is more pronounced: parks average 55.52 °C, compared to 57.41 °C at 500 m and 57.52 °C across the city (+2.00 °C). In autumn, parks average 32.12 °C, rising to 33.23 °C in the buffer zone and 33.38 °C citywide. Winter shows the largest nighttime SCI effect, with parks averaging 4.43 °C, compared to 3.84 °C at 500 m and 3.79 °C in the wider urban area.

Nighttime SCI behavior reveals a more nuanced thermal dynamic across seasons (Figure 7). In some cases, parks exhibit slightly higher nighttime temperatures than their immediate surroundings, particularly the 500 m buffer zone and the broader urban area. In spring, the average nighttime LST in parks is 22.56 °C, which is 0.27 °C higher than the 500 m buffer (22.29 °C) and 0.31 °C higher than the urban area (22.25 °C). This counterintuitive nighttime warming arises from seasonal park energy dynamics. In spring, incomplete canopies provide limited shading, allowing soils to absorb more heat during the day. Moist soils and young vegetation with high thermal inertia then release stored heat at night, while evapotranspiration, the primary cooling process, ceases after sunset [55]. Built-up areas, by contrast, cool more quickly through ventilation and lower heat storage, leaving parks relatively warmer despite their daytime cooling [56].

In summer, parks display a modest nighttime cooling effect, averaging 31.99 °C, 0.22 °C lower than the buffer (32.21 °C), and 0.19 °C below the urban area (32.18 °C). During autumn, nighttime LST in parks is 20.27 °C, slightly higher than the buffer (20.21 °C) and the urban area (20.17 °C). As previously noted, winter shows the strongest nighttime SCI effect, with parks averaging 4.43 °C compared to 3.84 °C and 3.79 °C in the buffer and urban areas, respectively.

In summary, while parks consistently maintain lower daytime LST values, the intensity of the nighttime SCI effect varies seasonally. At times, parks may be slightly warmer than surrounding areas after sunset due to the vegetation phenology and soil thermal properties. These findings underscore the complexity of urban thermal dynamics, particularly during nighttime, but also reaffirm the year-round contribution of parks to localized cooling in arid urban settings.

3.1.2. SCI Extent

The SCI Lmax of the parks of Mexicali indicates that the cooling effect of parks extends further during the day compared to the night, and its extent is generally higher during spring and autumn compared to summer and winter (

Table 4. (a) Descriptive statistics of the SCI Lmax of the urban parks of Mexicali across seasons and time periods; and (b) Seasonal and annual Lmax footprint area and its proportion relative to the total urbanized area of Mexicali.

(a)						(b)
Lmax (m)		Min	Max	Avg.	Std. Dev.	Lmax Footprint Area
						Area (ha)	% Urban Area
Spring	Day	0.00	460	126.09	116.17	4778.44	21%
	Night	0.00	360	48.21	40.71	1293.50	6%
Summer	Day	0.00	490	119.03	119.32	4692.09	21%
	Night	0.00	430	52.71	40.05	1432.27	6%
Autumn	Day	0.00	500	114.32	121.54	4584.53	20%
	Night	0.00	370	52.48	46.31	1450.61	6%
Winter	Day	0.00	500	121.15	121.95	4500.22	20%
	Night	0.00	330	38.00	50.3	1067.92	5%
Annual	Day	0.00	500	120.15	119.74	7206.85	32%
	Night	0.00	430	47.85	44.90	2404.67	11%
	Total	0.00	500	84.00	97.37	7639.89	34%

). During the day, the average Lmax is highest in spring (126.09 m) and autumn (114.32 m), while the smallest daytime extent occurs in summer (119.03 m). At night, the average Lmax reaches its peak in spring (48.21 m), with the lowest values recorded in autumn (52.48 m) and winter (38.00 m).

The frequency distribution of parks by Lmax shows a marked contrast between day and night (

Table A4. Frequency distribution of parks by SCI Lmax across seasons and time of day. Table (a) reports the absolute number of parks per Lmax interval, and Table (b) shows the corresponding percentage based on the full park sample (N = 435).

(a)
Lmax (m) n = 435	Spring		Summer		Autumn		Winter
	Day	Night	Day	Night	Day	Night	Day	Night
0	122	99	133	83	153	111	133	190
10–100	94	309	96	337	85	287	94	210
110–200	113	23	111	12	97	32	105	27
210–300	71	3	57	1	64	4	60	6
310–400	28	1	27	1	27	1	33	2
410–500	7	0	11	1	9	0	10	0
(b)
0	28%	23%	31%	19%	35%	26%	31%	44%
10–100	22%	71%	22%	77%	20%	66%	22%	48%
110–200	26%	5%	26%	3%	22%	7%	24%	6%
210–300	16%	1%	13%	0%	15%	1%	14%	1%
310–400	6%	0%	6%	0%	6%	0%	8%	0%
410–500	2%	0%	3%	0%	2%	0%	2%	0%

). During daytime across all seasons, a substantial proportion of parks exhibit Lmax values between 110 and 200 m. For instance, 26% of parks in spring and summer, and 24% in winter, fall within this range, highlighting a consistent and measurable daytime cooling footprint.

At night, however, the SCI extent is considerably shorter. In summer, 71% of parks exhibit Lmax values between 10 and 100 m, and in winter, this proportion is 48%. These values reflect a more spatially limited nocturnal cooling influence (Table A4).

Moreover, a significant proportion of parks exhibit no SCI effect (Lmax = 0) during both day and night. In summer, 31% of parks show no SCI effect during the day, while this percentage drops to 19% at night. In spring, 28% of parks record no SCI effect during the day, compared to 23% at night. In autumn, 35% of parks show no cooling during the day, while 26% exhibit no SCI effect at night. Finally, in winter, 31% of parks lack a cooling effect during the day, but this increases to 44% at night. These null cases largely reflect ecological and morphological constraints: Many parks lacked sufficient vegetation since they were defined by typology rather than strictly as green areas, while small or irregular shapes and adjacency to impervious surfaces (e.g., roads, dense built-up zones) further limited their cooling potential, especially during seasons with low vegetation activity.

From a spatial perspective, the cumulative SCI extent across all parks and seasons results in a measurable cooling footprint (Figure 8). During the daytime, the SCI effect reaches over 7206.85 hectares annually, covering approximately 32% of the total urbanized area (Figure 9). In contrast, the nighttime SCI footprint is considerably smaller, totaling 2404.67 hectares (11%)(Table 4b). This stark disparity highlights the diminished cooling capacity of parks after sunset, likely due to reduced thermal inertia and lower heat storage capacity in vegetated surfaces during the night.

Seasonal differences in the daytime SCI footprint are also evident. In both spring and summer, the daytime cooling footprint exceeds 4700 hectares (21% of the urban area). Slightly smaller extents are observed in autumn (4584.53 ha) and winter (4500.22 ha), representing 20% each. At night, SCI coverage remains consistently lower, around 6% of the urban area in most seasons. Winter shows the least extensive nocturnal cooling footprint at just 1067.92 hectares, equivalent to 5% of the city’s urban footprint.

3.1.3. SCI Intensity

Results on the SCI ΔTmax indicate a consistently stronger cooling effect during daytime across all seasons, while the nighttime influence is notably weaker. On an annual basis, the average ΔTmax during daytime is 0.81 °C, with a maximum of 6.41 °C. At night, the annual average ΔTmax is 0.43 °C, reaching a maximum of 1.94 °C (

Table 5. Descriptive statistics of SCI intensity (ΔTmax, in °C) across seasons and time of day for 435 urban parks in Mexicali.

ΔTmax (°C) n = 435	Spring		Summer		Autumn		Winter		Annual
	Day	Night	Day	Night	Day	Night	Day	Night	Day	Night	Total
Min	0	0	0	0	0	0	0	0	0	0	0
Max	5.82	1.62	6.41	1.94	4.49	1.83	4.42	1.94	6.41	1.94	6.41
Avg	1.01	0.43	0.9	0.56	0.56	0.47	0.72	0.27	0.81	0.43	0.62
Std. Dev.	1.15	0.4	1.07	0.47	0.47	0.43	0.89	0.35	1	0.43	0.79

).

Daytime SCI intensity exhibits seasonal variability. The highest average ΔTmax is observed in spring (1.01 °C), followed by summer (0.90 °C), while autumn and winter present lower mean values of 0.56 °C and 0.72 °C, respectively. The maximum observed daytime ΔTmax of 6.41 °C occurred in summer. Notably, all seasons include parks with ΔTmax = 0 °C, indicating an absence of detectable cooling.

The frequency distribution (

Table A5. Frequency distribution of parks by SCI ΔTmax across daytime and nighttime periods. ((a,b): Counts; (c,d): Percentages).

(a)					(b)
ΔTmax (°C)	Day				ΔTmax (°C)	Night
n = 435	Spring	Summer	Autumn	Winter	n = 435	Spring	Summer	Autumn	Winter
0	122	133	153	133	0	99	83	111	190
<1.28	168	175	208	298	<0.39	120	97	101	108
<2.57	98	92	61	4	<0.78	130	109	103	101
<3.85	34	26	11	0	<1.17	65	94	90	27
<5.14	9	7	2	0	<1.56	18	44	26	6
<6.42	4	2	0	0	<1.95	3	8	4	3
(c)					(d)
ΔTmax (°C)	Day				ΔTmax (°C)	Night
n = 435	Spring	Summer	Autumn	Winter	n = 435	Spring	Summer	Autumn	Winter
0	28%	31%	35%	31%	0	23%	19%	26%	44%
<1.28	39%	40%	48%	69%	<0.39	28%	22%	23%	25%
<2.57	23%	21%	14%	1%	<0.78	30%	25%	24%	23%
<3.85	8%	6%	3%	0%	<1.17	15%	22%	21%	6%
<5.14	2%	2%	0%	0%	<1.56	4%	10%	6%	1%
<6.42	1%	0%	0%	0%	<1.95	1%	2%	1%	1%

) shows that 28–35% of parks presented no daytime SCI effect, particularly during autumn (35%) and winter (31%). Parks with moderate cooling intensity (ΔTmax between 1.28 °C and 2.57 °C) were more common in spring (39%) and summer (40%). Parks exhibiting very high SCI intensity (ΔTmax > 3.85 °C) were rare and only observed in a few cases during spring and summer.

At night, SCI intensity was generally lower across all seasons. The highest average nighttime ΔTmax was recorded in summer (0.56 °C), followed by autumn (0.47 °C), spring (0.43 °C), and winter (0.27 °C). Although maximum nighttime ΔTmax values reached 1.94 °C in both summer and winter, a substantial proportion of parks showed no nighttime cooling effect: 44% in winter, 26% in autumn, 23% in spring, and 19% in summer. Most parks that exhibited nighttime cooling recorded low to moderate ΔTmax values (<0.78 °C), particularly in spring and summer. High nighttime SCI intensities (ΔTmax > 1.17 °C) were rare across all seasons, with only a small subset of parks exceeding these thresholds.

3.2. Physical Characteristics of the Parks

3.2.1. Park Size and Shape Characteristics

The spatial characteristics of urban parks in Mexicali indicate a predominance of small, simply shaped areas with moderate structural complexity. Park sizes range from 0.03 to 55.32 hectares, with a mean of 1.02 hectares and a standard deviation of 3.77 hectares. Only 10 parks exceed 5 hectares, reinforcing the dominance of small-scale green spaces in the city. The LSI varies between 1.00 (a near-perfect circle) and 3.90, with an average of 1.28 and standard deviation of 0.31, while FD values range from 1.23 to 1.56, averaging 1.37 with a standard deviation of 0.06. These metrics reflect generally simple to moderately complex park geometries.

3.2.2. Land Cover Composition and Diversity

NDVI-based analysis shows clear seasonal variation in vegetation cover across the three spatial domains: within parks, the Lmax footprint area, and the surrounding areas up to 500 m (

Table A6. Seasonal distribution of land cover by NDVI ranges in (a) urban parks, (b) Lmax footprint areas, and (c) surrounding areas outside the Lmax footprint (up to 500 m buffer).

Total Area (ha)	No. of Parks	Season		NDVI Ranges Cover Percentage (%)
				<0.10	0.10 – 0.20	0.20 – 0.30	0.30 – 0.40	0.40 – 0.50	0.50 – 0.60	>0.60	Non-Vegetated (NDVI < 0.30)	Vegetation (NDVI ≥ 0.30)
(a) Parks
441.55	435	SP		12.8	24.2	16.2	14.4	11.7	10.2	10.6	53.1	46.9
	435	SU		24.8	19.4	15.3	13.4	11.2	10.2	5.8	59.5	40.5
	435	AU		15.9	18.6	13.3	12.2	10.3	9.4	20.4	47.7	52.3
	435	W		39.8	32.7	19.5	7.09	0.81	0.02	0.00	92.07	7.93
(b) Lmax footprint area
4706.97	313	SP	D	54.1	35.6	6.8	2.2	0.8	0.4	0.2	96.5	3.5
1258.38	336		N	50.2	36.7	8.5	3.0	1.1	0.4	0.2	95.4	4.6
4641.68	302	SU	D	78.2	15.4	4.2	1.4	0.5	0.2	0.1	97.8	2.2
1391.30	352		N	73.7	18.1	5.3	1.9	0.6	0.3	0.1	97.1	2.9
4510.06	282	AU	D	61.8	26.0	7.1	2.8	1.2	0.5	0.4	95.0	5.0
1420.69	324		N	58.3	27.0	8.6	3.4	1.6	0.6	0.4	93.9	6.1
4442.52	302	W	D	80.0	17.2	2.2	0.4	0.1	0.1	0.0	99.4	0.6
1048.14	245		N	79.0	18.1	2.5	0.4	0.0	0.0	0.0	99.6	0.4
(c) Outside Lmax area (up to 500 m)
12,945.77	313	SP	D	52.9	34.9	7.0	2.5	1.0	0.6	1.1	94.8	5.2
14,688.04	336		N	53.1	35.1	6.9	2.4	1.0	0.5	1.0	95.1	4.9
12,521.93	302	SU	D	76.9	15.6	4.5	1.6	0.7	0.3	0.4	96.9	3.1
15,631.78	352		N	77.2	15.6	4.3	1.5	0.7	0.3	0.4	97.1	2.9
12,024.99	282	AU	D	60.0	26.2	7.3	3.2	1.5	0.8	0.9	93.5	6.5
14,559.06	324		N	60.7	26.1	7.2	3.0	1.4	0.8	0.8	94.0	6.0
12,516.01	302	W	D	78.4	17.6	2.6	0.6	0.3	0.2	0.3	98.7	1.3
11,983.32	245		N	79.4	17.2	2.4	0.5	0.2	0.1	0.1	99.0	1.0

SP: spring; SU: summer; AU: autumn; W: winter; D: day; N: night.

). Inside parks, non-vegetated surfaces (NDVI < 0.30) dominate in all seasons, increasing from 53.1% in winter and spring to 59.5% in summer. NDVI > 0.30 cover is highest in autumn (52.3%), followed by spring and winter (46.9%), but drops significantly in summer (40.5%), highlighting vegetation stress under extreme arid conditions. In the Lmax areas, vegetation cover is limited but relatively stable, ranging from 10% to 15% across all seasons. In contrast, the areas beyond Lmax up to 500 m are overwhelmingly non-vegetated, with NDVI < 0.30 accounting for 90–95% of land cover, and peak vegetation presence reaching just 6% in spring.

The RF classification produced consistently higher estimates of vegetation cover compared to NDVI thresholding (

Table A7. Seasonal distribution of land cover by RF classification in (a) urban parks, (b) Lmax footprint areas, and (c) surrounding areas outside the Lmax footprint (up to 500 m buffer).

Total Area (ha)	No. of Parks	Season		Land Cover Classes by RF (%)
				Agriculture	Trees and Shrubs	Grass	Sandy Rustic Soil	Rocky Rustic Soil	Sandy Urban Soil	Water	Concrete	Industrial	Asphalt	Buildings	Non-Vegetated LC	Vegetation LC
(a) Parks
441.55	435	SP		0.1	28.7	31.6	5.2	0.8	9.6	4.1	7.2	0.8	2.1	9.8	60.4	39.6
	435	SU		0.4	27.3	28.8	3.9	0.3	11.5	4.6	3.0	0.4	5.2	14.6	56.5	43.5
	435	AU		1.3	38.6	21.1	3.9	1.6	8.6	2.6	5.3	0.2	3.3	13.6	61.0	39.0
	435	W		1.3	28.9	29.6	1.3	1.3	15.0	4.9	6.8	0.4	2.9	7.6	59.8	40.2
(b) Lmax footprint area
4706.97	313	SP	D	0.0	7.7	2.7	4.4	3.2	14.9	0.3	14.3	1.7	9.1	41.8	89.6	10.4
1258.38	336		N	0.0	8.0	3.5	0.8	1.4	11.1	1.8	17.1	0.9	15.7	39.6	88.5	11.5
4641.68	302	SU	D	0.0	3.4	3.4	3.0	1.2	22.0	0.2	12.4	1.7	11.4	41.3	93.2	6.8
1391.30	352		N	0.0	5.8	4.4	2.2	1.0	16.2	0.4	12.5	0.8	13.1	43.6	89.8	10.2
4510.06	282	AU	D	0.4	5.7	2.5	3.4	2.5	13.8	0.7	14.9	0.7	9.4	46.0	91.4	8.6
1420.69	324		N	0.6	8.8	2.0	1.8	1.4	7.0	1.1	17.6	0.4	9.9	49.4	88.7	11.3
4442.52	302	W	D	0.2	5.3	3.6	1.2	2.1	19.6	1.2	13.8	1.0	13.1	38.9	90.9	9.1
1048.14	245		N	0.0	7.8	2.9	0.3	1.0	8.6	1.7	15.9	0.5	16.9	44.4	89.3	10.7
(c) Outside Lmax area (up to 500 m)
12,945.77	313	SP	D	0.4	7.6	3.7	5.9	3.3	14.3	0.2	14.6	2.6	9.4	38.0	88.2	11.8
14,688.04	336		N	0.3	8.0	3.7	0.0	3.0	13.3	0.2	15.1	2.5	10.5	43.8	88.3	12.0
12,521.93	302	SU	D	0.1	3.9	4.1	0.0	1.5	24.5	0.3	13.5	2.1	11.6	38.4	91.9	8.1
15,631.78	352		N	0.1	3.8	4.1	0.0	1.4	24.4	0.3	13.3	2.1	11.4	39.7	92.6	8.1
12,024.99	282	AU	D	0.6	6.7	2.8	0.0	3.0	14.2	0.6	15.8	1.2	10.5	45.0	90.3	10.1
14,559.06	324		N	0.6	6.3	2.7	0.0	2.6	13.0	0.6	16.1	1.0	10.8	46.9	91.0	9.7
12,516.01	302	W	D	0.4	5.3	4.3	0.0	2.2	18.5	1.1	14.8	1.6	15.5	36.4	90.0	10.0
11,983.32	245		N	0.2	5.3	3.4	0.0	1.8	15.8	1.1	14.4	1.5	16.4	40.5	91.5	8.9

SP: spring; SU: summer; AU: autumn; W: winter; D: day; N: night.

). Within parks, vegetation is the dominant land cover type, with trees and shrubs comprising 27% to 38%, and grass contributing between 21% and 32%, peaking during autumn. In contrast, vegetation cover declines markedly within the Lmax footprint areas, where trees and shrubs account for only 7–12%, and grass for 4–8%. In these zones, non-vegetated surfaces—including urban soils, buildings, and concrete—dominate, collectively exceeding 80% of the total land cover. Outside the Lmax footprint area, extending up to 500 m from the parks, vegetation becomes increasingly sparse. Trees and shrubs contribute only 1–5%, while grass rarely exceeds 3.6%. In these outer zones, non-vegetated surfaces, primarily buildings, urban soils, and concrete, are predominant, making up 90–95% of the total land cover. This contrast highlights the critical role of parks in maintaining vegetation cover amidst the surrounding urban fabric.

3.2.3. Vegetation Cover Structure and Landscape Metrics

The NDVI > 0.30 analysis indicates that 364 parks (84%) contain some proportion of vegetation cover in spring, decreasing to 341 parks (78%) in summer, and peaking at 390 parks (90%) in autumn. Winter shows the lowest presence, with only 231 parks (53%) exceeding the vegetation threshold. In contrast, the RF classification consistently identifies a higher number of vegetated parks across all seasons, with limited seasonal variability. Specifically, vegetation is detected in 389 parks (89%) during spring, 379 (87%) in summer, 383 (88%) in autumn, and 387 (89%) in winter.

The combined LSI and FD metrics underline the seasonal dynamics and methodological differences in understanding the structural and compositional complexity of vegetation within urban parks. For NDVI > 0.30, LSI values < 2.0 suggest compact vegetation shapes, with greater irregularity in spring and autumn. FD values between 1.50 and 1.55 dominate in those seasons, while winter exhibits simpler geometries. RF classifications consistently detect more structurally complex vegetation patterns than NDVI-based methods, particularly during low-growth periods like winter.

In addition, the H′ index further underscores seasonal and methodological variation. NDVI-based H′ values peak in autumn, with a greater number of parks exhibiting moderate diversity levels (1.35–1.65) and dropping significantly in winter. Meanwhile, RF classifications consistently capture higher diversity levels across all seasons, with H′ values remaining stable and highlighting the method’s ability to detect nuanced variations. This stability reinforces the robustness of RF in providing a detailed and comprehensive view of vegetation diversity, particularly during periods of lower vegetation activity.

3.3. Influence of Physical Descriptors on the SCI

3.3.1. Bivariate Correlation Between SCI and Physical Descriptors

The bivariate correlation analyses confirm that the cooling effectiveness of parks, both in SCI ΔTmax and Lmax, is strongly determined by internal vegetation structure and, to a lesser extent, by the configuration of adjacent areas. High vegetative cover, low fragmentation, and greater spatial cohesion consistently support broader and stronger cooling performance.

Results first highlight a robust and statistically significant interdependence between SCI ΔTmax and Lmax across all seasons and time periods (r ≥ 0.497 **, p < 0.01), indicating that parks capable of producing greater temperature reductions also tend to extend their cooling influence farther into the urban matrix (

Table A8. Bivariate correlation (Pearson, r) between physical descriptors and SCI Lmax across seasons and time periods (day/night): (a) Within parks; (b) Lmax footprint areas; and (c) Surrounding areas outside the Lmax footprint (up to 500 m buffer).

SCI Lmax (m) n = 148		Spring		Summer		Autumn		Winter
		Day	Night	Day	Night	Day	Night	Day	Night
SCI indicators	SCI Lmax (m)	1.000	1.000	1.000	1.000	1.000	1.000	1.000	1.000
	SCI ΔTmax (°C)	0.876 **	0.659 **	0.827 **	0.548 **	0.868 **	0.497 **	0.874 **	0.644 **
(a) Park
Shape	Area (m2)	0.018	0.045	−0.042	0.014	−0.170 *	−0.093	−0.163 *	−0.013
	LSI	0.023	0.113	0.006	0.114	0.027	0.112	0.023	0.105
	FD	0.051	0.139	0.048	0.146	0.122	0.212 **	0.135	0.117
Location	Longitude (m)	−0.080	0.114	−0.049	0.083	0.025	0.138	−0.020	−0.007
	Latitude (m)	0.325 **	0.151	0.146	0.129	0.192 *	0.137	0.241 **	0.211 **
	Elevation (masl)	−0.110	−0.037	−0.106	−0.032	−0.084	0.023	−0.071	−0.092
Vegetation (NDVI > 0.30)	Proportion (%)	0.555 **	0.511 **	0.595 **	0.411 **	0.629 **	0.444 **	0.467 **	0.340 **
	Area (m2)	0.313 **	0.262 **	0.294 **	0.242 **	0.128	0.132	0.179 *	0.254 **
	LSI	0.022	0.081	−0.024	0.134	−0.266 **	0.040	−0.018	0.196
	FD	−0.439 **	−0.365 **	−0.433 **	−0.159	−0.570 **	−0.266 **	−0.166	−0.098
	H′	0.434 **	0.468 **	0.432 **	0.530 **	0.211 **	0.283 **	0.320 **	0.231 **
Vegetation (RF)	Proportion (%)	0.540 **	0.555 **	0.570 **	0.470 **	0.618 **	0.476 **	0.431 **	0.393 **
	Area (m2)	0.250 **	0.244 **	0.214 **	0.213 **	0.094	0.113	0.042	0.137
	LSI	−0.194 *	−0.105	−0.140	0.123	−0.338 **	−0.039	−0.274 **	−0.146
	FD	−0.518 **	−0.444 **	−0.500 **	−0.333 **	−0.568 **	−0.285 **	−0.391 **	−0.339 **
	H′	−0.309 **	−0.178 *	−0.220 **	0.024	−0.545 **	−0.261 **	−0.261 **	−0.186 *
(b) Lmax footprint area
Vegetation (NDVI > 0.30)	Proportion (%)	−0.075	−0.041	−0.130	0.196 *	−0.205 *	−0.022	−0.094	−0.005
	Area (m2)	0.514 **	0.220 *	0.337 **	0.692 **	0.400 **	0.466 **	0.175	0.222 *
	LSI	0.716 **	0.369 **	0.519 **	0.500 **	0.674 **	0.572 **	0.146	0.160
	FD	0.398 **	0.338 **	0.312 **	0.057	0.470 **	0.293 **	0.139	0.050
	H′	−0.097	−0.132	−0.209 *	0.170	−0.254 *	−0.040	−0.220 *	−0.104
Vegetation (RF)	Proportion (%)	−0.221 *	−0.117	−0.232 *	0.127	−0.264 **	−0.079	−0.289 **	−0.190
	Area (m2)	0.586 **	0.340 **	0.382 **	0.704 **	0.411 **	0.402 **	0.470 **	0.364 **
	LSI	0.855 **	0.547 **	0.708 **	0.528 **	0.704 **	0.554 **	0.792 **	0.534 **
	FD	0.372 **	0.277 **	0.409 **	0.080	0.369 **	−0.790	0.450 **	0.198
	H′	0.104	−0.100	0.001	0.157	0.003	0.003	−0.017	−0.214 *
(c) Outside Lmax area (up to 500 m)
Vegetation (NDVI > 0.30)	Proportion (%)	0.004	−0.102	0.057	−0.085	−0.103	0.023	−0.098	−0.054
	Area (m2)	−0.183	−0.062	−0.114	−0.059	−0.268 **	−0.088	−0.017	−0.041
	LSI	−0.474 **	−0.044	−0.396 **	0.015	−0.575 **	−0.060	−0.277 **	0.021
	FD	−0.238 *	0.063	−0.202 *	0.016	−0.334 **	0.049	−0.195	0.086
	H′	0.061	−0.059	0.023	−0.014	−0.165	0.059	−0.231 *	−0.040
Vegetation (RF)	Proportion (%)	0.027	−0.072	−0.003	−0.005	−0.095	−0.048	−0.289 **	−0.190
	Area (m2)	−0.010	−0.029	−0.084	0.035	−0.129	−0.056	−0.207 *	−0.123
	LSI	−0.547 **	0.101	−0.472 **	0.077	−0.572 **	−0.031	−0.550 **	−0.089
	FD	−0.411 **	0.082	−0.201 *	−0.051	−0.381 **	0.053	−0.161	0.068
	H′	0.127	0.065	0.006	0.096	0.051	0.008	−0.056	−0.179

** p < 0.01; * p < 0.05.

and

Table A9. Bivariate correlation between physical descriptors and SCI ΔTmax across seasons and time periods (day/night): (a) Within parks; (b) Lmax footprint areas; and (c) Surrounding areas outside the Lmax footprint (up to 500 m buffer).

SCI ΔTmax (°C) n = 148		Spring		Summer		Autumn		Winter
		Day	Night	Day	Night	Day	Night	Day	Night
SCI indicators	SCI Lmax (m)	0.876 **	0.659 **	0.827 **	0.548 **	0.868 **	0.497 **	0.874 **	0.644 **
	SCI ΔTmax (°C)	1.000	1.000	1.000	1.000	1.000	1.000	1.000	1.000
(a) Park
Shape	Area (m2)	−0.021	0.111	−0.019	−0.036	−0.124	−0.059	−0.134	0.031
	LSI	−0.035	0.012	−0.029	0.036	−0.047	0.01	−0.056	0.012
	FD	0.000	0.035	−0.009	0.122	0.027	0.081	0.03	0.032
Location	Longitude (m)	−0.071	0.064	−0.075	0.088	−0.034	0.082	−0.056	−0.008
	Latitude (m)	0.292 **	0.305 **	0.212 **	0.329 **	0.151	0.422 **	0.210 *	0.289 **
	Elevation (masl)	−0.062	−0.072	−0.071	−0.047	−0.092	−0.044	−0.022	−0.115
Vegetation (NDVI > 0.30)	Proportion (%)	0.611 **	0.831 **	0.637 **	0.903 **	0.569 **	0.877 **	0.478 **	0.512 **
	Area (m2)	0.322 **	0.504 **	0.388 **	0.448 **	0.161	0.362 **	0.335 **	0.336 **
	LSI	−0.087	0.005	0.007	0.049	−0.255 **	−0.133	0.009	0.230 *
	FD	−0.528 **	−0.738 **	−0.467 **	−0.634 **	−0.544 **	−0.737 **	−0.345 **	−0.227 *
	H′	0.374 **	0.518 **	0.452 **	0.696 **	0.128	0.316 **	0.261 **	0.274 **
Vegetation (RF)	Proportion (%)	0.558 **	0.805 **	0.593 **	0.896 **	0.546 **	0.870 **	0.416 **	0.505 **
	Area (m2)	0.234 **	0.423 **	0.274 **	0.341 **	0.124	0.307 **	0.075	0.207 *
	LSI	−0.266 **	−0.265 **	−0.175 *	−0.154	−0.337 **	−0.321 **	−0.275 **	−0.217 **
	FD	−0.565 **	−0.739 **	−0.567 **	−0.760 **	−0.558 **	−0.756 **	−0.415 **	−0.453 **
	H′	−0.409 **	−0.501 **	−0.274 **	−0.302 **	−0.475 **	−0.562 **	−0.321 **	−0.335 **
(b) Lmax footprint area
Vegetation (NDVI > 0.30)	Proportion (%)	0.070	0.277 **	0.011	0.240 **	−0.081	0.184 *	0.057	−0.008
	Area (m2)	0.436 **	0.343 **	0.313 **	0.168	0.377 **	0.222 *	0.229 *	0.057
	LSI	0.584 **	0.445 **	0.490 **	0.240 *	0.524 **	0.309 **	0.222	0.079
	FD	0.258 **	0.152	0.286 **	0.041	0.274 **	0.150	0.042	0.017
	H′	0.010	0.300 **	−0.042	0.250 **	−0.111	0.206 *	−0.086	−0.054
Vegetation (RF)	Proportion (%)	−0.096	0.218 *	−0.059	0.266 **	−0.134	0.257 **	−0.132	−0.029
	Area (m2)	0.502 **	0.317 **	0.355 **	0.157	0.379 **	0.258 **	0.422 **	0.165
	LSI	0.619 **	0.416 **	0.514 **	0.318 **	0.533 **	0.307 **	0.604 **	0.208 *
	FD	0.156	−0.007	0.199 *	0.016	0.177	0.257 **	0.266 **	0.011
	H′	0.064	0.047	−0.060	−0.097	0.067	−0.029	−0.015	−0.163
(c) Outside Lmax area (up to 500 m)
Vegetation (NDVI > 0.30)	Proportion (%)	0.028	−0.056	0.038	0.074	−0.100	0.025	−0.035	−0.027
	Area (m2)	−0.016	−0.022	−0.106	0.05	−0.233 *	0.097	0.057	−0.004
	LSI	−0.339 **	0.267 **	−0.176	−0.215 **	−0.470 **	0.170	−0.197	0.154
	FD	−0.214 *	0.192 *	−0.067	0.183 *	−0.238 *	−0.107	−0.215 *	0.125
	H′	0.093	0.027	0.090	0.009	−0.151	0.068	−0.119	−0.036
Vegetation (RF)	Proportion (%)	−0.009	−0.014	0.028	0.04	−0.032	0.153	−0.132	0.043
	Area (m2)	0.077	−0.020	0.000	0.086	−0.036	0.168	−0.015	0.064
	LSI	−0.415 **	0.164	−0.244 *	0.242 **	−0.447 **	0.153	−0.376 **	−0.046
	FD	−0.388 **	−0.100	−0.165	0.133	−0.270 **	0.076	−0.17	−0.041
	H′	0.162	0.201 **	0.600 *	−0.059	0.093	−0.078	0.016	−0.211 *

** p < 0.01; * p < 0.05.

).Vegetation coverage within the park emerged as the primary determinant of SCI effectiveness. The proportion and area of vegetated surfaces, whether measured by NDVI (threshold > 0.30) or RF-derived classifications, showed consistently strong positive correlations with both ΔTmax and Lmax in all seasons (e.g., NDVI-based vegetation proportion and ΔTmax: r = 0.903 ** on a summer night; RF-based vegetation proportion and Lmax: r = 0.618 ** on an autumn day).

In contrast, vegetation configuration metrics showed consistently strong negative correlations with both SCI indicators. FD exhibited particularly high negative values (e.g., RF-based FD and ΔTmax: r = −0.565 **; FD and Lmax: r = −0.518 **), suggesting that fragmented vegetation limits both cooling intensity and spatial reach. LSI also showed negative or weak associations (e.g., RF-based LSI and ΔTmax: r = −0.266 **), indicating that compact vegetation forms enhance the SCI efficiency. H′ index, derived from NDVI, was positively correlated with ΔTmax (e.g., r = 0.696 **, summer night), suggesting that internal heterogeneity may promote cooling. However, its relationship with Lmax was weak or negative, especially in RF-based metrics, likely due to mixed land cover influences.

In the Lmax footprint area, the vegetated area and LSI were positively associated with both indicators (e.g., LSI and Lmax: r = 0.855 **), highlighting the role of spatial continuity in cooling propagation (Table A8b and Table A9b). FD and H′ showed weaker or inconsistent effects, suggesting a more context-dependent influence.

In the surrounding buffer (up to 500 m), vegetation metrics had mostly negative correlations with SCI indicators (e.g., NDVI-based LSI and ΔTmax: r = −0.547 **; RF-based LSI and Lmax: r = −0.550 **), reinforcing that fragmented vegetation beyond parks may hinder SCI spread ( Table A8c and Table A9c). This highlights that vegetation must be not only present, but also spatially coherent to effectively support cooling performance.

3.3.2. Linear Regression-Based Prediction of SCI

The linear regression models revealed that the ability to predict SCI behavior varies significantly depending on both the indicator (Lmax vs. ΔTmax) and the type of vegetation retrieval method (NDVI vs. RF). Overall, SCI Lmax was more accurately predicted than SCI ΔTmax, with particularly strong results obtained using Random Forest-derived vegetation metrics.

The NDVI-based regression models (

Table 6. NDVI-based regression model for SCI Lmax.

Lmax (m); N = 148 NDVI-Based Variables		Spring		Summer		Autumn		Winter
		Day	Night	Day	Night	Day	Night	Day	Night
Constant		182.12	315.32	93.95	325.37	604.68	337.19	−936.80	111.95
Park Shape	Area (m2)	e.	e.	e.	e.	e.	0.001	e.	e.
	LSI	e.	−8.83	e.	−36.99	−80.55	−20.70	−143.23	e.
	FD	e.	e.	e.	216.28	679.91	e.	1120.02	e.
Park Location	Longitude (m)	e.	e.	e.	e.	e.	e.	e.	e.
	Latitude (m)	e.	e.	e.	e.	e.	e.	e.	e.
	Elevation (masl)	e.	e.	e.	e.	e.	e.	e.	e.
Park Vegetation NDVI > 0.30	Proportion (%)	e.	e.	e.	−16.57	−109.87	e.	−129.86	e.
	LSI	−15.51	−7.99	−11.86	−5.26	e.	e.	e.	e.
	FD	e.	e.	e.	e.	−367.74	e.	e.	e.
	H′	e.	e.	e.	e.	e.	e.	−73.48	e.
Lmax Vegetation NDVI > 0.30	Proportion (%)	e.	−137.52	−774.42	−831.08	e.	−204.14	−3570.29	−2319.44
	LSI	34.20	25.67	48.58	37.24	33.67	32.48	65.92	42.45
	FD	e.	−136.13	−303.96	−330.30	−346.57	−326.92	e.	e.
	H′	−114.53	−77.87	−100.62	e.	−128.72	−72.76	−194.01	−120.53
Outside Lmax Vegetation NDVI > 0.30	Proportion (%)	−477.64	−120.69	e.	e.	e.	e.	e.	e.
	LSI	−13.45	−1.93	−24.73	−1.77	−8.64	−3.07	−15.54	e.
	FD	e.	e.	372.47	e.	e.	164.09	e.	e.
	H′	117.54	34.36	104.54	e.	e.	36.83	146.58	e.
R2		0.87	0.59	0.83	0.66	0.88	0.80	0.74	0.41
Std. Error		37.87	15.22	43.99	19.09	34.26	17.71	54.54	41.31

e. excluded after backward elimination.

and

Table 7. NDVI-based regression model for SCI ΔTmax.

ΔTmax (°C); N = 148 NDVI-Based Variables		Spring		Summer		Autumn		Winter
		Day	Night	Day	Night	Day	Night	Day	Night
Constant		1.59	2.36	20.14	−1.84	136.39	−96.50	9.06	9.99
Park Shape	Area (m2)	e.	0.001	e.	e.	e.	e.	e.	e.
	LSI	e.	−0.14	e.	−0.50	e.	−0.11	−0.34	e.
	FD	e.	e.	−2.59	3.65	e.	e.	e.	e.
Park Location	Longitude (m)	e.	e.	0.00	e.	0.00	e.	e.	e.
	Latitude (m)	e.	e.	e.	e.	0.00	0.00	e.	e.
	Elevation (masl)	e.	e.	e.	e.	e.	e.	−0.06	e.
Park Vegetation NDVI > 0.30	Proportion (%)	0.89	0.82	1.11	1.09	e.	1.11	e.	0.80
	LSI	−0.24	e.	e.	0.13	e.	e.	e.	−0.11
	FD	e.	−1.24	e.	−1.41	−1.82	e.	e.	1.80
	H′	e.	−0.31	e.	−0.25	e.	−0.22	−0.81	e.
Lmax Vegetation NDVI > 0.30	Proportion (%)	e.	−2.23	e.	−3.41	e.	e.	e.	−24.11
	LSI	0.24	0.13	0.26	0.08	0.14	0.03	0.33	0.32
	FD	e.	e.	e.	e.	e.	1.58	e.	−4.38
	H′	e.	e.	−0.63	e.	e.	e.	−1.16	e.
Outside Lmax Vegetation NDVI > 0.30	Proportion (%)	e.	e.	−7.34	e.	e.	−2.03	e.	e.
	LSI	−0.13	e.	−0.16	e.	−0.09	−0.02	e.	0.13
	FD	e.	e.	e.	e.	e.	e.	−3.62	−3.58
	H′	e.	e.	1.77	e.	e.	0.43	e.	−0.74
R2		0.61	0.74	0.58	0.79	0.58	0.78	0.60	0.44
Std. Error		0.66	0.19	0.66	0.20	0.45	0.19	0.39	0.26

e. excluded after backward elimination.

) exhibited greater predictive capacity for SCI Lmax than for ΔTmax. Lmax models reached R² values of 0.88 (autumn day) and 0.87 (spring day), while ΔTmax models peaked at 0.79 (summer night) and 0.78 (autumn night). Performance declined during winter nights. Vegetation metrics within parks were significant predictors of both SCI indicators, though their direction and magnitude varied. For SCI Lmax, configuration metrics such as LSI and H′ frequently displayed negative coefficients (e.g., LSI = −80.55 in autumn day; H′ = −194.01 in winter day), suggesting that greater internal spatial complexity may inhibit the lateral spread of cooling. Conversely, ΔTmax models showed positive associations with vegetation proportion (e.g., 1.11 on a summer night), highlighting the role of vegetation density in peak cooling. However, FD and H′ were negatively associated with ΔTmax in several warm-season models (e.g., FD = −1.41 in summer night), suggesting that fragmentation limits cooling effectiveness.

In the Lmax footprint area, vegetation area and LSI consistently predicted Lmax positively (e.g., LSI = 65.92 in winter day), while vegetation proportion, FD, and H′ showed strong negative coefficients (e.g., proportion = −3570.29; FD = −367.74), reinforcing the importance of cohesive vegetation structure outside the park.

Beyond the SCI Lmax area, NDVI-based metrics had weaker and inconsistent effects. Lmax models showed negative coefficients for LSI and proportion (e.g., LSI = −15.54 on a winter day; proportion = −477.64 on a spring day), while ΔTmax models revealed sporadic positive associations, such as H′ = 1.77 on a summer day. These results suggest that broader vegetative heterogeneity may support localized cooling under certain conditions, but this effect diminishes in colder seasons and at night.

The RF-based models (

Table 8. RF-based regression model for SCI Lmax.

Lmax (m); N = 148 RF-Based Variables		Spring		Summer		Autumn		Winter
		Day	Night	Day	Night	Day	Night	Day	Night
Constant		823.18	163.94	823.18	4,248.07	16,467.37	5142.81	623.81	367.42
Park Shape	Area (m2)	e.	e.	e.	0.001	e.	0.001	e.	0.001
	LSI	e.	−17.44	e.	−8.79	e.	−25.58	e.	e.
	FD	e.	121.82	e.	e.	e.	e.	e.	e.
Park Location	Longitude (m)	e.	e.	e.	e.	e.	e.	e.	e.
	Latitude (m)	e.	e.	e.	0.00	0.00	0.00	e.	e.
	Elevation (masl)	e.	e.	e.	e.	e.	e.	e.	e.
Park Vegetation RF	Proportion (%)	e.	−20.63	e.	−21.28	e.	e.	e.	e.
	LSI	−23.03	−10.87	−23.03	−12.86	−29.90	e.	−43.36	e.
	FD	e.	e.	e.	e.	e.	e.	159.32	e.
	H′	e.	e.	e.	e.	e.	e.	e.	e.
Lmax Vegetation RF	Proportion (%)	−217.34	−74.54	217.34	−229.40	−341.92	−291.02	e.	e.
	LSI	33.04	18.31	33.04	34.60	37.17	35.55	36.37	35.43
	FD	−424.50	−153.21	424.50	−384.22	−531.80	−515.20	417.22	341.20
	H′	e.	−15.92	e.	e.	e.	−12.87	e.	−50.19
Outside Lmax Vegetation RF	Proportion (%)	e.	e.	e.	e.	e.	77.41	−221.63	−144.79
	LSI	−7.99	e.	−7.99	−1.63	−10.26	−2.53	−9.54	−4.00
	FD	e.	e.	e.	e.	e.	124.70	e.	157.49
	H′	e.	e.	e.	e.	e.	e.	e.	e.
R2		0.92	0.59	0.93	0.70	0.88	0.77	0.87	0.74
Std. Error (m)		27.73	15.48	27.73	17.35	34.31	19.08	38.43	23.48

e. excluded after backward elimination.

and

Table 9. RF-based regression model for SCI ΔTmax.

ΔTmax (°C); N = 148 RF-Based Variables		Spring		Summer		Autumn		Winter
		Day	Night	Day	Night	Day	Night	Day	Night
Constant		28.37	3.47	20.37	2.36	202.12	−50.37	19.18	0.42
Park Shape	Area (m2)	e.	0.001	e.	e.	e.	e.	e.	e.
	LSI	e.	−0.27	e.	−0.26	e.	e.	e.	e.
	FD	e.	2.30	e.	1.86	e.	−1.41	−2.02	e.
Park Location	Longitude (m)	0.001	e.	e.	e.	0.001	e.	0.001	e.
	Latitude (m)	e.	e.	e.	e.	0.001	0.001	e.	e.
	Elevation (masl)	0.02	e.	e.	e.	e.	e.	0.02	e.
Park Vegetation RF	Proportion (%)	e.	e.	e.	0.91	e.	0.90	e.	0.29
	LSI	e.	e.	e.	e.	e.	e.	e.	e.
	FD	−2.39	−2.46	−4.25	−0.96	−2.42	e.	e.	e.
	H′	−0.71	−0.17	e.	e.	e.	−0.14	−0.34	−0.20
Lmax Vegetation RF	Proportion (%)	e.	−0.36	e.	−1.35	−1.74	e.	e.	e.
	LSI	0.22	0.12	0.23	0.11	0.21	0.05	0.16	0.04
	FD	−5.44	−1.26	e.	−1.97	−5.18	e.	e.	e.
	H′	e.	e.	e.	e.	e.	−0.16	e.	e.
Outside Lmax Vegetation RF	Proportion (%)	−3.86	e.	−4.27	e.	−1.51	−0.60	e.	−0.58
	LSI	−0.10	e.	e.	e.	−0.06	e.	−0.07	e.
	FD	e.	e.	−7.62	e.	e.	e.	e.	e.
	H′	0.77	−0.19	−0.80	e.	0.53	e.	e.	e.
R2		0.64	0.73	0.50	0.75	0.62	0.74	0.55	0.31
Std. Error (°C)		0.65	0.20	0.72	0.21	0.43	0.21	0.47	0.25

e. excluded after backward elimination.

) demonstrated superior predictive capacity compared to NDVI-derived models. For SCI Lmax, RF models achieved R² values as high as 0.93 on a summer day, followed by 0.92 on a spring day and 0.88 on an autumn day, confirming robust performance under high thermal contrast conditions. The lowest Lmax performance was observed on a spring night (R² = 0.50), likely reflecting reduced radiative forcing and lower vegetation–atmosphere coupling at night. For SCI ΔTmax, R² values peaked at 0.75 in summer night, followed closely by 0.74 on an autumn night, while the poorest performance occurred during winter night (R² = 0.31), when surface thermal variability is minimal. These results underscore the enhanced ability of RF-derived vegetation descriptors to capture spatial heterogeneity and vegetation structure relevant to SCI dynamics across diverse seasonal and diurnal conditions.

Within the park, configuration metrics such as LSI and H′ were significant predictors of SCI Lmax (Table 8). LSI showed consistently negative coefficients in multiple daytime and nighttime models (e.g., LSI = −29.90 in autumn day; −43.36 in winter day), suggesting that irregular vegetation forms limit lateral cooling propagation. Similarly, H′ was negatively associated with Lmax in several nighttime models (e.g., −50.19 in winter night), indicating that high internal vegetation diversity may constrain SCI spread under low radiative conditions. In contrast, ΔTmax models (Table 9) showed weaker and more variable contributions from park-level configuration metrics. However, FD and H′ emerged as significant in some warm-season models (e.g., FD = −4.25 in summer day; H′ = −0.71 in spring day), suggesting that fragmented and heterogeneous vegetation may reduce cooling intensity, although these effects appear seasonally dependent and subtler than those observed for Lmax.

Within the Lmax footprint zone, RF-based vegetation metrics were the strongest and most consistent predictors of SCI Lmax (Table 8). LSI coefficients were uniformly positive across all day and night models (e.g., LSI = 37.17 in autumn day; 36.37 in winter day), reinforcing the importance of spatially cohesive vegetation patterns in facilitating cooling propagation. By contrast, FD and vegetation proportion had strong negative coefficients (e.g., FD = −531.80; proportion = −341.92 in autumn day), reflecting that fragmented or sparse vegetation in this zone constrains SCI extent.

For SCI ΔTmax (Table 9), vegetation metrics in the Lmax zone had limited influence. While some variables like FD (e.g., −5.18 in autumn day) and LSI (e.g., 0.21 in autumn day) showed moderate associations, overall contributions were smaller and less consistent than those observed for Lmax, reinforcing that cooling intensity is more strongly governed by in-park characteristics and atmospheric conditions.

In the area outside the Lmax footprint (up to 500 m), the vegetation structure showed mostly negative coefficients in the Lmax models, particularly for LSI (e.g., −10.26 in autumn day) and vegetation proportion (e.g., −221.63 in winter day), indicating that fragmented vegetation in the surrounding urban matrix may act as a barrier to cooling diffusion. ΔTmax models revealed some positive H′ coefficients in spring and summer (e.g., 0.77 in spring day), suggesting that heterogeneity in adjacent landscapes may occasionally enhance localized cooling, although this effect was minor and inconsistent across models.

In summary, these RF-based models confirm that SCI Lmax is more reliably predicted than SCI ΔTmax, and that vegetation configuration, particularly within the Lmax footprint, is critical for enhancing the cooling extent. Conversely, SCI ΔTmax appears to depend more on internal vegetation density and broader climatic factors. Promoting structured, cohesive vegetation patterns both inside and around urban parks may significantly improve the effectiveness of green infrastructure in mitigating urban heat.

4. Discussion

4.1. Design and Environmental Determinants of SCI

4.1.1. Vegetation Structure and Coverage Effects in Arid Cities

This study confirms that vegetation coverage and structural configuration are the most influential and consistent determinants of SCI performance in the urban parks of Mexicali, a desert city with extreme climatic conditions. Metrics derived from both NDVI thresholding and RF classification showed strong, statistically significant associations with SCI indicators, ΔTmax and Lmax, across seasons and day–night cycles (Figure 10). These findings reinforce previous research asserting that evapotranspiration, shading, and increased albedo, especially from trees and dense vegetation, are primary drivers of surface cooling in arid environments [57,58].

Across both modeling approaches, the proportion of vegetation within park boundaries emerged as the most stable and impactful predictor of daytime SCI intensity. According to the NDVI-based models, a 10% increase in vegetated area was associated with ΔTmax increases of approximately +0.09 to +0.11 °C, depending on the season, with the strongest effect observed in summer (+0.111 °C). Although this effect may appear modest, it is meaningful in the context of urban thermal mitigation, particularly during extreme heat periods. Similarly, within the Lmax footprint zone, a 10% increase in vegetation proportion corresponded to an increase of approximately +3.4 m to +6.6 m in Lmax, with the highest seasonal gain recorded in winter (NDVI-based model). These marginal gains compound with increased vegetation density, indicating that parks exceeding 60–70% vegetative cover are significantly more likely to produce measurable SCI effects. This threshold aligns with previous empirical findings and offers practical guidance for landscape architects aiming to maximize passive cooling in hot-arid environments.

Beyond overall vegetative cover, spatial configuration plays a critical role in SCI strength and propagation. Both FD and LSI exhibited negative associations with ΔTmax and Lmax in park cores and buffer zones. For instance, a 0.1-unit increase in FD—indicating greater shape fragmentation—was associated with a reduction of ~13 to 34 m in Lmax, and ~0.12 to 0.18 °C loss in ΔTmax, particularly during spring and summer. These results confirm that compact and cohesive vegetation patterns (i.e., low FD and LSI) enhance the efficiency of park-induced cooling by supporting continuous evapotranspiration and reducing heat exposure from impervious edges [8,59].

Land cover diversity (H′), produced more variable results. In NDVI-based regressions, H′ showed weak but positive associations with ΔTmax, potentially reflecting benefits from vertical layering or mixed vegetation types. In contrast, RF-based models often revealed negative or inconsistent associations, likely due to methodological sensitivity: RF classification may overestimate heterogeneity in sparsely vegetated parks, inflating H′ and diluting its predictive value for thermal outcomes. These results suggest that while diversity may improve ecological resilience, cooling performance benefits most from compact and connected vegetation cores, with heterogeneity best positioned peripherally or in non-thermal zones of the park [60].

Vegetation in the surrounding 500 m buffer had weak or negative associations with SCI. For instance, the RF model in summer found that a 10% increase in surrounding vegetation corresponded to a −0.076 °C decrease in ΔTmax, likely due to the reduced thermal contrast between parks and their environments. Although this may weaken SCI detectability, it does not imply diminished thermal comfort, and it reflects a diffuse cooling background that makes isolated park effects harder to quantify.

4.1.2. Park Size and Shape Influence

Contrary to conventional assumptions, the park area alone did not consistently predict the SCI intensity or extent. Correlations between total park size and SCI metrics were weak and seasonally inconsistent, suggesting that larger parks are not inherently more effective at reducing surface temperatures. This finding underscores the limitations of using area as a standalone indicator of thermal performance in urban green spaces.

In contrast, shape-based metrics (LSI and FD) demonstrated stronger and more consistent associations with SCI. Parks exhibiting high shape complexity or irregularity (higher LSI and FD) typically showed a reduced cooling performance. These results support previous findings indicating that compact, geometrically coherent green spaces yield stronger SCI effects compared to an elongated or fragmented parks of similar size [14,29].

Although park size had limited explanatory power, these observations suggest that spatial configuration plays a more decisive role in cooling effectiveness. In practice, this means that urban parks should prioritize compactness over irregularity, particularly in smaller or fragmented sites. A simple geometric layout not only enhances SCI performance but also facilitates more efficient irrigation, shading, and vegetation maintenance. Where irregular shapes are unavoidable due to urban constraints, designers should aim to concentrate dense vegetation cores in central areas, buffering them with lower-maintenance perimeter zones [61]. These findings reinforce prior recommendations to minimize edge-to-core ratios in arid green infrastructure planning.

4.1.3. Surrounding Land Cover Constraints

The SCI effect was found to be spatially confined, with limited propagation beyond park boundaries. Areas extending up to 500 m from the Lmax footprint area were dominated by impervious surfaces (concrete, asphalt, and rooftops) with vegetated land rarely exceeding 5%. This urban matrix reduces lateral thermal diffusion and confines cooling to park interiors.

Previous studies have shown that impervious surfaces elevate sensible heat flux, reducing the efficacy of adjacent vegetated areas in lowering temperatures [62]. By contrast, green corridors, vegetated streetscapes, or buffer zones can enhance thermal continuity and allow SCI effects to extend beyond individual parks [63].

Therefore, integrating parks into broader ecological networks is essential to maximize cooling benefits. Urban planning strategies that connect parks through continuous vegetation infrastructure could substantially enhance SCI coverage and improve neighborhood-level resilience to heat stress.

4.1.4. Geographic Location Effects

Location variables of latitude, longitude, and elevation showed limited but noteworthy associations with SCI behavior. Statistically significant correlations between latitude and Lmax during several daytime periods (e.g., summer: r = 0.325; winter: r = 0.241) suggest that parks situated further north in Mexicali exhibited broader SCI footprints. This spatial gradient may reflect differences in surrounding land use [9,32]: northern and central areas contain more residential vegetation, while southern zones are dominated by industrial facilities and impervious surfaces.

Longitude showed no consistent relationship with SCI, while elevation, given Mexicali’s uniformly flat terrain, presented no statistically significant correlations across seasons. Although these variables are not directly modifiable, their inclusion provides important spatial context for interpreting SCI variability across the urban landscape [7].

4.1.5. Climatic Considerations on SCI

Climatic factors, particularly rainfall and wind, remain external but influential variables in SCI dynamics. While satellite image selection avoided post-precipitation periods, Mexicali’s semi-arid climate (72–80 mm annual precipitation) naturally limits soil moisture and vegetative activity during most of the year [25].

Previous studies have shown that the cooling intensity of urban green spaces can increase after rainfall due to improved vegetation vitality and higher evapotranspiration rates [64]. Additionally, soil moisture availability plays a key role in supporting the temperature regulation functions of green spaces, particularly in dry climates [8].

Wind dynamics are another important variable influencing SCI behavior. Prevailing winds in Mexicali originate predominantly from the west and northwest, with average speeds ranging from 10 km/h in winter to 13–14 km/h in late spring [25]. These winds may disperse localized cool air masses generated by parks, particularly at night, thereby reducing the spatial detectability of SCI effects in surface temperature profiles. Similar dynamics were documented by Chow et al. [59] in Phoenix, where prevailing winds limited nighttime SCI extent through the horizontal advection and dilution of cool zones.

4.1.6. Nighttime SCI Behavior and Interpretation

Unlike the consistent daytime SCI patterns, nighttime cooling effects were more nuanced and often diminished, with some parks even exhibiting slightly higher LST than their surroundings, particularly in spring and autumn. This phenomenon can be attributed to several interrelated factors. First, evapotranspiration ceases after sunset, reducing the primary cooling mechanism active during the day [56]. Second, vegetated areas with moist soil and high thermal inertia tend to retain the heat accumulated during the day and release it gradually at night, potentially leading to localized warming [65,66].

Moreover, incomplete canopy development during transitional seasons like spring may result in exposed soil surfaces, which absorb more solar radiation during the day and slowly re-emit it at night [67,68]. In contrast, adjacent built-up areas may cool faster due to lower thermal mass, stronger radiative loss, or increased airflow through open urban canyons, especially in grid-planned cities like Mexicali [69].

Furthermore, the regression models showed weaker and less consistent relationships between vegetation metrics and nighttime ΔTmax, suggesting that vegetation structure alone is insufficient to guarantee cooling after dark. Instead, park layout, surrounding land use, and local wind conditions may play a more dominant role in nighttime thermal behavior [70,71].

These findings underscore the importance of integrating nocturnal thermal dynamics into climate-responsive design. Strategies such as increasing canopy density, using materials with lower thermal retention, and enhancing ventilation corridors should be considered to strengthen nighttime cooling in arid urban parks.

4.1.7. Null SCI Cases

A considerable proportion of parks in Mexicali exhibited no detectable SCI effect (Lmax = 0), with seasonal proportions ranging from 28 to 35% during the day and 19–44% at night. These null cases can be explained by several interacting factors. First, many of the sites classified as “parks” were included by typology rather than strictly by vegetation cover, and in fact, some parks showed no vegetation at all in both NDVI- and RF-based descriptors. Unsurprisingly, these barren spaces produced no measurable cooling effect. This aligns with the vegetation analysis, which showed strong positive correlations between SCI indicators and vegetation amount: NDVI-based vegetation proportion correlated up to r = 0.903 with ΔTmax in summer nights, while RF-based vegetation coverage explained much of the variability in Lmax across seasons. By contrast, fragmented or irregular vegetation structures (e.g., high FD or LSI) also weakened SCI detectability.

Overall, these findings confirm that null SCI outcomes are not methodological artifacts but rather reflect ecological, morphological, and typological constraints. The fact that some “parks” contained no vegetation underlines that the presence, quantity, and spatial configuration of vegetation are essential for SCI development.

4.2. Methodological Considerations for SCI Analysis

4.2.1. Limitations of the Radial Buffer and Isotropic Assumptions

The radial buffer approach used to extract SCI profiles assumes isotropic cooling, whereby the thermal influence of each park is expected to propagate uniformly in all directions. While this method enables standardized and scalable comparisons across sites, it does not fully account for contextual anisotropies, such as the presence of adjacent roads, water bodies, or other contrasting land covers, that may locally amplify or suppress LST patterns. Because some parks are bordered by sharply contrasting land uses, such as built-up areas and major roads on one side, and open, non-urbanized land on the other, uneven cooling patterns are likely to emerge across their boundaries. In such cases, the averaged radial profile may obscure distinct directional cooling or heating gradients, potentially leading to the under- or overestimation of SCI extent. For example, cooling may extend farther over vegetated or permeable surfaces and attenuate more quickly toward densely built or high-traffic areas.

This limitation has been noted in prior studies. For example, Wang et al. [72] and He et al. [73] demonstrated that urban LST exhibit strong directional anisotropy influenced by surface structure, wind, and solar angle. Similarly, Sun et al. [74] found that cooling distances vary with surrounding built context, while Jiang et al. [75] showed that cooling effects are enhanced when parks and riparian green spaces are positioned relative to wind direction and river morphology. Together, these studies reinforce the need to consider directional variability when evaluating SCI patterns.

Although the current method smooths out such heterogeneities by averaging over all azimuthal directions, future improvements could include directional SCI analysis, where radial LST profiles are extracted separately by quadrant or along specific land cover interfaces (e.g., using sector-based or polar transects). This would allow for a more nuanced assessment of spatial asymmetry in cooling effects and offer better insights into landscape-dependent LST dynamics.

An additional methodological consideration relates to the resampling of thermal data. To spatially align Landsat 30 m LST products with the 10 m resolution buffer system used in this study, LST values were resampled to 10 m using bilinear interpolation. While this step enhances geometric compatibility and allows finer-grained radial profiling, it introduces the risk of artificial smoothing and a false sense of spatial precision. It is important to note that Landsat’s native thermal resolution is 120 m (downscaled to 30 m by USGS through cubic convolution), and further resampling to 10 m may amplify spatial uncertainties and obscure small-scale thermal variation.

Acknowledging these limitations, the resampling method was applied uniformly across all observations to maintain internal consistency in comparative analysis. Prior studies support the feasibility of resampling Landsat-derived LST for urban climate assessments when properly constrained [76,77], particularly when the focus lies on general spatial trends rather than absolute pixel-level accuracy. While bilinear resampling improves dataset compatibility, future studies should incorporate real data validation using in situ measurements or high-resolution thermal imagery (e.g., ECOSTRESS, UAV-mounted sensors). Such validation would allow the quantification of resampling-induced error and enhance the precision and interpretability of spatial LST gradients, particularly when evaluating fine-scale SCI patterns.

Despite its simplicity, the radial buffer method remains technically valid for comparative assessments across heterogeneous urban contexts [14,53]; and the small proportion of manually corrected cases (<3%) suggests that the approach reliably captured general SCI patterns. However, recognizing and quantifying directional variability remains an important step for enhancing the ecological and planning relevance of SCI evaluations.

4.2.2. NDVI Threshold vs. RF Classification Approaches

This study employed two complementary methods to assess vegetation cover and structure: a simplified NDVI threshold approach and a supervised RF classification based on Sentinel-2 multispectral data. While both methods revealed significant associations with SCI indicators, their differences in complexity, interpretability, and sensitivity to landscape patterns merit comparison.

The NDVI threshold method offers a rapid, replicable way to distinguish vegetated from non-vegetated surfaces, using empirically established cutoff values. This approach allowed for a broad and consistent evaluation of green cover across parks. However, it is limited in its ability to differentiate between types of vegetation (e.g., trees vs. grass), and may misclassify areas with intermediate NDVI values, particularly in arid zones with sparse vegetation or mixed land covers. These limitations are consistent with findings in previous research, which noted that NDVI alone often fails to distinguish detailed vegetation structures, especially in semi-arid and urban landscapes [78].

In contrast, the RF method produced a more detailed land cover classification, distinguishing multiple surface categories and allowing for a nuanced analysis of vegetation configuration. This granularity was critical for calculating landscape metrics, such as LSI and FD, which showed significant correlations with SCI performance. RF’s effectiveness in integrating spectral and structural features has been demonstrated in similar applications, enhancing classification accuracy over simpler methods [78,79].

However, RF classification heavily depends on training data quality and spectral separability. Misclassification risks, especially among spectrally similar classes like dry soil, asphalt, and sparse shrubs, may introduce uncertainty into landscape metric calculations. Despite these trade-offs, the integration of both methods strengthened the study’s findings. NDVI thresholds provided consistency and replicability for cross-case comparisons, while RF classification contributed detail and ecological nuance.

4.2.3. Resolution and Integration Challenges

One of the principal methodological challenges in this study involved the integration of satellite data with different spatial resolutions and formats. Vegetation indicators were derived from Sentinel-2 data at 10 m resolution, while LST was extracted from Landsat 8–9 TIRS data with a native resolution of 100 m, resampled to 30 m. Additionally, SCI analysis required aligning these raster datasets with vector-based park polygons and buffer zones.

As observed in urban ecological studies, differences in pixel size can mask or distort ecological relationships, particularly in heterogeneous urban settings, as Tian et al. [80] demonstrated, coarse-resolution satellite imagery (e.g., 500 m MODIS) tends to amplify urban–rural differences in vegetation phenology due to mixed pixels, leading to misrepresented land-surface indicators; and Wang et al. [81] further confirmed that mixing high-resolution Sentinel-2 and Landsat data into coarser pixels significantly alters derived vegetation indices and phenological metrics, mischaracterizing urban ecological patterns, especially in fragmented green spaces.

Moreover, overlaying high-resolution vector polygons on coarser rasters can introduce positional errors at feature boundaries, especially when raster pixels span multiple land cover types. Collins and Dronova [82] showed how object-based image analysis (OBIA) on 30 m Landsat data improved classification accuracy of urban climate zones, highlighting that pixel boundaries often misrepresent true class extents. These mixed-pixel effects distort metrics like SCI near park edges.

The use of fixed-distance concentric buffers (e.g., 100–500 m) assumes isotropic cooling propagation, which oversimplifies real-world conditions influenced by wind direction, urban morphology, and surface albedo. These factors create spatially asymmetric cooling patterns that vary significantly across urban environments. For example, Jiang et al. [75] found that the cooling effectiveness of green spaces in Shanghai was highly dependent on their placement relative to prevailing winds and rivers, with greater cooling observed in leeward zones. Similarly, Kenawy et al. [83] demonstrated that surface albedo and urban morphology were key drivers of spatial SUHI patterns in Cairo, where reduced albedo and increased impervious surfaces led to stronger nighttime heat retention.

Future research should explore the integration of higher-resolution thermal data from sensors like ECOSTRESS or UAV platforms, which have demonstrated enhanced accuracy in capturing fine-scale thermal heterogeneity in urban and ecological environments [6,84]. Additionally, object-based image analysis (OBIA) has proven to be an effective method for urban mapping and classification from high-resolution imagery, reducing boundary uncertainties when linking spatial ecological patterns across resolutions [23].

5. Conclusions

In this study, the Surface Cool Island (SCI) effect was assessed across 435 urban parks in Mexicali, Mexico, using multi-seasonal remote sensing and spatial metrics. Two key SCI indicators, cooling intensity (ΔTmax) and spatial extent (Lmax), were calculated based on Landsat-derived land surface temperature (LST) and vegetation classification using both NDVI thresholding and Random Forest (RF) methods.

An average daytime SCI intensity of 0.81 °C was recorded, with a maximum ΔTmax of 6.41 °C observed during the summer season. In contrast, nighttime SCI effects were weaker, with an average ΔTmax of 0.43 °C and a maximum of 1.94 °C. The mean daytime Lmax was 120.15 m, while the nighttime mean Lmax was 47.85 m, with some parks exhibiting cooling extents of up to 500 m (day) and 430 m (night).

SCI variability was better explained by RF-derived vegetation metrics than by NDVI-based metrics. R² values of up to 0.93 for Lmax and 0.75 for ΔTmax were achieved using RF inputs, compared to 0.88 and 0.79, respectively, for NDVI-based models. These results highlight the improved performance of object-based vegetation classification methods in arid urban environments.

Parks characterized by greater internal vegetation cover (above 40%), compact shapes (LSI < 1.5), and low fragmentation (FD ≈ 1.35–1.45) were found to produce stronger and more extensive cooling effects. Conversely, up to 35% of parks exhibited no measurable SCI effect during certain seasons, often due to sparse vegetation (<30%), irregular geometries, or adjacency to impervious surfaces.

The spatial propagation of cooling was shown to be limited when vegetation was discontinuous or highly fragmented beyond park boundaries. In contrast, cohesive vegetative structures both within and around parks were associated with stronger SCI performance.

While robust results were obtained, certain methodological limitations should be acknowledged. The 30 m spatial resolution of Landsat imagery may constrain the detection of fine-scale thermal gradients, particularly in smaller or highly fragmented parks. In addition, classification accuracy may be affected by seasonal vegetation stress and the spectral similarity of arid surfaces to degraded vegetation. Future research should consider integrating higher-resolution thermal imagery, in situ temperature monitoring, and 3D vegetation structure data to improve model precision. Moreover, the nuanced behavior of nighttime SCI effects, such as cases where parks retained heat rather than cooled, should be further explored, particularly in relation to soil moisture retention, vegetation thermal inertia, and surface composition, to better understand post-sunset thermal dynamics in arid urban environments.

A scalable remote sensing framework was developed to support climate-adaptive green infrastructure planning in desert cities. By integrating multi-source satellite data and spatial metrics, this approach can be replicated in other arid urban contexts where thermal vulnerability is high, and vegetation is constrained.