The CUGA Method: A Reliable Framework for Identifying Public Urban Green Spaces in Metropolitan Regions

Abstract

This study addresses the challenge of reliably identifying Public Urban Green Spaces (PUGS) in metropolitan areas, a key requirement for advancing equitable access to green infrastructure and monitoring progress toward SDG 11.7 and WHO recommendations. In the absence of consistent local datasets, we propose the Candidate Urban Green Area (CUGA) method, which integrates OpenStreetMap and Copernicus Urban Atlas data through a structured, transparent workflow. The method applies spatial and functional filters to isolate green spaces that are publicly accessible, meet minimum size and usability criteria, and are embedded within the urban fabric. We validate CUGA in the Dublin Region using a stratified random sample of 1-ha cells and compare its performance against five alternative datasets. Results show that CUGA achieves the highest classification accuracy, spatial coverage, and statistical robustness across all counties, significantly outperforming administrative, crowdsourced, and satellite-derived sources. The method also delivers greater net spatial impact in terms of green area, catchment coverage, and residential land intercepted. These findings support the use of CUGA as a reliable and transferable tool for urban green space planning, policy evaluation, and sustainability reporting, particularly in data-scarce or fragmented governance contexts.

1. Introduction

1.1. Urban Green and Public Space: Background and Relevance

In 2015, the United Nations adopted the 2030 Agenda for Sustainable Development [1], which established 17 Sustainable Development Goals (SDGs). Within goal 11, “make cities and human settlements inclusive, safe, resilient and sustainable”, target 11.7 calls to “provide universal access to safe, inclusive and accessible, green and public spaces”. One year later, UN-Habitat adopted the New Urban Agenda [2], declaring its aim to “improve human health and wellbeing”, with multiple commitments regarding “green and quality public spaces” in cities and human settlements.

Simultaneously and accordingly, the World Health Organization Regional Office for Europe (WHO-ROE) highlighted the importance of providing access to urban green spaces (UGS) to the population [3,4]. According to this United Nations (UN) agency, it is key to recognize that: (a) UGS provide multiple benefits and constitute a necessary feature of healthy settlements; (b) their benefits can be maximized through adequate planning, design and evaluation; and (c) local authorities are responsible for their protection and maintenance [5].

UGS are widely recognized as vital for the health and liveability of cities [6]. They contribute significantly to the physical and mental well-being of urban populations, deliver crucial ecosystem services, and strengthen resilience to climate change [7]. Their importance is well recognized in the European Union’s (EU) Green Infrastructure Strategy [8], in the EU Biodiversity Strategy [9] and in the EU environmental strategy on spatial planning and land-use change [10].

As part of the Urban Agenda for the EU [11], the main objective of the Sustainable Land Use Partnership is “to ensure that the changes in urban areas (growing, shrinking and regeneration) are respectful of the environment, improving quality of life”, and one of the three approaches to achieve that goal includes “the protection of urban green areas to ensure more liveable conditions” [12]

Adequate provision of and accessibility to green spaces (GS) have been advocated for and promoted especially during the 21st century by national agencies in countries like the UK [13,14]. Currently, the WHO-ROE recommends that “residents should be able to access public green spaces of at least 0.5–1 hectare within 300 m’ linear distance (around 5 min’ walk) of their homes.” [5].

1.2. Urban Green Spaces, Equity and Accessibility: Policies, Planning and Research

UGS, however, are often distributed unequally. Access tends to be stratified along lines of income, race and ethnicity, age, gender, ability, and other social dimensions [15]. This disparity has increasingly been acknowledged as an issue of environmental justice, particularly as awareness has grown regarding the critical role of GS in supporting public health. A decade ago, scholars began calling for greater integration of environmental justice perspectives into the planning and governance of UGS in Europe [16]

UGS are key elements of green infrastructure (GI). Prior to the launch of the EU Green Infrastructure Strategy in 2013 [8], academic publications on green infrastructure were scarce; however, since then, the field has experienced exponential growth [17]. At the beginning of this decade, however, work was still limited on actual accessibility and most studies focused on quantifying GI availability [18]. Given the well-documented social and ecological heterogeneity of urban environments, identifying urban green infrastructure (UGI) elements at a relatively high spatial resolution is essential for integrating UGI accessibility into urban policy, planning, and management—enabling decisions and interventions to be effectively prioritized at the neighbourhood or parcel level [19]. The importance of interconnectivity among UGI elements also demands their accurate identification to properly address their management, transformation and future development [20].

Research on UGI accessibility has mainly focused on parks, examining how different socio-demographic groups experience this access and exploring the health implications of limited availability. Much of this work was originated in the United States, the United Kingdom, and Australia. However, in the last decade, there are an increasing number of studies in the EU context. They range from inter-European and national assessments [21,22,23,24] to more detailed studies on cities like Amsterdam, Barcelona, Bologna, Helsinki, Łódź, Rome, or Vitoria [25,26,27,28,29,30,31].

Although some cities adopted GI plans before the turn of the century, the introduction of the EU Green Infrastructure Strategy [8] has also driven a growing number of European cities to develop green infrastructure plans and public space strategies in which public urban green spaces (PUGS) play a central role. However, the evolution of GI planning has varied dramatically between European nations [32], and it is safe to assume that today most EU cities do not have one nor a proper geospatial record of their components. The situation is worse when addressing planning or datasets for functional urban areas (FUA) involving multiple administrations—which is usually the case.

1.3. Public Urban Green Spaces (PUGS) Identification: Challenges and Research Objective

Ensuring equitable access to UGS is best achieved by providing cities with an adequate supply of PUGS. To inform planning decisions related to PUGS and housing distribution, it is essential to assess their current provision and accessibility. This, in turn, requires accurate geospatial identification and documentation—an inherently challenging task in the absence of local data or on-site verification, particularly when it comes to determining whether these spaces are truly open and accessible to the public [33,34].

In such contexts, UN-Habitat has advocated for the use of open-source, generic datasets when more reliable or detailed maps are unavailable. Technical documents specifically reference OpenStreetMap (OSM) [35] and satellite-based Earth observation data—in Europe, generally the Copernicus Urban Atlas (CUA) [36] and other Landsat products—to identify potential public urban open spaces [37]. While these sources are widely used in the absence of more precise geospatial data, their reliability for identifying PUGS remains limited.

Low-to-medium resolution imagery (10–30 m) offers useful insights for urban planning [38], but it often underestimates small GS [39,40] and performs poorly in complex urban environments—e.g., streets under 10 m wide are typically undetected [41,42,43]. High-resolution data (finer than 10 m) are more suitable for mapping urban open spaces [40] and are commonly used in SDG 11.7.1 monitoring. However, it is costly and less feasible for large-scale applications, particularly in cloud-prone regions [44]. Moreover, even with high-resolution imagery, distinguishing which UGS are genuinely open and accessible to the public remains a challenge [45]—i.e., identifying PUGS.

In Europe, research involving PUGS mostly relies on thematic datasets provided by local administrations [25,26,31,46,47], the ‘green urban areas’ (GUA) class from the CUA [23,30,48,49,50], or the ‘park’ and park-related elements from OSM [24,34,51]. However, each one presents various limitations that are discussed in the following section.

This article presents a novel and more reliable method for identifying PUGS by combining the two most widely used open-access geospatial datasets—OSM and CUA. With the former being available globally and the latter covering most European Functional Urban Areas (FUAs) with populations over 50,000, this approach is potentially applicable across a broad range of urban contexts throughout Europe. Its improved reliability is supported by a verification system based on the random sampling of 1-hectare cells in the Dublin Region (DR), providing a robust foundation for broader application. However, further testing in other European locations with different urban and peri-urban form and vegetation characteristics may be required to fully assess its generalizability.

2. Materials and Methods

2.1. PUGS Identification: Limitations of Administration, CUA and OSM Databases

The use of thematic cartography provided by local administrations presents at least three major limitations, as will be demonstrated in our own case study: (a) many administrations do not have this type of digital mapping available; (b) when such data exist, they typically include only the main parks or PUGS managed by that specific administration, omitting other relevant PUGS—or conversely, they include all areas managed by the parks and gardens department without distinguishing which ones actually qualify as PUGS; and (c) in geographic areas involving multiple local administrations—as is the case in most Functional Urban Areas (FUAs)—there is no consistency or harmonization between datasets.

CUA is a highly valuable high-resolution dataset, and the use of its GUA class to identify PUGS is widespread. However, its reliability is limited due to at least three issues inherent to its production method [52]: (a) the declared accuracy is ≥0.85, although the available documentation lacks sufficient detail to fully understand the implications of this figure; (b) the resolution and precision (Minimum Mapping Unit: 0.25 ha; Minimum Mapping Width: 10 m) may be insufficient in certain contexts—for example, in vegetated areas narrower than 10 m along coastlines or water bodies, or when conducting pedestrian network accessibility analyses; and (c) the classification criteria may not align with those used to identify PUGS. For instance, the GUA class includes areas such as zoos, castle parks, and cemeteries, which—due to their nature or restricted access—do not qualify as PUGS. Conversely, patches of natural vegetation or agricultural land between 0.25 ha and 1 ha, enclosed by built-up areas and not managed as urban green areas, are classified as ‘land without current use’, even though they may function as PUGS.

OSM offers very high-resolution mapping, sometimes based on official cartography when licensing allows. Its spatial features are supported by an extensive alphanumeric tagging infrastructure with well-defined classification criteria. However, the most common practice when using OSM to identify PUGS is to rely solely on polygons tagged as ‘park’ under the ‘leisure’ key. As we will demonstrate, this criterion is overly restrictive and excludes many areas that do qualify as PUGS. This is a relevant issue considering that informal GS can provide comparable ecosystem services to cultivated urban parks [53].

2.2. PUGS Identification: Previous Methods Based on OSM and CUA

The first research article we have encountered that used OSM tags to identify PUGS was applied in Brussels and published in 2018. Le Texier et al. defined a two-step inclusion criterion: first, they selected UGS based on values in the ‘landuse’ (forest, grass, greenfield, meadow, orchard, recreation_ground, village_green, or vineyard), ‘leisure’ (common, dog_park, garden, golf_course, nature_reserve, park, or pitch), and ‘natural’ (fell, grassland, heath, moor, scrub, or wood) keys; and second, they filtered for PUGS using the ‘access’ key with values ‘yes’, ‘public’, or left empty [54].

A more advanced approach to assessing accessibility (public vs. private) expanded the list of potential UGS by including additional values in the ‘landuse’ (cemetery), ‘tourism’ (camp_site), and ‘amenity’ (grave_yard) keys. This method also refined the accessibility classification by using both the ‘access’ and ‘fee’ keys. Based on the ‘access’ key values, areas were considered non-public (no, private, official, military, agricultural, forestry, destination, restricted, delivery), public (yes, permissive, or public), or of unknown status (unknown). In terms of cost, based on the ‘fee’ key, areas were considered freely accessible (no) or paid (yes, a specific cost or interval) [33,55].

An alternative approach integrates OSM data with the CUA and the Copernicus Street Tree Layer (STL) to identify PUGS based on their physical accessibility [56]. This method classifies the following green elements as highly accessible UGS—i.e., PUGS: forests (CUA forest); urban forestry areas not formally classified as forests (CUA_GUA intersected with STL); GUAs (CUA_GUA); tree-covered areas within urban parks (CUA_GUA intersected with OSM parks); non-tree greenery within urban parks (CUA_GUA intersected with OSM parks); and trees located along roads and transport corridors. The classification relies primarily on the CUA_GUA layer, complemented by very limited OSM features—parks (the only ones considered to have high physical accessibility), and cemeteries and allotment gardens (classified as having medium physical accessibility).

Despite the increasing sophistication of these approaches, the inclusion and exclusion criteria for identifying potential PUGS remain questionable. This applies both to the specific values used in the ‘landuse’, ‘leisure’, ‘natural’, ‘tourism’, and ‘amenity’ keys, and to the interpretation of accessibility based on the ‘access’ and ‘fee’ tags. Regarding the former, from our perspective, limiting the selection to parks is overly restrictive, while the inclusion of features such as zoos, camp_sites, and golf_courses appears unnecessary, and others remain debatable. As for the latter, the assumption that a missing ‘access’ tag implies public availability is problematic and largely unverified. In addition to these methodological concerns, none of the reviewed studies provide a clear verification or testing protocol, which would be essential for their broader adoption

2.3. PUGS: Definition, Inclusion and Accessibility Criteria

There are no universally accepted definitions of GS, UGS, or PUGS [56], nor are there precise criteria regarding their size, vegetation cover, public or private status, or potential uses. Therefore, it is important to clarify how UGS and PUGS are defined in the context of any related study [57], including this one.

For UGS, we adopt the definition proposed by the WHO-ROE, which includes “all urban land covered by vegetation of any kind […] on private and public grounds, irrespective of size and function, and can also include small water bodies such as ponds, lakes or streams” [5].

The definition of PUGS is derived by applying a set of restrictions to the broader category of UGS, in line with the WHO-ROE’s recommendation that “residents should be able to access public green spaces of at least 0.5–1 ha” [5]. Regarding the minimum area requirement, we include only those green spaces that are at least 0.5 ha in size, free of sports pitches, and where it is possible to inscribe a circle with a diameter of at least 30 m, ensuring a minimum level of spatial functionality, as proposed by different local administrations.

In the absence of a more precise specification from the WHO regarding what constitutes ‘public’ we adopt the criteria proposed by UN-Habitat for defining public urban space: “accessible to the public without charge, and provides recreational areas for residents and helps to enhance the beauty and environmental quality of neighbourhoods.” This includes green public areas, riparian reserves, parks and urban forests, community gardens, and pocket parks, among others [58].

It is important to note that this definition differs in some respects from the classification of GUA used in the CUA, which includes “public green areas for predominantly recreational use such as gardens, playgrounds, zoos, parks, castle parks and cemeteries”; “suburban natural areas that have become and are managed as urban parks”; and “forests or green areas extending from the surroundings into urban areas when at least two sides are bordered by urban areas and structures, and traces of recreational use are visible.” However, private gardens within housing areas and patches of natural vegetation or agricultural land between 0.25 ha and 1 ha, enclosed by built-up areas and not managed as GUAs, are not considered GUAs [52].

We also adopt WHO-ROE’s recommendation of a 300-m linear distance as the catchment area for PUGS [5]. Since the publication of this guideline, the 300-m threshold has gradually become a widely accepted planning benchmark and is frequently used in UGS accessibility research across Europe [59]. It is also incorporated into the increasingly popular 3-30-300 rule for urban forestry, which advocates that every resident should be able to see at least three trees from their home, live in a neighbourhood with a minimum of 30% tree canopy cover, and be located within 300 m of a park or green space of at least 1 ha [60].

2.4. Study Area

This study focuses on the Dublin Region (DR), which comprises four administrative counties: Dublin City (DC), South Dublin (SD), Fingal (F), and Dún Laoghaire–Rathdown (DLR). This configuration allows for the examination of similarities and disparities in thematic datasets across neighbouring administrations. Notably, Dublin City has shown particular interest in its PUGS, as reflected in the publication of the Dublin City Parks Strategy in 2019 [61]. The city was also accepted as 1 of 16 cities bidding for the 2023 European Green Capital Award, although it was not selected as a finalist.

The region represents a particularly relevant case for identifying PUGS due to its distinctive urban morphology—with a significant proportion of low-density residential development, particularly in suburban areas—and diverse green infrastructure—where private gardens are a common feature.

According to a comprehensive ranking of the ecosystem services in the EU including all metropolitan areas with more than 100,000 inhabitants [21], the DR ended in a middle position (167 out of 305). Compared to all those EU urban regions, the DR had medium percentage of urban green spaces (2.60%), a low percentage of impervious surfaces (9%), and very low percentages of natural areas (10%) and protected areas (11.3%).

This spatial configuration and land use distribution make the DR a challenging setting to explore the distinction between publicly accessible green areas and private green spaces, a distinction that is often blurred in remote sensing and land cover datasets.

2.5. Data Sources

This study uses geospatial data from the following sources:

• Administrative datasets by the four counties in the DR providing official records of parks and other public spaces, mostly PUGS but not exclusively—namely DCC Parks and Open Spaces 2016 [62], Parks SDCC [63], Local/National Parks and Play Grounds FCC 2023 [64], and Main Parks DLR [65]
• CUA 2018 [36] provided by the EU Copernicus Land Monitoring Service offering harmonized land use and land cover data for 17 classes with the Minimum Mapping Unit (MMU) of 0.25 ha in urban areas (10 classes with the MMU of 1 ha in rural areas) for EU functional urban areas with more than 50,000 inhabitants, for the 2018 reference year in EEA38 countries (EU, EFTA, Western Balkans and Türkiye) and the UK. CUA 2018 was published in 2020 (reviewed in 2021) and was the first update of the CUA 2012. CUA is planned to be updated every 6 years, so the 2024 edition is expected to be published soon. CUA, like all Copernicus Land Monitoring Service products, is freely and openly accessible, provided users credit the source, disclose any modifications, and avoid implying EU endorsement.
• OSM [35] datasets derived from open crowdsourced geospatial data worldwide and offering a flexible and detailed dataset of urban land use and other relevant cartographic information. OSM features are labelled using a flexible tagging system, where each element is described by key–value pairs that define its type and attributes. For example, green spaces may be tagged as leisure = park, landuse = forest, or natural = grassland. This structure enables detailed classification but requires careful interpretation due to its community-driven nature. It is supposed to be regularly updated but accuracy is not guaranteed. OSM data are openly available under the Open Database License, allowing free use, adaptation, and redistribution, provided that proper attribution is given and any derivative works are shared under the same terms.

Besides these three sources, Google Earth satellite imagery [66], Google Maps information [67] and Google Maps Streetview imagery [68] were used for visual inspection and virtual exploration regarding urban green areas.

2.6. The Candidate Urban Green Area (CUGA) Method for PUGS Identification

This section describes the method developed to identify potential PUGS —here referred to as CUGAs—using open-access geospatial data. The approach integrates land use and land cover information from the CUA dataset and OSM, applying a structured sequence of spatial operations to isolate public green spaces with potential accessibility for the urban population (Figure 1).

2.6.1. Datasets and Preprocessing

The CUGA method is based on CUA [36] and OSM [35] datasets. From CUA, the following land classes are employed: residential urban fabric (classes 11100 to 11240), land transport infrastructure (12210 to 12230), GUAs (14100), and sports and leisure facilities (14200) —further information on the CUA dataset can be found in the Mapping Guide v6.3 for a European Urban Atlas [52].

From OSM, spatial features are extracted using QuickOSM [69], to create the following thematic layers:

• OSM_green_raw, containing polygons with potential to be PUGS, including the following key values: ‘landuse’ (basin, forest, grass, greenfield, meadow, recreation_ground, village_green), ‘leisure’ (common, garden, nature_reserve, park, pitch), and ‘natural’ (beach, grassland, heath, scrub, shrub, shrubbery, tree_row, water, wood);
• OSM_notPUGS, containing polygons with land uses that are GS but are excluded from PUGS classification, including the following key values: ‘landuse’ animal_keeping, cemetery, education, farmland, farmyard, greenhouse_horticulture, orchard, plant_nursery, vineyard, religious), ‘leisure’ (golf_course, horse_riding, marina, sports_centre, stadium); ‘amenities’ (animal_boarding, animal_breeding, animal_shelter, arts_centre, childcare, clinic, college, conference_centre, courthouse, embassy, events_venue, fire_station, grave_yard, hospital, monastery, nursing_home, parking, police, prep_school, prison, school, shelter, theatre, townhall, vehicle_inspecion, waste_transfer_station, youth_centre); ‘tourism’ (camp_site, caravan_site, theme_park, zoo);
• OSM_pitch, containing sports fields, ‘leisure’ (pitch);
• OSM_coastline, representing coastal line features, i.e., ‘natural’ (coastline);
• OSM_pedestrian, including pedestrian-related road segments, ‘highway’ (footway, path, pedestrian, residential, living_street).

The OSM_green_raw layer is clipped to the DR boundary. Polygons from OSM_notPUGS are subtracted, multipart geometries are converted to single parts, and polygons smaller than 50 m² are removed. The resulting layer, OSM_green_pro, serves as the base for the detection of CUGAs.

2.6.2. CUGAs Identification

The identification process is structured into five sequential selection steps. In each step, polygons meeting specific criteria are saved as a new CUGA layer. At the end of each step, all polygons of any size in OSM_green_pro matching the corresponding ‘leisure’ or ‘landuse’ key values and their nested geometries are tagged accordingly —e.g., park, garden, nature_reserve—to prevent re-selection in subsequent steps:

• Parks (CUGA_1_prk): Polygons with ‘leisure’ (park) and an area ≥ 0.5 ha are selected and saved as CUGA_1_prk.
• Gardens (CUGA__2_grd): From the untagged polygons, those with ‘leisure’ (garden), an area ≥ 0.5 ha, and access set to permissive or yes are selected and saved as CUGA_2_grd.
• Nature Reserves (CUGA_3_ntr): From the remaining untagged polygons, those with ‘leisure’ (nature_reserve), an area ≥ 0.5 ha, and intersecting the pedestrian network—suggesting greater likelihood of public foot access—are selected and saved as CUGA_3ntr.
• Recreation Grounds and Village Greens (CUGA_4_rgvg): Polygons with ‘landuse’ (recreation_ground, village_green), and an area ≥ 0.5 ha, are selected. CUA ‘other roads and associated land’ (12220) are subtracted —since some selected OSM features inaccurately extend over road areas—, multipart geometries are split, and polygons ≤ 0.5 ha are removed. The result is saved as CUGA_4_rgvg.
• Grass, Forest, Grassland, and Wood (CUGA_5_gfgw): From the remaining untagged polygons, those with ‘landuse’ (grass, forest), or ‘natural’ (grassland, wood), are selected. Polygons contained within others or overlapping CUA land transport infrastructure are removed. The following spatial relationships (SR) are then investigated to identify whether each polygon: (SR_a) intersects CUA GUAs; (SR_b) is contained in CUA GUAs; (SR_c) intersects the pedestrian network; (SR_d) is within 10 m of residential street segments; (SR_e) is surrounded by residential street segments; (SR_f) intersects with CUA dense residential fabric (classes 11100, 11210, 11220); or (SR_g) is within 25 m of the coastline. Polygons meeting any of these conditions are saved as CUGA_5-1 with registered binary information about their positive or negative spatial relationships.
• Then, a new filter is applied to CUGA_5-1 to select those polygons meeting any of the following combinations of the previously registered SR: SR_a, SR_c and SR_d; SR_a, SR_c and SR_f; SR_a, SR_d and SR_f; SR_b and SR_c; SR_e and ≥0.5 ha; SR_g. As can be inferred, this set of spatial conditions is designed to identify GUAs that are genuinely open and accessible to the public.
• The resulting polygons are processed with the CP6m script, which merges polygons closer than 6 m and removes narrow slivers. The output is saved as CUGA_5-2. From CUGA_5-2, polygons intersecting the pedestrian network, the 10 m residential buffer, or the CUA dense residential fabric are selected and intersected with CUGA_5-1. The result is saved as CUGA_5gfgw. CUA sports and leisure facilities and previously defined CUGAs (i.e., layers CUGA_1 to CUGA_4) are spatially subtracted. The CP6m script is applied again, followed by subtraction of the CUA transport infrastructure and removal of polygons < 0.5 ha. The result is saved as an updated CUGA_5_gfgw.

2.6.3. CUGAs Consolidation

All five CUGA layers are merged into a single dataset and the presence of sports fields is assessed for each CUGA polygon. A new attribute records whether the polygon contains sports fields but cannot accommodate a circle with a diameter of 30 m (‘pitch&C30′). Additionally, the proportion of area occupied by sports fields is calculated and stored (‘sport_ratio’). A final filtering step saves those polygons for which ‘pitch&C30′ is null and that are either tagged as parks or have a ‘sport_ratio’ ≤ 0.25 as CUGAs.

2.7. Datasets and Validation Design

2.7.1. Datasets: Initial Comparison via Spatial Visualization and Descriptive Statistics

This study evaluates the performance of the proposed CUGA method alongside five other datasets derived from county administrations, OSM, and CUA, all filtered not to include polygons <0.5 ha:

• Adm_prk: includes designated public parks from the datasets provided by each county.
• Adm_all: a broader dataset encompassing parks and other publicly accessible outdoor spaces, such as sports fields and recreational grounds. It includes all elements of Adm_prk.
• OSM_prk: contains OSM polygons tagged as parks.
• OSM_grn: an extensive dataset aiming to include potential PUGS (OSM_grn_raw). It includes all elements of OSM_prk.
• CUA_GUA: land classified as GUA by the CUA.
• CUGA: resulting from the proposed CUGA method

Before conducting the performance evaluation, we carried out a preliminary comparison of the datasets using the following materials: (a) different maps of the entire DR, including detailed insets for each county at closer scales, illustrating the spatial agreement and disagreement among datasets; and (b) basic descriptive statistics summarizing the extent and characteristics of the polygon areas (in ha) for each dataset, and the whole DR.

2.7.2. Urban Scope and Grid-Based Random Sampling

To focus on the overall motivation of the research—population accessibility to urban green areas—we defined an urban scope by applying a 300-m buffer around CUA residential urban land. This was used to delineate the urbanized area of interest, ensuring that the validation focused on zones where population density and potential demand for PUGS are relevant.

To test the performance of the proposed CUGA method, a 100 × 100 m grid was overlaid across the entire DR, resulting in 184,620 cells. Of these, 52,987 cells intersected the urban buffer and were retained for study population. From this urban subset, a random sample of 382 cells was selected for validation, so its size is statistically significant at a 95% confidence level with a 5% margin of error, based on the standard formula for finite population correction.

The spatial distribution of the 382 one-hectare cells across the four counties of the DR demonstrates a high degree of representativeness with respect to three key land use variables: total urban area, residential urban area, and areas of special interest. The sample size and distribution reflect the spatial distribution of urbanization and research focus across the region and ensure representativeness across the urbanized area and sufficient statistical power for a comparative analysis between counties (

Table 1. Distribution of randomly selected 1-hectare cells across the four counties and the whole DR in relation to urban area, residential urban area, and urban scope area.

County	Validation Cells [n]	Urban Area [km2]	Residential Urban Area [km2]	Urban Scope Area [km2]
Dublin City (DC)	99 (25.9%)	91.69 (24.9%)	51.06 (29.4%)	115.11 (22.8%)
Fingal (F)	129 (33.8%)	127.15 (34.5%)	49.42 (28.4%)	183.95 (36.4%)
Dún Laoghaire-Rathdown (DLR)	68 (17.8%)	65.84 (17.9%)	39.28 (22.6%)	89.05 (17.6%)
South Dublin (SD)	86 (22.5%)	83.65 (22.7%)	34.19 (19.7%)	116.86 (23.1%)
Dublin Region (DR)	382 (100.0%)	368.33 (100.0%)	173.95 (100.0%)	504.96 (100.0%)

).

2.7.3. Validation

To assess the effectiveness of the proposed CUGA method and each dataset in identifying actual PUGS, we developed a validation framework based on spatial sampling and manual ground-truthing. Each sampled 1-ha cell was examined to determine whether it contained a true PUGS using the following sources: (a) Google Maps (satellite imagery, Street View, and other available information); (b) additional online resources (e.g., local authority websites, community-generated maps); and (c) site visits, when previous methods were inconclusive. The outcome of this inspection served as the ground truth against which all datasets were evaluated (PUGS).

A validation table was constructed, providing binary indicators for the following variables:(1–6) intersection with each dataset (Adm_prk, Adm_all, OSM_prk, OSM_grn, CUA_GUA, CUGA); (7) ground truth label (PUGS); and (8) county in which the cell is located (County)

This table forms the basis for all subsequent statistical analyses, including performance metrics and significance testing.

2.8. Performance Evaluation and Metrics

To assess the effectiveness of each dataset in identifying PUGS, we employed a comprehensive set of classification metrics. These metrics were calculated by comparing the binary classification results of each dataset—indicating whether a cell intersects a potential PUGS polygon of each dataset—with the ground truth, establishing the number of true positives (TP), true negatives (TN), false positives (FP) and false negatives (FN).

The following metrics are expressed as indices ranging from 0 to 1, where higher values indicate better performance:

• Accuracy measures the overall proportion of correctly classified cells. While intuitive, it can be misleading in imbalanced datasets, where the majority class dominates the metric. Therefore, it is reported for completeness but not used as a primary evaluation criterion.

$$Accuracy={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}$$

• Precision quantifies the proportion of potential PUGS of each dataset that are correctly identified. This metric penalizes false positives and is particularly relevant when overestimating green space accessibility could lead to misleading conclusions in planning or policy.

$$Precision={\displaystyle \frac{TP}{TP+FP}}$$

• Recall (or sensitivity) captures the proportion of actual PUGS that are correctly identified by the dataset. It penalizes FN and is essential to ensure that areas with real GS access are not overlooked.

$$Recall={\displaystyle \frac{TP}{TP+FN}}$$

• Specificity, or true negative rate, complements recall by measuring the proportion of non-PUGS cells that are correctly classified. This metric is important for evaluating the model’s ability to avoid overestimating GS presence.

$$Specificity={\displaystyle \frac{TN}{TP+FP}}$$

• F1-score is the harmonic mean of precision and recall. It balances both types of error and is particularly useful in the presence of class imbalance. In this study, F1-score is used as the primary metric for ranking datasets.

$$F1score={\displaystyle \frac{2·Precision·Recall}{Precision+Recall}}$$

• Matthews Correlation Coefficient (MCC) provides a balanced measure that considers all four elements of the confusion matrix. It is regarded as one of the most informative single-value metrics for binary classification, especially under class imbalance.

$$MCC={\displaystyle \frac{TP·TN-FP·FN}{\sqrt{(TP+FP)(TP+FN)(TN+FP)(TN+FN)}}}$$

• Balanced Accuracy (BA) is the average of recall and specificity. This metric provides a more equitable assessment of performance across both classes, mitigating the bias introduced by class imbalance.

$$Balanced Accuracy={\displaystyle \frac{Sensitivity+Specificity}{2}}$$

• Jaccard Index (JI), or Intersection over Union, measures the overlap between predicted and actual UGAs. It is a stricter metric than F1-score and is widely used in spatial analysis to evaluate the similarity between predicted and reference areas.

$$Jaccard Index={\displaystyle \frac{TP}{TP+FP+FN}}$$

• Cohen’s Kappa (κ) quantifies the agreement between the dataset and the ground truth, correcting for agreement expected by chance. This metric is particularly relevant in validation contexts, as it reflects the reliability of the classification beyond random coincidence.

$$\kappa ={\displaystyle \frac{P_o-P_e}{1-P_e}}$$

• where P_o is the observed agreement and P_e is the expected agreement by chance.

All metrics were computed for each dataset at two levels: individually for each of the four counties and for the whole DR. This dual-level analysis enables the assessment of both local and regional performance consistency and robustness.

2.9. Performance Statistical Significance Testing

Following the computation of standard validation metrics, we conducted a statistical test to assess whether the observed superiority of the CUGA method over alternative datasets was statistically significant. For this purpose, we employed McNemar’s test, a non-parametric method specifically designed to compare the performance of two classifiers based on paired binary outcomes.

McNemar’s test is particularly appropriate in this context because it evaluates whether two classification methods differ significantly in their error patterns when applied to the same set of spatial units (i.e., the sampled 1-ha cells). Unlike global metrics that summarize performance, McNemar’s test focuses on discordant cases—those where one method is correct and the other is not—thus providing a direct statistical comparison of classification behaviour.

Let b represent the number of cases where CUGA correctly identifies a UGA and the comparison method does not, and c be the number of cases where the comparison method is correct and CUGA is not. The test evaluates the null hypothesis that both methods have the same probability of being correct (i.e., b = c.) The test statistic is based solely on these discordant pairs and is computed using the exact binomial version of McNemar’s test when the total number of discordant cases (b + c) is small (typically < 25), which ensures robustness in small samples.

We applied McNemar’s test to compare CUGA against each of the other datasets (Adm_prk, Adm_all, OSM_prk, OSM_grn, CUA_GUA) and the vegetation presence, both for each of the four counties and the entire DR. This analysis goes beyond descriptive performance metrics and formally tests whether the observed improvements in classification accuracy by the CUGA method are statistically meaningful.

2.10. Net Spatial Impact: Analysis and Statistical Testing

We assessed the spatial impact of each dataset by analysing three key dimensions:

• the net area correctly identified as Urban Green Area (UGA);
• its associated Individual Catchment Area (ICA), defined as the 300-m buffer surrounding the UGA perimeter; and
• the Intercepted Residential Land (IRL) within that catchment, calculated as the sum of the areas of intersected polygons from the CUA residential land layer.

For each dataset and spatial unit (i.e., each county and the entire DR), we computed the net area as the difference between the total area of true positives (TP) and false positives (FP) for each of the three spatial dimensions (UGA, ICA, IRL). This approach enables us to quantify not only classification accuracy but also the effective spatial benefit delivered by each method.

To statistically assess whether differences in net spatial impact between datasets were significant, we applied the Kruskal–Wallis test, a non-parametric method suitable for comparing multiple independent groups without assuming normality. This test was conducted separately for each area type (UGA, ICA, IRL), both at the county level and for the entire DR. When the Kruskal–Wallis test indicated statistically significant differences (p < 0.05), we performed post hoc pairwise comparisons using the Mann–Whitney U test.

2.11. Tools and Software

Spatial and statistical analyses were conducted using QGIS [70] and jamovi [71], respectively, ensuring methodological transparency and reproducibility throughout the study.

3. Results

3.1. Preliminary Dataset Comparison: Visualization and Basic Statistics

The initial comparison of datasets was conducted through spatial visualization and descriptive statistical analysis to assess the extent and distribution of potential PUGS across the DR. Figure 2 and Figure 3 present regional-scale and county-level maps illustrating the spatial coverage of six datasets: two from local administrations (Adm_prk and Adm_all), two derived from OSM (OSM_prk and OSM_grn), the GUA class from the CUA (CUA_GUA), and the proposed CUGA Method dataset.

Figure 2 shows that Adm_prk and Adm_all are limited in extent and tend to focus on formally designated parks and recreational areas, while also including others that do not serve as PUGS. These datasets present notable gaps in suburban and peri-urban zones, especially in the two counties with short datasets (F and DLR). CUA_GUA provides broader coverage but lacks fine detail, particularly in narrow or fragmented urban spaces, and includes areas that do not serve as PUGS, e.g., cemeteries. OSM_prk and OSM_grn offer more extensive mapping, especially in suburban areas, but differ in selectivity: OSM_prk is highly restrictive, while OSM_grn includes a wider range of green land uses, potentially introducing noise from non-public or ambiguous spaces. The county insets make these differences more apparent, revealing inconsistencies in coverage and classification across administrative boundaries.

Figure 3 focuses on the CUGA dataset. As expected, given the methodology used to generate this dataset, the map suggests a coverage similar to CUA_GUA, but with visually appreciable modifications. CUGA is more extensive than OSM_prk (which it almost entirely includes) but much more limited than OSM_grn, from which it clearly excludes green areas located far from residential land. Notably, CUGA captures numerous mid-sized and irregularly shaped green areas that are absent from other datasets—especially Admin_prk, CUA_GUA and OSM_prk—particularly in residential and mixed-use zones. The insets suggest that CUGA captures potential PUGS in both dense urban cores and low-density suburban areas, advancing an improved spatial equity in PUGS identification.

Descriptive statistics (

Table 2. Descriptive statistics for the area (ha) variable of each dataset for each county and the whole DR, including results of the Shapiro-Wilk test showing a lack of normality when p ≤ 0.05.

						Percentiles					Shapiro-Wilk
County	Dataset	n	Mean	SD	Min	p25	p50	p75	Max	Sum	W	p
DC	Adm_all	206	9.7	53.4	0.5	0.9	1.8	5.4	678.8	2005.8	0.130	<0.001
	Adm_prk	60	26.7	97.2	0.5	2.5	6.7	13.3	678.8	1602.4	0.248	<0.001
	OSM_grn	530	4.8	31.2	0.5	0.7	1.1	2.3	649.1	2524.8	0.090	<0.001
	OSM_prk	148	9.5	55.0	0.5	0.8	1.6	6.5	660.8	1406.5	0.117	<0.001
	CUA_GUA	329	4.4	12.0	0.5	0.8	1.2	2.8	157.7	1442.2	0.305	<0.001
	CUGA	263	7.0	44.0	0.5	0.8	1.2	3.1	660.8	1837.9	0.105	<0.001
DLR	Adm_all	16	22.4	21.9	1.1	7.9	13.0	35.0	84.0	358.5	0.822	0.005
	Adm_prk	15	23.8	21.9	1.6	8.3	13.4	36.4	84.0	357.4	0.823	0.007
	OSM_grn	677	5.7	25.3	0.5	0.7	1.2	3.0	470.6	3851.3	0.162	<0.001
	OSM_prk	93	5.6	11.2	0.5	0.9	1.9	4.6	81.5	517.2	0.453	<0.001
	CUA_GUA	214	3.2	6.5	0.5	0.8	1.2	2.6	51.3	678.3	0.400	<0.001
	CUGA	199	3.9	8.6	0.5	0.7	1.2	3.0	81.5	766.2	0.385	<0.001
F	Adm_all	35	16.3	30.1	0.6	1.2	3.6	14.0	129.1	569.6	0.575	<0.001
	Adm_prk	33	17.1	30.8	0.6	1.2	3.6	14.2	129.1	564.3	0.591	<0.001
	OSM_grn	2078	3.2	7.4	0.5	0.9	1.6	3.1	129.1	6556.7	0.278	<0.001
	OSM_prk	120	9.3	19.8	0.5	0.9	2.4	6.1	129.1	1113.0	0.475	<0.001
	CUA_GUA	302	3.4	8.5	0.5	0.7	1.2	2.1	73.5	1030.2	0.333	<0.001
	CUGA	311	5.2	13.2	0.5	0.7	1.4	3.5	129.1	1623.5	0.362	<0.001
SD	Adm_all	326	4.5	12.3	0.5	0.7	1.2	2.7	118.2	1452.3	0.313	<0.001
	Adm_prk	73	12.5	18.8	0.6	3.1	6.7	13.1	118.2	911.6	0.568	<0.001
	OSM_grn	1424	5.5	35.6	0.5	0.8	1.3	2.9	1218.0	7766.6	0.083	<0.001
	OSM_prk	83	11.5	20.2	0.5	1.6	4.4	11.5	130.2	952.8	0.535	<0.001
	CUA_GUA	316	3.3	8.5	0.5	0.7	1.2	2.5	118.0	1038.9	0.284	<0.001
	CUGA	301	4.3	11.6	0.5	0.7	1.2	2.5	130.2	1279.9	0.315	<0.001
DR	Adm_all	583	7.5	34.2	0.5	0.8	1.4	4.4	678.8	4386.2	0.159	<0.001
	Adm_prk	181	19.0	59.0	0.5	2.6	6.7	14.2	678.8	3435.6	0.257	<0.001
	OSM_grn	4709	4.4	24.7	0.5	0.8	1.4	2.9	1218.0	20,699.3	0.093	<0.001
	OSM_prk	444	9.0	34.8	0.5	0.9	2.2	7.0	660.8	3989.5	0.185	<0.001
	CUA_GUA	1161	3.6	9.3	0.5	0.7	1.2	2.5	157.7	4189.7	0.310	<0.001
	CUGA	1074	5.1	24.0	0.5	0.7	1.2	3.1	660.8	5507.5	0.138	<0.001

) further support these observations. Across all counties, CUGA exhibits a balanced distribution of polygon sizes, with median values typically between 1.2 and 1.4 ha, and a relatively low standard deviation compared to OSM_grn or Adm_prk. The Shapiro–Wilk test confirms non-normality in the area distributions for all datasets (p < 0.001), justifying the use of non-parametric methods in subsequent analyses. Notably, CUGA consistently captures a higher total area of potential PUGS than Adm_prk and CUA_GUA, while avoiding the already-known overestimation of OSM_grn.

These preliminary findings suggest that the CUGA method offers a more robust and spatially equitable identification of PUGS across the DR. The next section evaluates the classification performance of each dataset against ground-truth data to validate these preliminary observations.

3.2. PUGS Identification Performance Across Datasets

To evaluate the effectiveness of each dataset in identifying PUGS, we compared their classifications against ground-truth data derived from a stratified random sample of 1-ha cells across the DR. The analysis was conducted at two levels: first, for each of the four counties and then for the whole region. Performance was assessed using a comprehensive set of classification metrics.

Figure 4 provides a spatial overview of the validation framework. It shows the urban scope (grey cells), sampled 1-ha cells (green = true PUGS, red = non-PUGS), the complete spatial footprint of the CUGA dataset for the DR, and partial insets including all existing spatial features from the six investigated datasets. The visual distribution of green and red cells across the four counties shows the statistical representativeness of the sample and highlights the spatial heterogeneity of PUGS availability and detection.

The insets reveal significant differences between the polygons of the various datasets and how some of the potential PUGS in certain datasets result in FP. Beyond the necessary visual inspection provided by Figure 4 and the previous figures,

Table 3. Validation and performance results of datasets for each county and the whole DR. Validation columns show the number of true positives (TP), true negatives (TN), false positives (FP) and false negatives (FN). Performance columns show accuracy (Acc), performance (Prc), recall (Rec), specificity (Spc), F1-score (F1s), Matthews Correlation Coefficient (MCC), Balanced Accuracy (BA), Jaccard Index (JI) and Cohen’s Kappa (κ).

		Validation				Performance
County	Dataset	TP	TN	FP	FN	Acc	Prc	Rec	Spc	F1s	MCC	BA	JI	κ
DC	Adm_all	25	58	10	6	0.838	0.714	0.806	0.853	0.758	0.640	0.830	0.610	0.637
	Adm_prk	17	66	2	14	0.838	0.895	0.548	0.971	0.680	0.611	0.759	0.515	0.580
	OSM_grn	31	46	22	0	0.778	0.585	1.000	0.676	0.738	0.629	0.838	0.585	0.567
	OSM_prk	20	67	1	11	0.879	0.952	0.645	0.985	0.769	0.715	0.815	0.625	0.691
	CUA_GUA	26	57	11	5	0.838	0.703	0.839	0.838	0.765	0.649	0.838	0.619	0.643
	CUGA	30	67	1	1	0.980	0.968	0.968	0.985	0.968	0.953	0.977	0.938	0.953
DLR	Adm_all	4	45	1	18	0.721	0.800	0.182	0.978	0.296	0.287	0.580	0.174	0.200
	Adm_prk	4	46	0	18	0.735	1.000	0.182	1.000	0.308	0.361	0.591	0.182	0.231
	OSM_grn	21	33	13	1	0.794	0.618	0.955	0.717	0.750	0.629	0.836	0.600	0.588
	OSM_prk	11	46	0	11	0.838	1.000	0.500	1.000	0.667	0.635	0.750	0.500	0.575
	CUA_GUA	12	37	9	10	0.721	0.571	0.545	0.804	0.558	0.354	0.675	0.387	0.354
	CUGA	21	46	0	1	0.985	1.000	0.955	1.000	0.977	0.967	0.977	0.955	0.966
F	Adm_all	5	104	0	20	0.845	1.000	0.200	1.000	0.333	0.410	0.600	0.200	0.287
	Adm_prk	5	104	0	20	0.845	1.000	0.200	1.000	0.333	0.410	0.600	0.200	0.287
	OSM_grn	24	66	38	1	0.698	0.387	0.960	0.635	0.552	0.470	0.797	0.381	0.381
	OSM_prk	12	101	3	13	0.876	0.800	0.480	0.971	0.600	0.556	0.726	0.429	0.532
	CUA_GUA	16	96	8	9	0.868	0.667	0.640	0.923	0.653	0.572	0.782	0.485	0.572
	CUGA	22	104	0	3	0.977	1.000	0.880	1.000	0.936	0.925	0.940	0.880	0.922
SD	Adm_all	15	58	6	7	0.849	0.714	0.682	0.906	0.698	0.597	0.794	0.536	0.597
	Adm_prk	9	64	0	13	0.849	1.000	0.409	1.000	0.581	0.583	0.705	0.409	0.507
	OSM_grn	21	43	21	1	0.744	0.500	0.955	0.672	0.656	0.547	0.813	0.488	0.482
	OSM_prk	11	64	0	11	0.872	1.000	0.500	1.000	0.667	0.653	0.750	0.500	0.598
	CUA_GUA	15	56	8	7	0.826	0.652	0.682	0.875	0.667	0.549	0.778	0.500	0.549
	CUGA	20	64	0	2	0.977	1.000	0.909	1.000	0.952	0.939	0.955	0.909	0.937
DR	Adm_all	49	265	17	51	0.822	0.742	0.490	0.940	0.590	0.500	0.715	0.419	0.483
	Adm_prk	35	280	2	65	0.825	0.946	0.350	0.993	0.511	0.510	0.671	0.343	0.430
	OSM_grn	97	188	94	3	0.746	0.508	0.970	0.667	0.667	0.560	0.818	0.500	0.492
	OSM_prk	54	278	4	46	0.869	0.931	0.540	0.986	0.684	0.644	0.763	0.519	0.608
	CUA_GUA	69	246	36	31	0.825	0.657	0.690	0.872	0.673	0.554	0.781	0.507	0.553
	CUGA	93	281	1	7	0.979	0.989	0.930	0.996	0.959	0.946	0.963	0.921	0.945

offers the validation and performance metrics for each dataset in each of the four counties and across the full validation sample (n = 382).

In DC, CUGA achieved near-perfect performance, with an F1-score of 0.968 and MCC of 0.953. It correctly identified 30 of 31 true PUGS and misclassified only one non-PUGS cell. In contrast, OSM_grn achieved perfect recall (1.000) but had a much lower precision (0.585), indicating a high false positive rate. Adm_prk and Adm_all showed high precision (0.895 and 0.714, respectively) but low recall (0.548 and 0.806), reflecting their limited spatial extent. CUA_GUA and OSM_prk performed moderately well, but both missed a significant number of true PUGS.

CUGA also outperformed all alternatives in DLR, with an F1-score of 0.977 and MCC of 0.967. It correctly identified 21 of 22 true PUGS and had no false positives. Other datasets struggled: Adm_prk and Adm_all had very low recall (0.182), and OSM_prk, while achieving perfect precision, only identified half of the actual PUGS. CUA_GUA and OSM_grn showed moderate recall but suffered from low precision, leading to weaker overall performance.

CUGA maintained strong performance in F (F1-score = 0.936, MCC = 0.925), with perfect precision and high recall (0.880). Adm_prk and Adm_all again had excellent precision (1.000) but extremely low recall (0.200), indicating that they missed most true PUGS. OSM_grn achieved very high recall (0.960) but with poor precision (0.387), resulting in a high false positive rate. CUA_GUA and OSM_prk performed moderately, with F1-scores of 0.653 and 0.600, respectively.

In SD, CUGA achieved the highest F1-score (0.952) and MCC (0.939), with perfect precision and a recall of 0.909. OSM_prk and Adm_prk again showed high precision but low recall, while OSM_grn had high recall (0.955) but low precision (0.500). CUA_GUA performed moderately (F1-score = 0.667), with balanced but suboptimal precision and recall.

Across the entire DR, CUGA demonstrated superior performance across all metrics. It achieved the highest F1-score (0.959), MCC (0.946), and Cohen’s Kappa (0.945), with precision and recall both above 0.93. This indicates that CUGA not only correctly identified most true PUGS but also minimized false positives.

In contrast, Adm_prk had high precision (0.946) but very low recall (0.350), leading to an F1-score of just 0.511. OSM_prk performed better (F1-score = 0.684) but still missed nearly half of the actual PUGS. OSM_grn achieved the highest recall (0.970) but had the lowest precision (0.508), resulting in a high false positive rate and a moderate F1-score (0.667). CUA_GUA showed balanced but lower performance (F1-score = 0.673), reflecting its limitations in resolution and classification accuracy.

These results confirm that CUGA offers a more accurate and consistent method for identifying PUGS across diverse urban contexts. Its high precision ensures that identified areas are truly public and green, while its high recall guarantees that few actual PUGS are missed—an essential balance for urban planning and policy applications.

3.3. Statistical Superiority of CUGA

The results of McNemar’s test (

Table 4. McNemar’s test results from the paired comparison between CUGA and each of the analysed datasets at county level and for the whole DR. Columns show the number of discordant pairs, b and c (with b being the number of cases where CUGA correctly identifies a UGA and the comparison method does not, and c being the number of cases where the comparison method is correct and CUGA is not), and p-value (being statistically significant if p ≤ 0.05).

County	Dataset	b	c	p-Value
DC	Adm_all	14	0	<0.001
	Adm_prk	15	1	<0.001
	OSM_grn	21	1	<0.001
	OSM_prk	11	1	0.006
	CUA_GUA	15	1	<0.001
DLR	Adm_all	18	0	<0.001
	Adm_prk	17	0	<0.001
	OSM_grn	14	1	<0.001
	OSM_prk	10	0	0.002
	CUA_GUA	18	0	<0.001
F	Adm_all	17	0	<0.001
	Adm_prk	17	0	<0.001
	OSM_grn	38	2	<0.001
	OSM_prk	13	0	<0.001
	CUA_GUA	14	0	<0.001
SD	Adm_all	11	0	<0.001
	Adm_prk	11	0	<0.001
	OSM_grn	21	1	<0.001
	OSM_prk	9	0	0.004
	CUA_GUA	15	2	0.002
DR	Adm_all	60	0	<0.001
	Adm_prk	60	1	<0.001
	OSM_grn	94	5	<0.001
	OSM_prk	43	1	<0.001
	CUA_GUA	62	3	<0.001

) provide strong statistical evidence that the CUGA method significantly outperforms all other evaluated datasets in correctly identifying PUGS. For each pairwise comparison between CUGA and the alternative methods, the number of cases where CUGA correctly identified a UGA while the other method failed (b) was consistently and substantially higher than the number of cases where the other method was correct and CUGA was not (c). This pattern held true across all four counties and, therefore, for the entire DR.

The p-values obtained in all comparisons were below the 0.05 threshold, indicating that the differences in classification performance are statistically significant. Importantly, McNemar’s test does not merely indicate that the methods differ; it specifically tests whether one method is more frequently correct in cases where the other is not. Therefore, the consistent finding that b > c across all comparisons supports the conclusion that CUGA is not just different, but more accurate in identifying PUGS than each of the other datasets.

These results reinforce the findings from the validation metrics and confirm that the methodological approach used to construct the CUGA dataset yields a statistically superior capacity to detect PUGS across every county and for the whole region.

3.4. Net Area Analysis and Statistical Differences Between Methods

Beyond classification accuracy, we aimed to assess the spatial impact of each dataset in terms of the actual area of green space identified and its implications for urban accessibility. Table 5 shows the net areas (those of the polygons associated to TP minus those of the polygons associated to FP) of three key spatial dimensions:

• Urban Green Area (UGA): area of potential PUGS.
• Individual Catchment Area (ICA): area of 300-m buffer around each potential PUGS
• Intercepted Residential Land (IRL): area of residential land intersected by each ICA, representing potential population exposure.

Subsequent columns in the table rank each method based on the amount of net area for each county and for the entire DR. Across all four counties—except for UGA in SD, where Adm_prk identified 300.4 ha while CUGA identified 289.3 ha—the CUGA dataset consistently ranked first in all three spatial dimensions (UGA, ICA, IRL), outperforming all other datasets. For example, in SD, CUGA identified 289.3 ha of UGA, 1429.7 ha of ICA, and 914.4 ha of IRL—substantially more than any other dataset. In F, CUGA achieved 497.9 ha of UGA and 1560.6 ha of ICA, intercepting 1062.2 ha of residential land, compared to just 388.6 ha of UGA and 665.8 ha of ICA from Adm_prk. In DC, CUGA identified 3718.5 ha of UGA and 2848.2 ha of IRL, surpassing OSM_prk and Adm_all by over 500 ha in IRL. In DLR, CUGA again led with 240.9 ha of UGA and 1532.8 ha of IRL, while CUA_GUA showed negative net area values due to false positives.

Table 5. Urban Green Area (UGA), Individual Catchment Area (IGA) and Intercepted Residential Land (IRL) for the potential PUGS intercepted by the sampled 1-ha cells. Datasets result for each county and the whole DR. Columns display the number of net positives (NP = TP − FP), net area values in hectares, and rankings (according to net area values).

			Net Areas [ha]			Rankings
County	Dataset	NP	UGA	ICA	IRL	UGA	ICA	IRL
SD	Adm_all	9	121.5	671.4	595.0	5	4	4
	Adm_prk	9	300.4	1096.9	723.0	1	2	2
	OSM_grn	0	71.6	311.5	484.3	6	6	5
	OSM_prk	11	273.4	1033.7	597.8	3	3	3
	CUA_GUA	7	183.2	585.0	433.3	4	5	6
	CUGA	20	289.3	1429.7	914.4	2	1	1
F	Adm_all	5	388.6	665.8	272.4	2	4	4
	Adm_prk	5	388.6	665.8	272.4	2	4	4
	OSM_grn	−14	86.0	−269.5	431.8	6	6	3
	OSM_prk	9	332.7	803.3	548.7	4	2	2
	CUA_GUA	8	274.1	728.3	167.3	5	3	6
	CUGA	22	497.9	1560.6	1062.2	1	1	1
DC	Adm_all	15	3136.2	3122.2	2415.7	3	2	2
	Adm_prk	15	3120.2	2587.0	1851.2	4	4	4
	OSM_grn	9	2096.3	1286.5	1466.9	5	5	5
	OSM_prk	19	3690.4	3112.5	2324.0	2	3	3
	CUA_GUA	15	325.1	1035.1	835.4	6	6	6
	CUGA	29	3718.5	3553.6	2848.2	1	1	1
DLR	Adm_all	3	232.3	480.7	307.9	3	5	5
	Adm_prk	4	233.6	508.1	345.8	2	4	4
	OSM_grn	8	182.3	735.2	1019.3	5	3	2
	OSM_prk	11	217.6	797.3	748.8	4	2	3
	CUA_GUA	3	−63.5	−19.2	94.5	6	6	6
	CUGA	21	240.9	1288.0	1532.8	1	1	1
DR	Adm_all	32	3878.7	4940.0	3590.9	4	3	3
	Adm_prk	33	4042.9	4857.8	3192.5	3	4	5
	OSM_grn	3	2436.2	2063.8	3402.3	5	6	4
	OSM_prk	50	4514.0	5746.8	4219.2	2	2	2
	CUA_GUA	33	718.8	2329.3	1530.5	6	5	6
	CUGA	92	4746.5	7831.9	6357.5	1	1	1

Table 5. Urban Green Area (UGA), Individual Catchment Area (IGA) and Intercepted Residential Land (IRL) for the potential PUGS intercepted by the sampled 1-ha cells. Datasets result for each county and the whole DR. Columns display the number of net positives (NP = TP − FP), net area values in hectares, and rankings (according to net area values).

			Net Areas [ha]			Rankings
County	Dataset	NP	UGA	ICA	IRL	UGA	ICA	IRL
SD	Adm_all	9	121.5	671.4	595.0	5	4	4
	Adm_prk	9	300.4	1096.9	723.0	1	2	2
	OSM_grn	0	71.6	311.5	484.3	6	6	5
	OSM_prk	11	273.4	1033.7	597.8	3	3	3
	CUA_GUA	7	183.2	585.0	433.3	4	5	6
	CUGA	20	289.3	1429.7	914.4	2	1	1
F	Adm_all	5	388.6	665.8	272.4	2	4	4
	Adm_prk	5	388.6	665.8	272.4	2	4	4
	OSM_grn	−14	86.0	−269.5	431.8	6	6	3
	OSM_prk	9	332.7	803.3	548.7	4	2	2
	CUA_GUA	8	274.1	728.3	167.3	5	3	6
	CUGA	22	497.9	1560.6	1062.2	1	1	1
DC	Adm_all	15	3136.2	3122.2	2415.7	3	2	2
	Adm_prk	15	3120.2	2587.0	1851.2	4	4	4
	OSM_grn	9	2096.3	1286.5	1466.9	5	5	5
	OSM_prk	19	3690.4	3112.5	2324.0	2	3	3
	CUA_GUA	15	325.1	1035.1	835.4	6	6	6
	CUGA	29	3718.5	3553.6	2848.2	1	1	1
DLR	Adm_all	3	232.3	480.7	307.9	3	5	5
	Adm_prk	4	233.6	508.1	345.8	2	4	4
	OSM_grn	8	182.3	735.2	1019.3	5	3	2
	OSM_prk	11	217.6	797.3	748.8	4	2	3
	CUA_GUA	3	−63.5	−19.2	94.5	6	6	6
	CUGA	21	240.9	1288.0	1532.8	1	1	1
DR	Adm_all	32	3878.7	4940.0	3590.9	4	3	3
	Adm_prk	33	4042.9	4857.8	3192.5	3	4	5
	OSM_grn	3	2436.2	2063.8	3402.3	5	6	4
	OSM_prk	50	4514.0	5746.8	4219.2	2	2	2
	CUA_GUA	33	718.8	2329.3	1530.5	6	5	6
	CUGA	92	4746.5	7831.9	6357.5	1	1	1

At the regional level, CUGA identified: 4746.5 ha of UGA, compared to 4514.0 ha by OSM_prk and just 718.8 ha by CUA_GUA; 7831.9 ha of ICA, far exceeding all other datasets; and 6357.5 ha of IRL, indicating the highest potential population exposure to green space. These figures suggest that CUGA provides the most extensive and equitable spatial footprint of PUGS across the DR.

The Kruskal–Wallis test (

Table 6. Results of the Kruskal–Wallis (K-W) test for each type of computed net area (UGA, ICA, IRL) across four counties and DR. The table reports the K-W statistic and the associated p-value (statistically significant if p < 0.05).

Area Type	County	K-W	p-Value
UGA	DC	20.38	0.001
	DLR	46.69	<0.001
	F	113.29	<0.001
	SD	35.97	<0.001
	DR	178.20	<0.001
ICA	DC	19.17	0.002
	DLR	46.00	<0.001
	F	107.56	<0.001
	SD	32.98	<0.001
	DR	167.84	<0.001
IRL	DC	21.86	<0.001
	DLR	49.44	<0.001
	F	87.49	<0.001
	SD	31.84	<0.001
	DR	158.89	<0.001

) revealed statistically significant differences (p < 0.05) in net area values across all datasets for each spatial dimension and county. Post hoc Mann–Whitney U tests (

Table 7. Results of Mann–Whitney U test pairwise comparison between methods, for the UGA, ICA and IRL net areas. The matrix displays corresponding p-values (*) statistically significant if p < 0.05.

UGA	Adm_all	Adm_prk	OSM_grn	OSM_prk	CUA_UGA	CUGA
Adm_all	—	0.018 *	<0.001 *	0.779	0.001 *	0.010 *
Adm_prk	0.018 *	—	<0.001 *	0.040 *	<0.001 *	<0.001 *
OSM_grn	<0.001 *	<0.001 *	—	<0.001 *	<0.001 *	<0.001 *
OSM_prk	0.779	0.040 *	<0.001 *	—	<0.001 *	0.004 *
CUA_UGA	0.001 *	<0.001 *	<0.001 *	<0.001 *	—	0.517
CUGA	0.010 *	<0.001 *	<0.001 *	0.004 *	0.517	—
ICA	Adm_all	Adm_prk	OSM_grn	OSM_prk	CUA_UGA	CUGA
Adm_all	—	0.019 *	<0.001 *	0.791	0.001 *	0.008 *
Adm_prk	0.019 *	—	<0.001 *	0.040 *	<0.001 *	<0.001 *
OSM_grn	<0.001 *	<0.001 *	—	<0.001 *	<0.001 *	<0.001 *
OSM_prk	0.791	0.040 *	<0.001 *	—	<0.001 *	0.004 *
CUA_UGA	0.001 *	<0.001 *	<0.001 *	<0.001 *	—	0.604
CUGA	0.008 *	<0.001 *	<0.001 *	0.004 *	0.604	—
IRL	Adm_all	Adm_prk	OSM_grn	OSM_prk	CUA_UGA	CUGA
Adm_all	—	0.015 *	<0.001 *	0.783	<0.001 *	0.004 *
Adm_prk	0.015 *	—	<0.001 *	0.031 *	<0.001 *	<0.001 *
OSM_grn	<0.001 *	<0.001 *	—	<0.001 *	<0.001 *	<0.001 *
OSM_prk	0.783	0.031 *	<0.001 *	—	<0.001 *	0.002 *
CUA_UGA	<0.001 *	<0.001 *	<0.001 *	<0.001 *	—	0.511
CUGA	0.004 *	<0.001 *	<0.001 *	0.002 *	0.511	—

) confirmed that CUGA significantly outperformed all other datasets except CUA_GUA in pairwise comparisons.

Interestingly, while CUA_GUA showed comparable performance in some ICA comparisons (e.g., p = 0.517 vs. CUGA), it consistently underperformed in UGA and IRL metrics. OSM_grn, despite its high recall, was statistically inferior to CUGA in all dimensions due to its low precision and inflated false positive rates, which undermined its performance in the net area analysis.

These findings reinforce the conclusion that CUGA not only excels in classification accuracy but also delivers superior spatial outcomes, making it a robust tool for urban green space planning and accessibility analysis.

4. Discussion

4.1. Comparative Performance of CUGA and Existing Datasets

The results presented in the preceding section demonstrate that the CUGA method outperforms all other evaluated datasets in identifying PUGS across the DR. This superiority is evident across multiple dimensions: classification accuracy, spatial coverage, and statistical robustness.

From a classification standpoint, CUGA achieved the highest F1-scores and MCC in all four counties and at the regional level. Its ability to simultaneously maintain high precision and high recall is particularly noteworthy. While administrative datasets (i.e., Adm_prk and Adm_all) and OSM_prk exhibited high precision, they consistently failed to detect a large proportion of actual PUGS, resulting in low recall and moderate F1-scores. Conversely, OSM_grn achieved high recall but suffered from low precision, leading to inflated FP rates and reduced overall reliability.

CUA_GUA, despite offering harmonized coverage and a moderate balance between precision and recall, was limited by its resolution and classification scheme. It frequently misclassified non-public or inaccessible green areas as PUGS and failed to detect smaller or irregularly shaped spaces, particularly in dense urban environments.

The statistical significance of CUGA’s performance advantage was confirmed through McNemar’s test, which showed that CUGA was significantly more accurate than all other datasets in nearly every pairwise comparison. This was consistent across all counties and the entire DR, with p-values well below the 0.05 threshold and a consistent pattern of discordant wins in favor of CUGA.

These findings underscore the robustness of the CUGA method as a classification tool. Its design—based on a structured integration of OSM and CUA data, combined with spatial filtering and accessibility criteria—enables it to overcome the limitations of both overly restrictive and overly inclusive approaches. As a result, CUGA provides a more reliable and equitable representation of publicly accessible green spaces, which is essential for urban planning, environmental justice, and health-related research.

4.2. Spatial Robustness Across Urban Contexts

One of the most compelling strengths of the CUGA method is its consistent performance across diverse urban morphologies and administrative contexts. The DR, comprising four counties—DC, DLR, F, and SD—offers a varied landscape of urban density, land use patterns, and GI typologies. This heterogeneity provides a robust testing ground for evaluating the spatial adaptability of PUGS identification methods.

CUGA demonstrated high classification accuracy and spatial coverage in all four counties, regardless of their urban form. In DC, characterized by dense built-up areas and fragmented GS, CUGA successfully identified small and irregularly shaped PUGS that were missed by coarser datasets such as CUA_GUA. In contrast, in suburban counties like F and SD, where GS are more dispersed and often embedded within low-density residential zones, CUGA maintained high recall and precision, outperforming both administrative and crowdsourced datasets.

The method’s robustness is further evidenced by its ability to harmonize data across administrative boundaries. Unlike Adm_prk and Adm_all, which suffer from inconsistencies in coverage and classification between counties, CUGA applies a uniform set of inclusion and accessibility criteria, ensuring comparability and continuity across the region. This is particularly important in FUA contexts, where robust analysis and planning decisions should necessarily span multiple jurisdictions.

Moreover, the spatial filtering and contextual tagging embedded in the CUGA workflow allow it to adapt to local conditions—such as proximity to pedestrian networks, residential buffers, and coastline adjacency—without relying on locally curated datasets. This enhances its transferability to other metropolitan regions with similar data constraints.

CUGA’s spatial robustness across urban contexts makes it a reliable tool for regional-scale assessments of GS accessibility. Its consistent performance in both dense urban cores and sprawling suburban landscapes supports its applicability in diverse planning scenarios and reinforces its value for comparative studies across EU cities.

4.3. Strengths and Limitations of the Validation Framework

The validation framework employed in this study was designed to ensure methodological rigor and spatial representativeness while remaining feasible within the constraints of available data and resources. By combining stratified random sampling, manual ground-truthing, and a comprehensive set of performance metrics, the framework provides a robust basis for evaluating the reliability of PUGS identification methods.

One of the key strengths of the approach is the use of a grid-based random sample of 1-ha cells across the DR. This spatial unit balances granularity with manageability, allowing for meaningful comparisons across urban contexts while maintaining statistical power. The sample size (n = 382) was calculated to achieve a 95% confidence level with a 5% margin of error, ensuring that the findings are generalizable to the broader urban scope.

The validation process itself relied on multi-source visual inspection, including Google satellite and Street View imagery, online information, and site visits. This triangulation of methods enhances the reliability of the ground truth, particularly in the absence of standardized public datasets. The inclusion of site visits for ambiguous cases further strengthens the credibility of the validation outcomes.

However, the framework is not without limitations. First, the reliance on visual inspection introduces a degree of subjectivity, particularly in borderline cases where public accessibility is unclear. While efforts were made to standardize interpretation criteria, some variability is inevitable. Second, the use of Google-based imagery may introduce temporal inconsistencies, as satellite and street-level data are updated at different intervals and may not reflect recent changes in land use or access.

Additionally, the validation was confined to a single metropolitan region. While the DR offers a diverse urban landscape, further testing in cities with different planning regimes, vegetation types, and cultural contexts would be necessary to fully assess the generalizability of the findings.

Despite these limitations, the validation framework provides a solid foundation for assessing the performance of geospatial datasets in identifying PUGS. Its combination of statistical rigor, spatial representativeness, and multi-source verification makes it a valuable model for future studies in urban green space accessibility.

4.4. Methodological Innovations and Transferability

The CUGA method introduces some methodological innovations that enhance the reliability, scalability, and applicability of PUGS identification in metropolitan contexts. These innovations address key limitations in existing approaches and offer a replicable framework for other urban regions, particularly where high-quality local datasets are unavailable or inconsistent.

A central innovation lies in the structured extensive integration of OSM and CUA data. Rather than relying on a single source or a narrow set of tags (e.g., ‘leisure = park’), CUGA employs a multi-step filtering process that combines land use, land cover, and accessibility attributes. This allows for the inclusion of a broader and more realistic range of GS while excluding areas that do not meet public accessibility or spatial functionality criteria.

The method also incorporates context-sensitive spatial filters, such as proximity to pedestrian networks, adjacency to residential areas, and minimum shape functionality (e.g., ability to inscribe a 30-m circle). These criteria ensure that identified PUGS are not only green and public but also usable and accessible in practice—an aspect often overlooked in automated mapping approaches.

Another key contribution is the modular and transparent design of the workflow. Each step—from raw data extraction to final polygon consolidation—is explicitly defined and reproducible using open-source tools. This transparency facilitates peer review, adaptation, and extension by other researchers or planning authorities.

Importantly, the method is transferable across EU cities and potentially beyond. The reliance on globally available OSM data and CUA coverage ensures that the approach can be applied in most FUAs with populations over 50,000. Moreover, the method’s adaptability to different urban forms—demonstrated in the DR—suggests its suitability for diverse planning contexts, from compact city cores to dispersed suburban landscapes.

In sum, the CUGA method offers a replicable and scalable solution for identifying PUGS with improved accuracy and spatial relevance. Its methodological transparency and reliance on open-access data make it a valuable tool for urban researchers, planners, and policymakers seeking to monitor and enhance GS accessibility in line with sustainability and equity goals.

4.5. Policy and Planning Implications

The findings of this study have direct implications for urban policy, spatial planning, and sustainability monitoring. As cities across Europe and beyond strive to meet the targets of the 2030 Agenda for Sustainable Development—particularly SDG 11.7 on universal access to green and public spaces—reliable, scalable, and equitable methods for identifying PUGS are urgently needed.

The CUGA method offers a practical solution to this challenge. Its reliance on open-access datasets and transparent criteria makes it particularly well-suited for municipalities with limited geospatial resources or fragmented administrative structures. By providing a harmonized and verifiable inventory of accessible GS, CUGA can support the development of evidence-based GI strategies, inform land-use planning, and guide investment in underserved areas.

Moreover, the method aligns with emerging policy frameworks such as the EU GI Strategy, the New Urban Agenda, WHO-ROE recommendations, and the 3–30–300 rule for urban forestry. Its ability to identify not only large parks but also smaller, functionally relevant PUGS enhances its utility for neighbourhood-scale interventions and equity-focused planning.

CUGA also holds potential for monitoring and reporting. Its compatibility with SDG indicator 11.7.1 and its capacity to generate reproducible metrics (e.g., ICA, IRL) make it a valuable tool for tracking progress toward sustainability goals. In this sense, it can complement or even substitute for official datasets in contexts where such data are outdated, incomplete, or inconsistent.

Finally, the method’s adaptability opens the door to cross-city comparisons and regional benchmarking, enabling policymakers to assess disparities in GS provision and accessibility across metropolitan areas. This can foster more coordinated and inclusive approaches to urban greening, particularly in FUA where governance is often fragmented.

4.6. Method Limitations and Future Research

While the CUGA method demonstrates strong performance and broad applicability, several limitations should be acknowledged to contextualize its findings and guide future research.

First, the method’s reliance on open-access datasets—namely OSM and CUA—means that its accuracy is inherently tied to the quality and completeness of these sources. Although CUGA mitigates many of their individual shortcomings through integration and filtering, it cannot fully compensate for missing or outdated data, particularly in cities with low OSM contributor activity or limited CUA coverage.

Second, the definition of PUGS adopted in this study, while grounded in WHO and UN-Habitat guidelines, remains operational and context-dependent. The 0.5-ha threshold, the 30-m circle criterion, and the partial exclusion of sports fields (when not publicly accessible or not allowing other uses) are all reasonable but ultimately normative choices. Future work could explore the sensitivity of results to alternative definitions, especially in cities with different cultural or spatial norms regarding GS use.

Third, the validation framework, although robust, was applied in a single metropolitan region and further validation could be costly. While the DR offers a diverse urban landscape, further testing in other EU FUA would be necessary to confirm the method’s generalizability. This includes cities with different public space cultures, GI characteristics, and urban forms.

Additionally, the current method does not incorporate temporal dynamics. UGS are subject to change due to development, regeneration, or natural succession. Integrating time-series data or change detection techniques could enhance the method’s utility for longitudinal monitoring and policy evaluation.

Finally, while the method is designed for replicability and scalability, its implementation still requires a degree of technical expertise in GIS and spatial analysis. Future work could focus on developing user-friendly tools or plugins that automate the workflow and make it accessible to a broader range of practitioners.

In summary, while we believe that CUGA represents a significant methodological advance, ongoing refinement and broader application are needed to fully realize its potential as a tool for equitable and sustainable PUGS planning.

5. Conclusions

This study introduces and validates the CUGA method as a robust, scalable, and transferable approach for identifying PUGS in metropolitan contexts. By integrating OSM and CUA data through a structured, transparent workflow, CUGA addresses key limitations of existing datasets—namely, their restricted scope, inconsistent classification, and limited spatial resolution.

Tested in the DR, CUGA consistently outperformed administrative, crowdsourced, and satellite-derived datasets across all counties and evaluation metrics. It demonstrated superior classification accuracy, spatial coverage, and statistical robustness, while maintaining high precision and recall. Its ability to harmonize data across jurisdictions and adapt to diverse urban morphologies underscores its potential for broader application.

Beyond technical performance, CUGA offers significant value for urban policy and planning. It supports evidence-based decision-making aligned with SDG 11.7 and the EU GI Strategy, and provides a replicable model for cities seeking to enhance GS accessibility and equity. Its compatibility with open data and open-source tools makes it particularly suitable for municipalities with limited resources or fragmented governance.

Importantly, CUGA also contributes to the scientific advancement of urban spatial analysis. It demonstrates how combining open geospatial data with spatial logic and accessibility criteria can yield more accurate and equitable representations of UGI. This opens new avenues for research on environmental justice, health equity, and urban resilience.

Future work should explore the method’s applicability in other geographic and institutional contexts, assess its sensitivity to alternative PUGS definitions, and extend it to dynamic monitoring through time-series analysis or integration with AI-based classification. Additionally, coupling CUGA with participatory mapping or citizen science initiatives could further enhance its relevance and legitimacy in local planning processes.

In sum, CUGA represents a meaningful step forward in the identification and monitoring of public GS, offering a practical and policy-relevant tool for advancing more inclusive, sustainable, and livable cities.