Unveiling Hidden Green Corridors: An Agent-Based Simulation (ABS) of Urban Green Continuity for Ecosystem Services and Climate Resilience

Highlights

What is the main finding?

• This paper introduces a novel method to evaluate urban green infrastructure performance by combining multi-species agent simulation and space syntax analysis. Specifically, the results of the Agent-based simulation of ecological behaviors reveal hidden green networks that are not aligned with existing green space layouts. Moreover, the spatial overlaps between pollinator intensity and thermal vulnerability expose coupled ecological–climatic risks in urban areas.

What are the implications of the main findings?

• Integrating dynamic ecological behavior simulation with temporal performance monitoring enables more precise identification of priority intervention zones for green infrastructure planning.
• The dual-agent framework combining ABS and AI interpretation provides a scalable approach for diagnosing and designing urban ecological resilience.

Abstract

Urban green spaces are essential for mitigating the heat island effect, supporting ecosystem services, and maintaining biodiversity. The distribution, fragmentation, and connection of the green spaces significantly impact the behavior of species in cities, serving as key indicators of environmental resilience and ecological benefits. However, current studies, as well as planning standards, often prioritize green spaces independently through their coverage or density, overlooking the importance of continuity and its impact on thermal regulation and accessibility. In this research, urban “hidden green corridors” refer to the unrecognized but functionally significant pathways that link fragmented green spaces through ecological behaviors, which enhance both biological and human habitats. This research focuses on developing an agent-based simulation (ABS) model based on the Physarealm plugin in Rhino, which can assess the effectiveness of these hidden corridors in different urban settings by integrating geographic information systems (GIS) and space syntax. Based on three case studies in Italy (Lambrate District, Bolognina, and Ispra), the simulation results are further interpreted through the AI agentic workflow “SOFIA”, developed by IMM Design Lab, Politecnico di Milano, and compared using manual analysis as well as mainstream large language models (ChatGPT 4.0 Web). The findings indicate that the “hidden green corridors” are essential for urban heat reduction, enhancement of urban biodiversity, and strengthening ecological flows.

1. Introduction

In an era of growing urban detachment from nature, the biophilic hypothesis reminds us that humans possess an evolutionarily ingrained instinct for seeking connections with living systems [1]. As proposed by Wilson [2], our affinity for nature is understood as an instinct shaped by millennia of human–nature interaction. However, modern urbanization has fractured this relationship, resulting in an invisible extinction of nature experience, which undermines human physical, psychological, and social well-being [3,4]. Reconstructing that connection through accessible, practical urban green infrastructure is not only desirable but essential.

1.1. Fragmentation Crisis and the Disruption of Urban Green Systems

Accelerated urbanization has intensified the competition for land [5,6,7], leading to an increasing fragmentation of urban green spaces into isolated patches that are disconnected and functionally inefficient [8]. Ecosystem integrity is thus undermined, creating spatial and functional disconnections, which are particularly detrimental to species that require large, continuous areas of habitat, like birds and small mammals [9]. Such fragmentation disrupts the integrity of ecosystems and reduces the capacity of urban landscapes to regulate microclimates and maintain hydrological cycles [10]. The discontinuity of vegetation corridors induces a series of interlocking ecological degradation, resulting in a chain effect of diminished biodiversity and malfunction in genetic movement among urban species and pollinator activities [11,12].

Fragmented green spaces exhibit low ecological efficiency and reduce the effectiveness in mitigating the urban heat island (UHI) effect. Research in Asian and European cities shows that poorly connected green spaces correlate with higher land surface temperature and restricted air circulation [13,14]. Additionally, when lack of a comprehensive planning and management, it is limited to forming effective thermal regulation at the community scale, resulting in exacerbated heat island effects in high-density urban areas [15,16]. These areas often coincide with inadequate infrastructure, degraded habitats, and isolated species populations, which further intensify both environmental and biological vulnerabilities [6,17], ultimately leading to greenspace deprivation.

1.2. Green Continuity as a Framework for Urban Resilience

While the fragmentation of green networks reveals their spatial rupture, green continuity is representative of their functional restorative capacity [18,19]. This concept refers not only to the existence and proximity of vegetation but also to the ability of ecological processes to persist in highly urbanized environments. This concern is further compounded by the understanding that urban green infrastructure should be viewed as “distributed ecological flows” rather than static land cover [1]. Even in the absence of habitat area reductions, fragmentation itself can significantly disrupt species migration and reduce biodiversity [9]. Green continuity, therefore, emphasizes not the existence of green space, but the ability to support species dispersal, gene flow, and climate regulation [15,20]. The concept of “green continuity” has emerged across diverse strands of ecological and planning literature, although not yet been formalized as a standard term. Synthesizing insights from studies on functional connectivity, landscape ecology, and urban biodiversity, it can be understood as the spatio-temporal ecological capacity of green infrastructures to sustain multispecies interactions, microclimatic regulation, and urban resilience [12,19,21,22,23,24].

This paper proposes “green continuity” over “green connectivity” to highlight the systemic and structural aspects of green infrastructure and emphasizes not just the presence of connections between green spaces, but the integrated network that serves as the backbone of the urban ecosystem. Rather than static green metrics, it emphasizes the characteristics of ecological flows, behavioral pathways, and biotic support across fragmented urban environments. Compared with “connectivity”, which mainly refers to the extent of landscape features that allow species to move or ecological flows, “continuity” addresses to the functional and physical integration of natural assets across scales, guaranteeing long-term ecological processes, habitat provision, and ecosystem services [25,26]. Under this definition, Green Continuity highlights the integration of green infrastructure, including core areas, corridors, and restoration zones, into an interconnected system that supports biodiversity, climate adaptation, and urban resilience [27,28]. This approach aligns with the concept of green infrastructure as a strategically planned network that sustains ecological functions and human well-being over time, not merely as a set of irregularly linked patches, but as a structural and functional continuum essential for urban resilience [29].

At the functional level, green continuity ensures that urban green space operates as ecological infrastructure. The configuration and distribution of green spaces, encompassing not only the area itself but also its physical characteristics, have been demonstrated to contribute to the efficacy of cooling [20,30]. Meanwhile, the temporal dimension of landscape connectivity is an important predictor of current species richness [12], but this history-related factor is often overlooked in the context of rapidly changing urban regeneration [31,32]. Therefore, green continuity should be comprehended in both spatial and temporal dimensions, as a form of ecological memory that links past connectivity to present function.

For species with limited dispersal abilities, such as pollinators and small birds, green continuity plays an especially critical role. The mobility and networking of pollinators are significantly constrained by the distribution and quality of green spaces [31], as well as their temporal instability [33]. The improvement of urban diversity depends mainly on the connectivity between spaces instead of the total area [34]. On the other hand, the resilience of the ecosystem may receive misjudgment if the range of species’ behavioral activities is ignored [35]. Therefore, green space continuity should be understood as an ecological support system that responds to the needs of multiple species, rather than just a structural indicator.

1.3. The Ecological Blind Spot: Planning Inertia and the Invisibility of More-than-Human Needs

The governance of urban green infrastructure depends not only on its spatial layout; it is deeply rooted in the cognition of “what infrastructure is”. In mainstream planning systems, “legibility” is key to the inclusion of governance: only those elements that are visible, measurable, and manageable are likely to be recognized by the system. This tendency toward infrastructural simplification often marginalizes the complex ecological relationships embedded in everyday life, effectively labeling them as “non-infrastructural [36,37,38].” For non-human species whose ecological functions depend on diffusion, rhythmic patterns, or dynamic interactions rather than fixed spatial locations, such logic renders them systematically invisible [39,40]. The institutional prioritization of spatial legibility further reinforces the exclusion of actors and processes that do not fit within the gridded logic of urban form.

This blind spot is not merely an incidental flaw; instead, it is the consequence of planning inertia. The universalist model of modern urbanism has the effect of flattening both regional and ecological particularities, thereby causing the complexity of more-than-human territorialities to be erased. [41,42]. Fix and Arantes (2021) criticize the “technocratic monoculture” in Brazilian urban studies, which systematically ignores informality, temporal fluctuations, and multi-species interactions [43]. In this context, green infrastructure not only performs inadequately but is also structurally constrained in the categories perceived and intervened upon by urban planners. Addressing this ecological blind spot requires not only better data or tools, but a reframing of planning paradigms to account for the political invisibility of non-human beings [41,44].

In order to address the ecological blind spots of the current situation, this paper presents green continuity as a conceptual lens, which moves beyond traditional measures of connectivity to reframe green infrastructure as an integrated and adaptive system by emphasizing the temporal and spatial dynamics of ecological flows. This lays the basis for the following sections, which explore how green continuity can be systematically evaluated and operationalized in urban settings.

2. Theoretical Foundation and Research Landscape

2.1. Reframing Urban Green Value Through Ecosystem Services

Over the past two decades, the concept of ecosystem services (ESS) has gradually evolved from a descriptive term to a structural operational framework guiding environmental policy and planning practices. Ecosystem is commonly categorized into provisioning, regulating, cultural, and supporting services [45] and gradually used as a framework for quantifying the multifunctionality of green infrastructure [46]. This classification was later expanded and critically refined by initiatives such as The Economics of Ecosystems and Biodiversity (TEEB) [47] and the Intergovernmental Science-Policy Platform on Biodiversity and Ecosystem Services (IPBES) [48]. These frameworks put emphasis on both the direct value of ecosystem functions and relational/intangible benefits to human and nonhuman beings [49]. This conceptual shift led to the recognition of ecosystem services as a governance logic, or as a language that makes the value of nature visible, measurable, and actionable. However, scholars such as Parker and Simpson (2018) point out that while the ecosystem services framework provides a basis for policy implementation, it can also oversimplify ecological complexity and the interdependencies between ecosystems by reducing them to manageable units [1].

Although the ESS framework has gained traction at the policy level, its application in planning practice still faces many fractions. Compared with rural or nature reserves, urban green spaces are facing multifunctional loads and subject to complex socio-political pressure [41,43]. Quantifying these services in such contexts often gives rise to epistemological and methodological tensions [31,50]. Still, most assessments of ecosystem services rely on anthropocentric logic, assuming humans are the primary beneficiaries while ignoring the experiences of or disruption to non-humans [45,51,52]. This epistemological bias hinders the accurate assessment of ecological functions [53] such as pollination, seed dispersal, and species migration, which differ significantly from the behaviors of non-human agents.

Lindborg and Eriksson (2004) noted that “historical landscape connectivity” is a better predictor of current species richness than static vegetation maps [12], suggesting that ecosystem function has a “temporal signature.” Yet, most ESS assessment models remain entrenched in the “spatial presentism” paradigm, neglecting how “ecological memory” [54,55] including seasonal cycles, historical connectivity, and intergenerational learning among species influences current ecosystem performance. This issue is especially evident in ecological services, such as soil restoration and trophic stabilization, which are complicated to measure and quantify in short-term urban planning. As a consequence, many essential multispecies services remain “invisible” in assessments, not due to a lack of functionality, but because they do not align with existing assessment epistemologies [56,57,58].

In response to existing conceptual and practical constraints in ecosystem service assessments, the notion of green continuity provides a perspective to comprehend ecological value in both urban and natural settings by emphasizing the relational and temporal dimensions of ecological functions, particularly those influenced by species movement, behavioral adaptation, and historical landscape dynamics. P. Herreros-Cantis and T. McPhearson (2021) developed a method to map ecosystem service supply and demand to evaluate spatial distribution and environmental justice [59]. Wang and Pei (2020), using Gaoyou in Jiangsu Province as a case, assessed green corridor connectivity through landscape pattern indices, showing how spatial structure can be analyzed [19]. Recent research by Convertino and Wu (2025) on ecological corridors introduces a new perspective. The ecological corridor design and model (ECORR), based on machine learning, incorporates indices like probability of connectivity (PC) and integral index of connectivity (IIC) to identify potential patches and corridors [60].

However, these planning and decision-making measures and tools still rely on static spatial proxies such as land cover ratios, vegetation indices, or buffer-based accessibility, while underestimating the influence of ecological flows, interspecies cooperation, and delayed climate effects. This mismatch creates a clear methodological gap in how we understand green continuity. It is necessary to “map the research landscape of nature-based solutions in urbanism”, providing a clearer framework for future studies on green continuity [61].

2.2. A Review of Methods for Assessing Urban Green Space Continuity

Static, patch-based indicators have long been used to describe the extent, fragmentation, and morphology of urban green spaces, which are rooted in quantifying spatial patterns of vegetation or monitored by remote sensors. Metrics such as normalized difference vegetation index (NDVI), class area (CA), patch density (PD), landscape shape index (LSI), and largest patch index (LPI) are commonly integrated to describe the extent of green space and its morphological complexity.

As first introduced by Rouse et al. (1974), NDVI was developed to assess vegetation coverage [62] by calculating the difference in reflectance between the near-infrared and red light bands. Subsequently, with the openness of data sharing, more researchers like Weng and Yang (2004) began to use NDVI for research [63], such as the correlation between the urban heat island (UHI) effect and plant growth conditions. Indicators like CA, PD, LPI, and LSI are among the most common patch-based metrics in landscape ecology and have been extensively applied to assess urban green spaces at different scales [64,65,66,67]. These indicators are extensively utilized in landscape ecology and spatial planning because of their compatibility with remote sensing and GIS platforms.

Nonetheless, these indicators primarily represent the physical arrangement of green spaces, which is mainly reliant on static graphical or morphological features, rather than the ecological processes or species interactions, which limit the capacity to capture functional connectivity via dynamic behaviors, movement barriers, or ecological flows [8,68]. In recent years, assessments of green continuity have gradually expanded from static indicators to functional metrics that reflect species movement and landscape permeability. Among these, graph theory models are widely used to depict the structural connections and potential functional links between urban green spaces [69,70]. By abstracting landscapes into “nodes” (e.g., habitat patches) and “edges” (e.g., corridors), the integrity of ecological networks can be assessed by connectivity indices (such as the probability of connectivity, PC). Least-cost path (LCP) analysis combined with graph theory can identify optimal species movement routes based on land-use-derived resistance surfaces to pinpoint key ecological corridors. To more effectively capture multi-path ecological dynamics, circuit theory implemented through tools such as Circuitscape [71] models species movement by simulating ecological flow as electrical current across resistance surfaces.

Despite the continuous development of structural and functional assessment models, current methods still face challenges in capturing dynamic, species-specific, and more-than-human-scale ecological processes [54,72]. Static land-use classifications or generic resistance values are still relied upon by dominant methodologies, which are indeed insufficient to adequately reflect the temporal heterogeneity or behavioral feedback present in actual ecosystems [73,74,75]. For example, bees and other urban pollinators have foraging behaviors that are greatly localized and seasonally fluctuate [33,76]. These behaviors are difficult to predict by using only NDVI or patch-based indicators [73]. Moreover, current models often lack the capacity to simulate ecological interactions among multiple species or their behavioral responses to climate and urban pattern changes [77].

These limitations underscore a persistent methodological gap. Current assessment tools are inadequate in comprehensively capturing the dynamic, behavioral, and multispecies dimensions of green continuity. Consequently, many critical ecological functions remain unexplored, not due to their absence, but because tools to observe them have not yet been designed.

2.3. The Potential of Agent-Based Simulation (ABS) in Urban Green Continuity Assessment

Across broad domains, agent-based simulation (ABS) has evolved into a mature modeling tool that is widely applied in fields like transportation systems [78], planning evacuations [79], robotic systems [80], and autonomous driving [81]. It is capable of modeling decentralized decision-making behavior and individual heterogeneity, helping to build complex adaptive systems (CASs), and revealing emergent phenomena. It has been utilized to optimize the distribution of citizens during emergency evacuation, simulate vehicle interactions, and, more frequently, train autonomous driving models with environmental adaptation.

Even though ABS has been widely introduced in these fields, it still remains a potential tool for use in ecological space planning, especially in urban green infrastructure and biodiversity conservation, and their management. Qiu and Fricker (2023) developed an ABS model of bumblebee foraging behavior in Helsinki. This model utilized real urban morphology and patch resistance to assess foraging routes and pollination efficiency [82]. Another study published in PLOS One simulated pollinator behavior in urban micro-farms, proposing the existence of a “moderate zone” determined by patch size and floral composition that maximizes pollination services [83]. Convertino et al. (2025) proposed the concept of “sentinel species”, employing butterflies as agents to assess urban ecosystems’ adaptability, connectivity, and climate responsiveness [84].

These studies show that ABS can imitate transforming ecological processes, like avoiding certain behaviors, remembering paths, or choosing habitats, which static indicators fail to do. ABS provides a more ecologically accurate way to investigate green continuity by focusing on species-level movement logic. Its methodological advantages, like agent heterogeneity, environmental feedback, and spatiotemporal autonomy, offer a new perspective for breaking away from traditional patch statistics.

By simulating species movement, adaptive foraging, or social-environmental negotiation processes, ABS can reveal hidden ecological bottlenecks and service gaps in fragmented urban systems and suggest “stepping stones” [20] for bridging the network of connected habitats for species, especially pollinators, facilitating their movement and promoting genetic exchange among plant populations. This gives the chance to effectively translate the simulation output into the discourse of ecological planning. In light of the growing severity of climate and biodiversity challenges, the integration of ABS into green continuity research facilitates more precise measurements and signifies a transition toward a more profound comprehension and the development of practical designs.

3. Methodology—Diagnosing and Reframing Urban Green Continuity Through Agent-Based Simulation

This section aims to consolidate the concept of “green continuity” proposed by IMM Design Lab, Politecnico di Milano, into a set of operational actions through dynamic process simulation from the perspectives of multiple-species groups using agent-based simulation (ABS).

3.1. From Concept to Measurement: Operationalizing Green Continuity

Simulation of green continuity focuses on integrating green-blue infrastructures (core areas, ecological corridors, and restoration nodes) into a systemic network that supports biodiversity, climate adaptation, and urban resilience. This approach emphasizes the importance of “morphology–performance interactions”, no longer viewing urban green spaces as a static collection of patches, but as a dynamic foundation supporting ecological processes, by proposing a diagnostic method based on behavioral simulation. To incorporate seasonal dynamics into the spatial analysis, this study employs urban monitoring data from two representative periods—spring (April) and late autumn (November) 2024—as input layers for simulation. By comparing pollinator activity patterns across these seasons, the model captures temporal variations in ecological mobility and reveals dynamic patterns of hidden urban corridors. Through incorporating ABS, a multi-agent model is developed to simulate the movement paths, resource perception, obstacle avoidance behavior, and finally, temperature responses of insects, small urban mammals, and birds within urban green space networks with different time slices. After obtaining ecological mobility patterns, the result unveils hidden green corridors embedded within fragmented urban landscapes, offering a multi-layered diagnosis of spatial continuity and seasonal variations in ecological behavior.

This study presents a detailed diagnosis of three selected urban zones in Italy—Lambrate (Milan), Bolognina (Bologna), and Ispra— representing three different spatial characteristics and administrative contexts. Lambrate is a highly populated district in a large compact city (Milan), Bolognina is a transformed zone designated as a cultural epicenter, while Ispra is a village in a rural environment. The method included processing both spatial and behavioral data, demonstrating the design logic and operational workflow of the ABS model. Based on the simulation results, this study develops an integrated diagnostic framework that captures both green continuity and urban climate adaptability. Combining species mobility with climate risk exposure, the methodology offers a novel approach to identifying ecological discontinuities and provides a grounded basis for guiding targeted green space interventions in urban planning.

Green continuity is defined through four measurable dimensions, as follows:

• Integration analysis of agent mobility network;
• Local connectivity analysis of agent moving paths;
• Agent aggregation probability;
• Spatial overlap of thermal vulnerability.

This operational approach transforms the abstract notion of green continuity into a set of spatially explicit, behavior-sensitive measurements and enables the identification of hidden ecological connections and environmental vulnerabilities within the urban fabric.

3.2. Study Area and Data Preparation: The Urban Mosaic of Milan, Boragna, and Ispra

The districts of Lambrate (Milan), Bolognina (Bologna), and Ispra were deliberately selected as case study locations in this research due to their unique ecological, spatial, and institutional characteristics, which provide contrasting yet complementary insights into urban green space continuity in local and medium-scale European urban environments.

Lambrate is a district in the northeast of Milan, where the urban fabric blends industrial heritage with characteristics of regenerative development. A major railway line runs through the center of the site, creating a strong physical and ecological division. Bolognina is a district in northern Bologna, adjacent to the Central Station (Stazione Centrale). In recent years, it has transformed into an urban regeneration pilot zone through building renovations, public space upgrades, and community-driven governance. Its densely interwoven secondary streets form closed or semi-closed neighborhoods. While informal green spaces and potential ecological corridors exist along the railway lines, physical barriers result in low connectivity, creating a typical “high-potential, low-connectivity” fragmented urban area. Ispra, situated on the eastern shore of Lake Maggiore in northern Italy, represents a low-density urban settlement embedded within a rich ecological landscape. The urban morphology of Ispra is characterized by dispersed built forms, extensive peri-urban green areas, and high proximity to Natura 2000 protected zones.

From a comparative perspective, Lambrate, Ispra, and Bolognina offer distinct typologies of green infrastructure governance and spatial configuration. Lambrate provides the chance to examine the influence of grey infrastructure (railway) on environmental performance, Ispra emphasizes ecological adjacency and peri-urban continuity, while Bolognina embodies the dynamics of inner-city fragmentation and redevelopment. The inclusion of these cities thus enables a multiscalar and typologically diverse analysis of urban green space continuity, in an approach that aligns with recent calls in the literature for ecological urbanism to be grounded in regional heterogeneity and context-specific urban ecologies [85,86].

3.3. Agent-Based Simulation Design: Modeling Multi-Agent Ecological Interactions

The agent-base simulation (ABS) model used in this study was developed based on the Physarealm plugin for Rhino Grasshopper. Physarealm operates as a multi-agent system in which agents follow local gradients within a diffusive–decaying trail field, reinforcing paths through deposition; the emergent global network pattern arises as a self-organized outcome of these iterative local rules [87,88]. It pre-processes the geometry of urban components in the input layers (

Table 1. Data collection and resource links.

Data Layer	Data Resource	Repository Link
Administrative Boundary (in vectors)	Geoportale della Lombardia	https://www.geoportale.regione.lombardia.it/en-GB/home (accessed on 1 May 2024)
Volume (in vectors)	Geoportale della Lombardia	https://www.geoportale.regione.lombardia.it/en-GB/home (accessed on 1 May 2024)
Green Space (in vectors)	Regione Emilia-Romagna	https://geoportale.regione.emilia-romagna.it/ (accessed on 2 May 2024)
Normalized Difference Vegetation Index (NDVI)	Landsat 8–9	https://earthexplorer.usgs.gov/ (accessed on 22 January 2025)
Land Surface Temperature (LST)	Landsat 8–9	https://earthexplorer.usgs.gov/ (accessed on 22 January 2025)

) in QGIS and samples the attributes of urban green spaces through 50 by 50 m hexagons. Physarealm consists of modules including emitter, food, and environment, and the central processing core. The principle is that by locating the emitters’ position through sampling points, the emitters send agent points to move around to find food points within the environment (Figure 1b).

The total distance traveled by each agent must not exceed the $R_{Group}$ which is differentiated based on the researched groups, like bees or birds (Formula (1)); otherwise, the agent’s movement fails and disappears. Since Physarealm is a dynamic simulation, once the agents’ movement stabilizes, time slices can be taken for further analysis of the movement network and location.

$$\sum _{i=1}^n{R}_i^{\left(A\right)}\le R_{Group}$$

(1)

Based on the categorization of agent groups and required input layers,

Table 2. Input parameters of ABS in Physarealm.

Input Parameter		Lambrate	Bologna	Ispra
General Setting	Food–Agent Ratio (TrRat)	50	50	50
	Total Green Space Area	1,220,427 m2	372,985 m2	6,738,090 m2
	Starting Agent Numbers (PopS)	2000	1000	1000
	Number of Green Space Sample Points in Spring (Food and Emitter)	2974	957	4113
	Number of Green Space Sample Points in Autumn (Food and Emitter)	2164	665	3093
	Number of Building and Railway Geometries (Obstacles)	1513	1372	3109
Agent Groups	Radius of Ground-to-Canopy Pollinator Movement (Ddis = RGround)	200 m	200 m	200 m
	Radius of Avian Pollinator Movement (Ddis = RAvian)	500 m	500 m	500 m

summarizes the corresponding parameters for running the ABS model in Phyrarealm. The starting agent number (PopS) was locally defined in proportion to the spatial extent of each study area and dynamically evolved as the simulation processed stably. Since pollinator richness tends to peak at intermediate NDVI values (approximately 0.4–0.6), aligning with the intermediate productivity hypothesis [89] under static conditions, the NDVI data from Landsat 8–9 were used to estimate the initial distribution and starting agent number (PopS), at a resolution of 50 × 50 m. Instead of statically analyzing the initial PopS, the ratio between the remaining agent points and PopS was measured to understand the support/exclusion of the urban context.

3.4. Modeling Multi-Species Movement: Ground and Avian Pollinators as Agents

To analyze urban green continuity, this study categorizes the pollinator agents into two main types: ground-to-canopy species (bees, butterflies, small mammals) and avian agents (common urban birds). These pollinators play a key role in maintaining the diversity and resilience of urban ecosystems. Ground pollinators, such as bees and butterflies, primarily engage in local plant pollination and support the reproduction of plant communities. Small urban mammals such as squirrels also demonstrate high spatial mobility in search of food, contributing indirectly to ecological interactions in the ground strata in most seasons. Urban birds, with their larger activity radius, enable long-distance seed dispersal and ecological connectivity between green patches. To accurately reflect their ecological interactions in fragmented green networks, this study adopts differentiated radii, which helps to deepen the understanding of urban green continuity from a multi-species perspective, identifying “hidden ecological corridors” and connection gaps that are difficult to reveal in single-species models.

To define species-specific foraging radii in urban pollinators, the key determinants are species identity, habitat quality, and the spatial distribution of resources (Table 2) [90]. The mobility radius of most bees and butterflies is between 200 and 500 m [91]. Urban roads and buildings restrict pollinators’ movement, leading to a reduction in their activity radius [92]. Therefore, in this study, the activity radius R_Ground for the ground pollinator group was set to 200 m. The activity radius of birds varies even more, depending on diet, body size, and habitat requirements. Small songbirds such as sparrows and titmice typically have an activity range of 100–500 m [93], while medium-sized birds such as starlings and turtle doves can have an activity radius of 1–3 km [94]. Considering that the research areas were at district and village scale, which should be localized, this study set the activity radius R_Avian for avian pollinators to 500 m. The output was shown as polygon networks and points, which were further analyzed by Space Syntax (DepthMapX) for a better understanding of the integration and local connectivity.

In the spatial processing stage, DepthmapX segment analysis was used to topologically characterize the continuity backbone extracted from the ABS outputs, to evaluate pollinator networking performance [95]. The core idea treats segments as nodes and their intersections as edges (which is different from conventional graph theory), computing accessibility by topological steps rather than Euclidean distance. Mean depth (Formula (2)) denotes the average topological distance from a given segment to all other segments in the network; lower values indicate higher global Integration, meaning the location is more easily reached and contributes more strongly to overall green continuity. This enables the identification of high-integration hotspots (candidate primary corridors/backbone) and low-integration weak belts (potential breaks or edge zones).

$$M{D}_i={\displaystyle \frac{1}{\left(k-1\right)}}\sum _k{d}_{ik}$$

(2)

where $k$ stands for the number of segments in the system, and $d$ stands for the depth. $\sum _k{d}_{ik}$ stands for the total depth sum from the root node $i$ to all other spaces in the researched zone.

Regarding connectivity (Formula (3)), the degree of a segment (number of intersecting segments) reflects local connectivity density and potential activity clustering. It is used to delineate pollinators’ local activity ranges and to identify micro-scale bottlenecks and stepping stones.

$$C_i=\sum _{j=1}^k{a}_{ij}$$

(3)

where $a_{ij}$ refers to the adjacency value between node $i$ and node $j$, equal to 1 if $i$ and $j$ intersect (or directly connect), and 0 otherwise; $k$ is the total number of segments in the system.

Due to the involvement of temporal variables, the initial number of pollinators changes with time and food quality, resulting in seasonal variations of aggregation patterns. Therefore, agent aggregation possibility (AAP) was introduced to quantify the degree of environmental support for pollinators and to identify intervals where human intervention may be required. This indicator highlights zones with higher potential for overlapping reach and shared use, offering insight into spatial equity, green space competition, and resource concentration. It reveals spatial differences between the ground and avian agent groups, as well as between different seasons, reflecting how spatial layout influences reachability and potential behavioral clustering. By capturing implicit attraction and balanced usage patterns beyond the network, it provides support for the structural optimization of green infrastructure and the balancing of ecological services.

The calculation method is given in Formula (4).

$${AAP}_i={\displaystyle \frac{n_i}{N}}$$

(4)

where $N$ represents the total number of remaining agents in the system, while $n_i$ denotes the number of agents whose movement radii cover the i-th hexagon.

3.5. From Simulation to Diagnosis: Uncovering Urban Ecological–Thermal Synergies Through Spatial Cross-Mapping

This section aims to translate the outputs from the ABS into actionable spatial diagnostics with land surface temperature (LST) weighted. To identify spatial patterns where ecological activity intensity overlaps with urban heat stress, a 4 × 4 cross-classified overlay model was developed. First, land surface temperature (LST) derived from Landsat 8–9 imagery, using the thermal infrared band (Band 10), was classified into four levels. Cold zones (<15 °C) are where most pollinating insects (such as bees, butterflies, and moths) begin to experience difficulty flying, which essentially ceases below 10 °C [96]. Optimal zones (15–30 °C) are where pollen release is vigorous and insect metabolic efficiency is highest, making it the best temperature range for green infrastructure to provide ecological services [97]. Thermal stress zones (30–40 °C) are where heat stress begins to affect insects and birds; foraging efficiency decreases significantly beyond 35 °C, with behavioral thermoregulation becoming evident [98]. Thermal risk zones (>40 °C) are where the urban heat island (UHI) effect is most severe [99].

Agent network density was defined as the frequency each grid cell was traversed by simulated agent paths, reflecting spatial preferences within the agent-based movement network. Given the highly skewed and long-tailed distribution of this result, which is common in path-based simulations, an exponential binning method was used to classify network density into four relative levels: isolated, low, medium, and high. Notably, these thresholds were not standardized across study areas but calculated locally for each case, considering the specific PopS of agents and the spatial characteristics. This locally adaptive classification enables a meaningful diagnosis of intra-case spatial intensity while avoiding misleading cross-case comparisons [100,101].

The resulting 4 × 4 spatial grid typology reveals diverse combinations such as cold–high activity, hot–low activity, etc., to support targeted ecological interventions, UHI mitigation, and urban resilience planning.

3.6. AI-Augmented Diagnosis: Interpreting Spatial Agency Through the SOFIA Agentic Workflow

Agent-based simulation (ABS) for green continuity developed through QGIS and Grasshopper offered a refined and high-resolution diagnosis. It visualized urgent environmental challenges in urban systems, such as ecological fragmentation, spatial inequality, and thermal risk, highlighting zones in need of design intervention. However, this level of analytical granularity introduced a new bottleneck because each hexagon contained complex, multi-attribute spatial data layers, making it difficult to derive fine-grained design actions through conventional urban design workflows. This gap between high-resolution spatial diagnosis and implementable planning decisions became increasingly evident, especially under the constraints of human interpretive capacity.

To address this gap, SOFIA (Systemic Options Framework for Intelligent Activation) was developed by IMM Design Lab. SOFIA responds through a modular AI-based workflow that interprets GeoJSON outputs from ABS and GIS platforms and generates context-aware, narrative planning recommendations. Built on the N8N platform, the AI agent SOFIA consists of several core modules (Figure 2). The process begins with geospatial preprocessing in QGIS, where spatial attributes are structured and exported as GeoJSON inputs. The input module parses these features and attributes and then followed by a retrieval-augmented generation (RAG) module, where DeepSeek LLM is paired with a Pinecone-based local vector database to retrieve the design literature and case precedents as a local knowledge base.

The interpretive reasoning engine implemented via OpenAI’s API serves as the cognitive core, integrating spatial attributes and retrieved references to construct semantically grounded spatial judgments. The output module translates this reasoning into a structured narrative for each spatial unit, including actionable suggestions, cited sources, case analogies, and stakeholder-oriented strategies. Beyond its main pipeline, SOFIA incorporates two embedded knowledge bases; the urban ontology module encodes typologies, spatial semantics, and historical planning contexts, and the IMM knowledge engine operationalizes IMM’s indicators, principles, and transformation logic to anchor the system’s outputs in methodological coherence.

4. Results—Spatial Manifestations of Urban Green Continuity

4.1. Agent-Based Movement Patterns: Identifying Ecological Corridors and Spatial Accessibility

By visualizing agent paths derived from ABS, high-frequency corridors are visualized in Figure 3, Figure 4, Figure 5, Figure 6, Figure 7 and Figure 8, revealing potential ecological linkages and accessibility patterns for different species. These movements serve as the behavioral basis for diagnosing structural gaps in urban green continuity.

Figure 3 shows the spatial integration of the agent movement, processed by using DepthMapX, which illustrates the overall accessibility structure across the whole system. In general, the Autumn groups (Figure 3d–f,j–l) show a higher or equal level of integration due to the reduction of high NDVI zones, which leads to a concentration of pollinators’ movement. As seen from the diagrams, the high-integration zones (yellow lines) cover smaller areas but have increased density, which is most obvious in the case of Ispra. This is also consistent with the research of Zaninotto et al. (2021) regarding the seasonal activity habits of pollinators [102].

Among all three cases, Lambrate was found to be the least centralized; its high integration zone is present in the mid-eastern open space, at 2.7–3.5 (high efficiency), indicating a potential for introducing landing islands for insects. The railway track acts as a major barrier, dividing the movement of ground pollinators into eastern and western segments, thereby reducing overall continuity. However, scattered green patches within the rail yard offer limited stepping-stone opportunities, allowing partial cross-barrier movement for the ground-to-canopy pollinator group and preserving the integrity of the network.

The avian pollinator group possesses a higher moving radius (R_Avian = 500 m), which allows agents to move longer distances to reach neighboring resources. The railway in Lambrate works as a barrier for birds as well, and the patches of green space are insufficient for the landing of pollinators. In the Bolognina case, the primary green spaces are situated along the district boundary, embracing the main built-up area and linked by linear trees. In spring, bird movements form a more evenly distributed spatial network, whereas in autumn, this network becomes more centralized around the urban core.

The case of Ispra presents a different pattern (Figure 3c,f,i,l). With the highest green coverage (72.6%) among the three study areas, the movement paths of pollinators in both the ground and avian groups exhibit a more balanced spatial distribution in both spring and autumn. As previously discussed, the decline in vegetation within residential gardens during autumn, reflected by a decrease in NDVI, leads pollinators to concentrate around forested areas and the southern agricultural lands, which results in the high integration zone shifting slightly southward.

4.2. Local Structural Connectivity: Unpacking Micro-Scale Mobility from Pollinator Networks

In the previous section, the overall ecological movement trends and major corridors were identified by general integration analysis, highlighting the macro-scale mobility patterns of pollinators. However, such global analysis alone is insufficient to reveal the behavioral reaction of local movement in micro-scale spatial structures. To address this, this section also includes local connectivity analysis by space syntax, to examine the local structural accessibility of the pollinator movement network, focusing on the role of path intersections within fragmented green infrastructures. The connectivity value, calculated using DepthMapX (Formula (3)), represents the link’s degree of spatial intersection and local path density, reflecting a location’s permeability, accessibility, and navigability within the fine-scale ecological network. This result is particularly relevant for ground-level pollinators, whose movement depends on the availability of continuous habitat links and spatially legible pathways.

The connectivity analysis in Figure 4 highlights localized spatial clusters where pollinator movement intensifies. In Bolognina, a clustering of movements occurs adjacent to rather than within green spaces suggests the likelihood of pollinators navigating between closely positioned patches, in contrast to the more interior-focused patterns observed in Lambrate. In Ispra, the southwestern corner exhibits a pronounced aggregation of movement relative to built-up areas, indicating the presence of a potential high-attraction zone for pollinator activity.

In the local connectivity analysis, Lambrate and Bolognina show a more balanced network distribution with a lower density of network but a bigger coverage of high-value zones. In Ispra, a second high-value zone appears in the middle of Figure 4i. The aggregation of movement clusters presents a double-edged condition. On one hand, it indicates frequent pollinator passage and visitation; on the other hand, the concentration of high-value zones reveals an uneven distribution within the whole movement network, reflecting the underlying fragmentation of urban green space.

Due to seasonal shifts, all three cases exhibit a clear pattern of centralization in November, which aligns with the spatial integration patterns discussed earlier. Compared to natural forests, spaces within residential zones show a pronounced decline in plant diversity during autumn and winter, as reflected by the reduction in NDVI values. In Lambrate, this effect is particularly evident. The clustering observed on the west side of the railway in spring disappears in November, while the eastern cluster becomes more intense. In Bolognina, spatial clusters demonstrate increased centripetal tendencies; several scattered clusters present in spring become consolidated into a more compact group in the mid-southern constructed zone. In contrast, Ispra, characterized by the largest and most evenly distributed green coverage, shows a slight reorganization of its initially balanced clusters into two dominant groupings: one centered in the southern forest zone and the other near the lakeside.

Figure 4. Multi-agent mobility network analysis—local connectivity for both pollinator groups. Brighter colors indicate higher local connectivity, reflecting denser intersections and potential activity clusters, whereas darker colors denote lower connectivity and more isolated segments. (a–c) and (d–f) depict the Ground-to-Canopy agent network in April and November 2024, respectively. (g–l) depict the Avian agent network for the same respective periods. The specific research area corresponding to each panel is denoted in its sub-figure title.

Figure 4. Multi-agent mobility network analysis—local connectivity for both pollinator groups. Brighter colors indicate higher local connectivity, reflecting denser intersections and potential activity clusters, whereas darker colors denote lower connectivity and more isolated segments. (a–c) and (d–f) depict the Ground-to-Canopy agent network in April and November 2024, respectively. (g–l) depict the Avian agent network for the same respective periods. The specific research area corresponding to each panel is denoted in its sub-figure title.

4.3. Spatial Coverage and Aggregation Probability: Quantifying Reachability of Green Spaces

To complement the dynamic path-based simulation, this section presents the results of a spatial coverage analysis using the agent aggregation probability (AAP) indicator. AAP quantifies the proportion of agents whose reachable radius includes each grid cell, based on circular buffers generated from agent origin points. Figure 5 shows the spatial distribution of the ground and avian groups after the ABS simulation reaches a steady state, with corresponding buffer zones generated based on their respective activity radii (R_Group in Table 2). By counting the number of agent buffers covering each hexagon cell and calculating the ratio of this number to the total number of remaining agents in the group, the probability of the cell being visited is obtained (Formula (4)). This ratio ensures comparability across different cases and avoids biases caused by different PopS.

A high aggregation probability (e.g., >40%) indicates an intensive visitation of multiple agents in this area, suggesting a higher ecological functionality or spatial attractiveness, potentially due to green resources, terrain accessibility, or network connectivity. Conversely, a lower overall aggregation value (e.g., <20%) shows that agents are distributed more evenly, without forming highly overlapping hotspots. This phenomenon can be interpreted as a positive signal indicating a relatively balanced green space layout and the absence of significant spatial fragmentation.

Figure 5. Agent aggregation probability analysis. Higher probabilities (e.g., >40%) indicate intensive agent visitation and greater ecological attractiveness, while lower values (e.g., <20%) reflect more evenly distributed movement without clear hotspots. (a–c) and (d–f) depict the Ground-to-Canopy agent statistics in April and November 2024, respectively. (g–l) depict the Avian agent statistics for the same respective periods. The specific research area corresponding to each panel is denoted in its sub-figure title.

Figure 5. Agent aggregation probability analysis. Higher probabilities (e.g., >40%) indicate intensive agent visitation and greater ecological attractiveness, while lower values (e.g., <20%) reflect more evenly distributed movement without clear hotspots. (a–c) and (d–f) depict the Ground-to-Canopy agent statistics in April and November 2024, respectively. (g–l) depict the Avian agent statistics for the same respective periods. The specific research area corresponding to each panel is denoted in its sub-figure title.

As shown in

Table 3. Summary statistics of agent aggregation probability and retention rate across study areas.

Research Case		Time	Q1 (%)	Median (%)	Q3 (%)	Mean (%)	Retained Agents (N)	Retention Rate (%)	Active Zone Rate (%)
RGround = 200 m	Lambrate	April	1.18	2.52	4.87	3.35	595	29.75	82.51
		November	1.11	3.11	6.44	4.48	459	22.95	78.11
	Bolognina	April	2.3	4.2	6.3	4.31	650	65.00	95.48
		November	2.32	4.64	7.45	5.00	604	60.40	97.37
	Ispra	April	1.07	1.68	2.29	1.76	655	65.50	90.37
		November	0.9	1.7	2.5	1.78	636	63.60	89.06
RAvian = 500 m	Lambrate	April	6.14	14.71	25.7	16.78	537	26.85	99.58
		November	5.99	14.98	28.8	18.42	434	21.70	98.03
	Bolognina	April	14.74	22.77	29.93	22.15	685	68.50	100.00
		November	14.57	23.22	34.75	24.41	659	65.90	100.00
	Ispra	April	5.97	8.06	10.16	7.82	620	62.00	98.49
		November	4.75	8.2	11.15	7.99	610	61.00	97.65

Retained Agents (N): Number of agents remaining after the simulation reaches equilibrium. Retention Rate (%): Percentage of agents retained relative to the initial agent count. Active Zone Rate (%): Proportion of hexagonal zones covered by agent activity.

, the aggregation probability exhibits significant spatial differentiation across study regions and radius settings. During the processing of ABS, the number of active agents gradually decreases as they respond to environmental conditions and the availability of food, eventually reaching a dynamic stable state of emission and disappearance. The retained agents value represents the total number of agents that persist in this state. The retention rate indicates the percentage of agents that remain compared to the initial agent population, reflecting the spatial and ecological supportiveness of the environment. Meanwhile, the active zone rate captures the spatial extent of the simulation by calculating the proportion of hexagons with agent activities, offering insight into the dispersion and territorial reach of ecological movement.

Across all study areas and both agent groups (R_Ground and R_Avian), November generally shows higher median and Q3 values for agent aggregation, indicating more concentrated clustering behavior in colder seasons. However, the retention rate (%) in November tends to decrease slightly, especially in denser urban areas like Lambrate and Bolognina, suggesting that fewer agents are retained due to environmental limitations. Meanwhile, active zone rates remain consistently high across both months, especially for the avian pollinator group.

Taking the April scenario as example, in the small radius group (200 m), Bolognina and Ispra demonstrate highly similar aggregation probability levels (average of 4.31% and 1.76%, respectively, with remaining agent proportions of 65.00% and 65.50%), while Lambrate’s mean and retention rate are both significantly lower (mean = 3.35%, retention rate = 29.75%), reflecting limited accessibility due to insufficient green space service coverage and dispersed environmental structures. In the large-radius group (500 m), the overall aggregation probability significantly increased, and the differences between groups became more pronounced. The average probability of the Bologna group reached 22.15%, with Q3 as high as 29.93%, indicating a significant clustering trend. This high value is not due to the adsorption effect of large, concentrated green spaces, but rather because the green spaces are distributed surrounding the building complex. This allows agents to approach these edge green spaces from different directions, forming aggregated and overlapping paths, thereby revealing a multi-point concurrent aggregation network. This spatial structure compensates for the disadvantage of insufficient green space area, demonstrating strong spatial flexibility and service connectivity.

Figure 6. Boxplot of Seasonal Variation in Agent Aggregation Probability.

Figure 6. Boxplot of Seasonal Variation in Agent Aggregation Probability.

These trends are further validated in the box plots in Figure 6. The overall box plots for the larger radius group are significantly elevated, with Bologna exhibiting the highest median and upper quartile values, which reflects its stronger aggregation and ecological viscosity. In contrast, Lambrate exhibits a downward shift in the box plot and a dispersed point distribution in both radius groups, further validating the diagnostic findings of weaker spatial service capacity and insufficient agent network support.

From spring to late autumn, the spatial behavior of agents exhibits significant differences due to seasonal changes in green space quality and vegetation density. Figure 7 illustrates the trends in pollinator aggregation across different periods by calculating the ΔPossibility (November–April) for each hexagon. Overall, changes in the ground group (R_Ground) are relatively mild, with aggregation and dispersion areas primarily exhibiting local fluctuations. In contrast, aggregation changes in the avian group (R_Avian) are more pronounced, with the range of concentration trends significantly expanding.

Figure 7. Spatial difference in agent aggregation possibility (AAP) (April to November) across percentage ranges. Green areas indicate an increase in aggregation possibility (>0), representing zones where agents entered from April to November, while red areas (<0) denote a decrease, indicating zones of agent departure. (a–c) illustrate the AAP variation of the Ground-to-Canopy agent group, while (d–f) show the Avian agent group. The specific research area for each panel is indicated in its sub-figure title.

Figure 7. Spatial difference in agent aggregation possibility (AAP) (April to November) across percentage ranges. Green areas indicate an increase in aggregation possibility (>0), representing zones where agents entered from April to November, while red areas (<0) denote a decrease, indicating zones of agent departure. (a–c) illustrate the AAP variation of the Ground-to-Canopy agent group, while (d–f) show the Avian agent group. The specific research area for each panel is indicated in its sub-figure title.

In the Bolognina and Lambrate areas, aggregation zones cluster toward large contiguous green spaces in the south and east, respectively, indicating that fragmented, patched green spaces no longer hold a significant attractiveness advantage in autumn, and ecological behavior tends to occur in structurally intact, highly contiguous green spaces. This spatial convergence trend reflects agents’ stronger aggregation preferences and sensitive responses to habitat quality under autumn ecological stress.

Meanwhile, the red dispersal areas are mostly located at the edges of neighborhoods or high-density blocks, further validating the influence of seasonal ecological condition changes on aggregation behavior. This reveals the significant intervention of seasonal dynamics on the distribution of ecological behavior and highlights the spatial amplification of green space structure at different scales in shaping pollinator behavior.

4.4. Thermal–Ecological Cross-Mapping: Diagnosing the Coupling of Green Continuity and Urban Heat

This section presents a 4 × 4 spatial diagnostic map based on the superposition of agent movement network intensity and land surface temperature (LST) to identify synergistic or fragmented patterns between green space connectivity and urban heat exposure. LST was classified into four thermal zones—cold (<15 °C), optimal (15–30 °C), stressed (30–40 °C), and risk (>40 °C)—reflecting escalating levels of urban heat stress. Given the highly skewed and long-tailed nature of agent path data, network density was binned into isolated, low, medium, and high levels using a locally adaptive exponential classification. Through a cross-analysis of network density and remotely sensed heat zone divisions, the resulting cross-mapping reveals how ecological behavior clusters coincide or diverge from thermal conditions in each study area, offering spatial insights into climate–ecological mismatches and opportunities for targeted resilience planning.

Figure 8. Thermal–ecological typology and spatial distribution across Lambrate, Bolognina, and Ispra.

Figure 8. Thermal–ecological typology and spatial distribution across Lambrate, Bolognina, and Ispra.

Figure 8 presents a seasonal diagnostic combining land surface temperature (LST) and agent movement network density. Each hexagon is classified into one of 16 bivariate categories from “Dormant Patch” to “Overload Core” based on its thermal condition (cold to risk) and movement density (isolated to high). This bivariate classification reveals how ecological behaviors interact with climatic stress across both time and space. Seasonal changes in thermal conditions or movement patterns can be interpreted either through shifts in scatterplot distribution or the corresponding spatial maps below.

The lower panel illustrates the spatial distribution of LST–network coupling typologies across seasons. In April, both Lambrate and Bolognina show pervasive high-temperature conditions (30–45 °C), with prominent urban heat island effects. In these contexts, many agent aggregation areas fall into ‘thermal conflict’ zones, where pollinator activity persists despite thermal stress. In contrast, Ispra presents the most favorable ecological conditions; its “optimal” zones align well with areas of high agent movement, forming a continuous “green corridor”, and the extents of “stressed” and “risk” zones are notably smaller. In November, seasonal divergence becomes more pronounced. Bolognina undergoes a sharp thermal decline, with many areas falling into the “cold” category (<15 °C), which severely limits ground-level pollinator activity. Meanwhile, Lambrate and Ispra both retain considerable “optimal” zones. Ispra continues to exhibit a high spatial match between agent activity and thermal suitability, suggesting a stable ecological-thermal fit. In contrast, Lambrate reveals a fragmented structure; pollinators appear more aggregated, resulting in a spatial pattern dominated by “Untapped Potential” and “Weak Link” categories. These seasonal and structural differences highlight distinct vulnerabilities and adaptation potentials across the three urban contexts.

The identified continuity patterns align with the green corridors outlined in the Milan PGT 2030 masterplan (the Green Rays/Raggi Verdi project) [103] and show broad consistency with the biodiversity intactness index (BII) map at a 1 km resolution [104]. While these comparisons cannot be quantified due to scale mismatches between neighborhood-level diagnostics and city- or regional-scale indicators, they nevertheless provide external validation of the plausibility of the findings.

4.5. Evaluating Urban Diagnostic Interpretations: Traditional Expertise, SOFIA Agent, and General-Purpose Large Language Model

In this sub-section, a cross-platform comparison reveals how different analytical logics interpret the same spatial indicators in varying ways, leading to diverse intervention strategies.

Table 4. Multi-source diagnostic comparison under IMM: Manual, SOFIA, and ChatGPT analysis.

Location	Cell ID	Platform	Catalyst	Design Ordering Principle (DOP)	Action	Case Reference	Academic Support	Stakeholders
Lambrate	1192	Manual Analysis	Porosity	DOP 01—Balance the ground land-use of the city considered as a whole ecosystem consisting of both natural and socio-economic components.	Action 1.4—Reduce urban land cover and reduce urban soil sealing.	NbSouth via retrofitting for Climate Adaptation, Brazil	Villa et al. (2020) [105]	Local municipality, water management department, academic researchers
		SOFIA	Permeability	DOP 05—Create connected open space system, activate urban metabolism.	Action 5.3—Develop a green infrastructure strategy that identifies priority areas for new green infrastructure.	Biotope Verbund System, Berlin, Germany	Kowarik et al. (2016) [106]	Municipal green space department, local residents, and entomology researchers
		ChatGPT 4.0 Web	Continuity	DOP 05—Create connected open space system, activate urban metabolism,	Action 5.4—Design a connected network that integrates green infrastructure, such as green roofs, green walls, bioswales, rain gardens, and nature-based solutions, with the urban built environment.	Cheonggyecheon Stream restoration, South Korea	Yu et al. (2020) [13]	Municipal planners, landscape architects, community associations
	3231	Manual Analysis	Accessibility	DOP 08—Change from multi-modality to inter-modality concept.	Action 8.1—Develop and implement a connected, systems approach to support inter-modality. This includes building bike lanes, pedestrian walkways, transit lanes, and dedicated parking areas for bike and car-sharing.	Tertiary Forest Parking Lot, France	Parison et al. (2023) [107]	Municipal planners, Academic Institutions, Meteorological and Environmental Monitoring Agencies, Public User Feedback
		SOFIA	Effectiveness	DOP 07—Balancing the public transportation potential.	Action 7.2—Rebalance public service distribution through accessibility mapping.	Biodiversity Action Plans (BAP), UK	Goddard, Dougill, and Benton (2010) [108]	Ornithological societies, park managers, urban farmers
		ChatGPT 4.0 Web	Diversity	DOP 04—Make biodiversity an important part of urban life.	Action 4.1—Enhance ecological restoration in cities by restoring degraded and underutilized land to improve local ecosystem health and wildlife habitat.	Biblioteca degli Alberi, Milan, Italy	Fahrig (2003) [9]	Ecologists, urban forest planners, biodiversity NGOs
Bolognina	462	Manual Analysis	Diversity	DOP 04—Make biodiversity an important part of urban life.	Action 4.7—Measuring and monitoring biodiversity.	Valona Strategic Planning, Albania	IMM Internal Source [5]	Local municipality, academic researchers, and urban planners
		SOFIA	Continuity	DOP 04—Make biodiversity an important part of urban life.	Action 4.1—Enhance ecological restoration in cities by restoring degraded and underutilized land to improve local ecosystem health and wildlife habitat.	Green Axes, Spain	Pino and Marull (2012) [109]	City planners, botanical gardeners
		ChatGPT 4.0 Web	Continuity	DOP 05—Create connected open space system, activate urban metabolism.	Action 5.1—Create a connected, continuous, and integrated system of urban open green spaces.	Green Belt of Vitoria-Gasteiz, Spain	Parker and Simpson (2018) [1]	Ecological planners, municipal green departments, mobility engineers
	1528	Manual Analysis	Porosity	DOP 01—Balance the ground land use of the city considered as a whole ecosystem consisting of both natural and socio-economic components.	Action 1.4—Reduce urban land cover and reduce the urban soil sealing.	NbSouth via retrofitting for Climate Adaptation, Brazil	Villa et al. (2020) [105]	Local municipality, water management department, academic researchers
		SOFIA	Interface	DOP 03—Balance the distribution of key types of uses and foster multifunctional spaces.	Action 3.4—Encourages multi-functional buildings.	Chicago Nature and Wildlife Plan, United States	Albrecht and Haider (2013) [110]	Property owners, architects, beekeeping associations
		ChatGPT 4.0 Web	Interface	DOP 06—Promote walkability, cycling and reinforce their integration with public transportation.	Action 6.5—Optimize spatial connectivity.	Pollinator Pathway, Seattle, United States	Crooks and Heppenstall (2012) [101]	Urban designers, pollinator conservation NGOs, active mobility advocates
Ispra	5610	Manual Analysis	Accessibility	DOP 08: Change from multi-modality to inter-modality concept.	Action 8.1—Develop and implement a connected, systems approach to support inter-modality. This includes building bike lanes, pedestrian walkways, transit lanes, and dedicated parking areas for bike and car-sharing.	Tertiary Forest Parking Lot, France	Parison et al. (2023) [107]	Municipal planners, academic institutions, meteorological and environmental monitoring agencies, public user feedback
		SOFIA	Diversity	DOP 01—Balance the ground land-use of the city considered as a whole ecosystem consisting of both natural and socio-economic components.	Action 1.5—Implement green open spaces like community gardens, urban farms, and pocket parks to reduce soil sealing, mitigate urban heat islands, absorb stormwater, provide wildlife habitats, and offer recreational opportunities for urban residents.	Munich’s “Blühstreifen” (Flowering Strips) Initiative, Germany	Hülsmann et al. (2015) [111]	Agricultural extension services, schools, nature reserves
		ChatGPT 4.0 Web	Proximity	DOP 03—Balance the distribution of key types of uses and foster multifunctional spaces.	Action 3.2—Prioritize compact, mixed-use, and walkable development patterns.	Paris’ Vision for 15-min City, France	Moreno et al. (2021) [112]	Local planners, mobility departments, citizen cooperatives
	8814	Manual Analysis	Porosity	DOP 04—Make biodiversity an important part of urban life.	Action 4.6—Integrate both green and non-green infrastructures in a holistic landscape approach.	Urgenda Case, Netherlands	Matthijs Jan Terstegge (2020) [113]	Local municipality, academic researchers, and urban planners
		SOFIA	Proximity	DOP 02—Implement Permeability to facilitate urban flow.	Action 2.2—Design Pedestrian-friendly streets: Incorporating the pedestrian and cycling paths with open spaces and greenery.	Honey Highway, Netherlands	Scheper et al. (2014) [114]	Transport agencies, utility companies, environmental NGOs
		ChatGPT 4.0 Web	Effectiveness	DOP 07—Balancing the public transportation potential.	Action 7.3—Activate a joint land use–transit model for integrating land use and transit planning.	Freiburg-Vauban District, Germany	Beatley (2016) [85]	Regional mobility agency, urban development board, sustainability task force

presents a comparative analysis of research interpretations by traditional urban planning, the SOFIA agent, and a general-purpose LLM (ChatGPT 4). Both the manual analysis and SOFIA are grounded in the IMM theoretical framework, resulting in outputs that are more accurate and actionable. SOFIA leverages the querying capacity of its AI core to significantly reduce analysis time, offering more rapid and comprehensive support for planning projects. In contrast, manual analysis relies heavily on the individual knowledge and experience of planners and designers, with a strong dependence on previously published works and projects. ChatGPT, drawing on online knowledge, complements this by providing external case references and academic sources not yet integrated into local databases, highlighting areas for improvement in both human-driven processes and SOFIA’s locally maintained knowledge base.

In the hexagon ID: 1528 (Bolognina), all three parties identified an increasing ecological potential (with a significant increase of the Avian Agent activities in November), but the area still lacks actual green spaces. The manual, SOFIA, and ChatGPT analyses positioned interface as the catalyst to stimulate the design intervention, emphasizing the need for spatial designs along street boundaries and building edges to enhance accessibility and ecological connectivity. SOFIA further analyzed the area from the perspectives of species behavior and seasonal dimensions, identifying a lack of bird resting and foraging infrastructure in the fall. It proposed vertical ecological interventions through rooftop greening, highlighting its capacity to synthesize multi-source knowledge.

At hexagon ID: 1192 (Lambrate), while both the human-made and ChatGPT approaches focused on repairing surface fragmentation and spatial discontinuity from the perspectives of porosity and continuity, SOFIA shifted the focus to the microstructural level of permeability. It identified high temperatures (LST) and low NDVI as forming an ecological “thermal resistance barrier”, suggesting the creation of a micro-permeability network through shaded small green spaces and linear corridors. This diagnostic logic emphasizes the coupling between spatial thermal environments and ecological behavior, serving as a powerful complement to traditional macro-connectivity approaches.

It should be noted that this study selected only six typical points for the comparative analysis (Table 4), intending to verify the consistency of diagnostic logic. In fact, SOFIA’s core advantage lies in its efficiency in processing large-scale spatial data and its ability to integrate multi-source knowledge. When the analysis scope expands to the urban scale (e.g., thousands of grid/hexagon cells), traditional manual methods face time-consuming and labor-intensive bottlenecks. SOFIA can rapidly generate structured intervention recommendations through automated workflows and combine local knowledge bases with case repositories to enable cross-scale reasoning. This capability will be further validated in subsequent research through full-region diagnostic cases.

5. Conclusion: Unveiling Design Insights from Hidden Green Continuity

This study takes “green continuity” as the core concept, addressing the multiple challenges posed by the fragmentation of urban green spaces, particularly the degradation of ecosystem services, spatial disparities in climate heat risks, highlighting the complex relationship between behavioral pathways, heat exposure, and ecological embeddedness in ecosystems, addressing the core tensions in the transition from landscape ecology to design-based urban ecology. By integrating agent-based simulation (ABS), land surface temperature data, and AI agentic workflow, a spatial diagnostic framework has been developed to identify urban ecological potential and structural inequalities.

The results reveal that ecological movement paths formed by agents often diverge from the conventional layout of green areas. These paths indicate a hidden green network that crosses multiple functional zones and is largely absent from current planning systems. When overlaid with LST data, many high-frequency paths were found to coincide with areas of thermal vulnerability, highlighting a spatial overlap between ecological intensity and heat exposure. This co-occurrence suggests a coupled system where ecological needs and climate vulnerability interact, indicating the gaps in current urban heat management strategies that overlook ecosystem structures.

The three case areas demonstrated the applicability of the IMM framework in translating complex ecological–thermal patterns into actionable intervention clues. Based on the behavioral density and risk overlaps revealed by ABS and spatial cross-diagnostics, a corresponding logic between types and strategies was formed. This structured identification method provides new tools and cognitive pathways for urban governance guided by ecological equity, breaks through the traditional ecological assessment logic based on land use types or green space boundaries, and supports a more mechanism-based “ecological behavior-based planning unit division.”

Methodologically, this study expands the scope of the “agent” concept. On one level, non-human movement patterns were modeled by agentic points to reconstruct their potential activities in a complex urban context. On the other level, the AI agent SOFIA was employed to interpret large-scale diagnosis results, producing structured intervention suggestions and reference cases. This dual-agent framework enhances the operational depth of ABS and illustrates a potential direction for “AI for science” (AI4S) in urban spatial research.

There are, however, several limitations. This study focused on three urban areas in Italy, which, although diverse in local characteristics, share similar environmental and urban contexts at the national scale. In this study, the input parameters of the ABS, such as starting agent numbers (PopS) and group activity radius, were determined based on the literature and expert experiences rather than long-term field-based biodiversity monitoring. Consequently, localized analyses have yet to integrate behavioral characteristics of particular species (e.g., specific butterflies or birds), which may create biases and constrain the capacity to comprehensively understand species-specific dynamics. Future research could combine ABS parameters with long-term monitoring data (like pollinator censuses) to enhance the ecological validity of the simulations.

The resolution of the data limited the ability to detect fine-scale green elements and microclimatic variations. While Landsat satellite data is mostly accurate, there are still inaccurate results due to spectral mixing, inadequate resolution, and incomplete atmospheric correction. Local-scale information such as the edges of pocket gardens or small patches of vegetation (e.g., street greenery and green walls) may not be fully captured. This could result in the differences between simulation and real ecological patterns.

Moreover, urban ecosystems are inherently dynamic, whereas this study only used April and November 2024 as temporal slices, without covering full annual cycles or enabling real-time monitoring. This constrains the ability to assess and predict ecological responses under extreme weather events. Future research should address this limitation by integrating digital twin technology, thereby extending temporal coverage and capturing the dynamic and uncertain nature of ecological processes more comprehensively.

Furthermore, the SOFIA system remains in a prototype stage and does not yet support fully automated analysis or generalization across regions. Sensitivity analysis or probabilistic modelling should be incorporated to better assess the uncertainty between parameters (GeoJSON) and outcomes (strategies and references), thereby enhancing the validity. Future development should also work on improving system interactivity and integration with real-time or participatory planning platforms like Mapbox.

In summary, green continuity is not only a spatial condition but also a conceptual lens for rethinking the intertwined relationships between urban ecology, social dynamics, and climate systems. Amid rising pressures on biodiversity and increasing heat inequities, this research offers a methodological contribution and design perspective that may inform future approaches to equitable, multispecies, and resilient urban spaces.