Research on Spatial Adaptation and Optimization Technology for “Region-City” Systems Based on ResGAT and LLM-MARL Models

Abstract

As China’s urbanization enters a stage of steady development while confronting diverse safety risks and challenges, coordinating “development” and “safety” has become a central proposition in territorial spatial governance. Conventional urban spatial simulation models rely heavily on trend extrapolation and lack multi-scale linkage mechanisms across regional networks and local land use. To address these challenges, this paper proposes a multi-scale spatial optimization framework linking regional and city tiers. The framework integrates three components: (1) a region–city linked adaptability assessment encompassing livability, safety, and efficiency–equity dimensions; (2) a hierarchical constraint transmission mechanism structured across “zoning–structure–pattern” levels; and (3) a city-level spatial adaptability optimization workflow driven by AI agents. Demonstrative simulations of China’s urban system toward 2035 indicate that incorporating topological scaling-law constraints increases global network efficiency from 0.42 to 0.50 (+19.0%) while reducing cascade vulnerability from 0.28 to 0.15 (−46.4%). Concurrently, city-level land-use structural adjustments project a 1.23% increase in urban blue-green open space nationwide by 2035, with increments approaching zero in mature eastern coastal agglomerations. While the top-down constraint transmission is fully operationalized, the bottom-up parameter aggregation is formulated as a methodological blueprint. This framework provides an exploratory decision-support paradigm for balancing high-quality development and systemic safety in regional spatial planning.

Graphical abstract

Keywords

Spatial adaptation optimization / Multi-scale coordination / Adaptation assessment / Constraint propagation / Urban network simulation / Urban renewal / Multi-agent reinforcement learning

Highlight

	● China’s urbanization formed dense city clusters, strong transport connectivity, and a nested urban-nature mosaic.
	● Urban spatial optimization should coordinate development efficiency and safety resilience across regional and city levels.
	● Construction land growth is projected to plateau around 2035–2040, while blue-green open increases from 27.44% (2020) to 28.67% (2035).
	● The proposed framework integrates adaptability assessment, hierarchical constraint transmission, and multi-scenario optimization using LLM + ResGAT + MARL.

Over the past four decades, China has undergone one of the largest and most rapid urbanization processes in global history [1,2]. Throughout this process, China’s urban spaces have progressively developed spatial structural patterns with distinctive characteristics [3]. While accomplishing the world’s largest rural-to-urban population migration, this transformation has simultaneously driven a profound restructuring of regional and urban spaces. Three principal features characterize this process: first, the intensive agglomeration of factors toward urban areas has given rise to a high-density urban network system anchored by city clusters and metropolitan areas. The agglomeration effects of megacities are particularly pronounced: the share of national urban population residing in megacities with populations exceeding 10 million surged from 10.5% in 2010 to 19.3%. Second, the space–time compression driven by transport infrastructure networks is highly significant. By 2024, China had constructed 48,000 kilometers of high-speed railway, reaching 96% of cities with populations above 500,000 and essentially achieving a spatial configuration in which provincial capital cities can access prefecture-level cities within their provinces within two hours. High-density transport infrastructure, serving as a catalytic medium for the evolution of urban network systems, has substantially compressed spatiotemporal distances and reduced barriers to the flow of regional factors. Third, the intensive interaction between urban expansion and natural substrates has produced a highly nested “urban–nature” mosaic. Owing to prolonged rapid spatial expansion and urban-rural ecological degradation, built-up land patches and natural ecological elements have become deeply interwoven and mutually interpenetrating.

However, this rapid expansion has long been subject to the rigid constraints of ecological carrying capacity and resource limits, with the contradiction between development demands and safety thresholds becoming increasingly manifest in the urbanization process. First, against the backdrop of intensifying global climate change, the growing frequency of extreme weather events poses severe challenges to urban safety. Approximately 13% of newly expanded urban spaces (relative to 1990 baselines) are currently exposed to moderate-to-high systemic risk [4]. The compound risks of typhoons and flooding in eastern coastal regions, the superimposition of heat waves and flooding in central regions, and complex geological hazards in western regions subject the resilience defenses of high-density human settlements to unprecedented tests. Concurrently, China’s highly space–time compressed urban network has given rise to increasingly prominent issues of regional economic resilience. Under extreme disaster impacts, flows of people, traffic, and information are susceptible to instantaneous disruption at critical network nodes and associated nodes [5], resulting in substantial losses. Furthermore, due to the high degree of interconnectedness in intercity economic networks, localized natural disasters or supply chain disruptions can generate significant risk spillover along spaces of flows, causing traditional assessments to underestimate systemic economic losses by an average of 52.5% to 57.5% [6]. In the eastern Yangtze River Delta and the Guangdong–Hong Kong–Macao Greater Bay Area, where urban spatial continuity is particularly pronounced, isolated urban heat islands have gradually coalesced into regional heat island archipelagos [7]. Consequently, the coordinated governance of “development” and “safety” has emerged as a central scientific proposition in contemporary territorial spatial governance.

Scale projection and spatial structure simulation constitute essential technical underpinnings for territorial spatial governance. Existing urban spatial simulation models employ Future Land Use Simulation Model (FLUS), Multi-Objective Programming coupled with Patch-generating Land Use Simulation Model (MOP-PLUS), and similar frameworks, utilizing historical land cover data for random forest training and trend extrapolation [8,9]. Alternatively, some studies have integrated system dynamics (SD) models (for macro-level quantitative projection) with PLUS models, exploring methodological approaches that fuse macro-level objectives with micro-level land patch prediction [10,11]. These methods have provided valuable support for land use prediction and urban spatial pattern assessment during the rapid urbanization phase. Nevertheless, they exhibit two principal limitations: first, they fail to integrate the dynamic mechanisms operating at regional and urban scales, with insufficient attention to the differential factors linking regions and cities across diverse geographic contexts; second, they demonstrate inadequate responsiveness to nonlinear risk perturbations arising from global climate change and macroeconomic fluctuations. Moreover, these methods insufficiently account for the feedback effects whereby micro-level changes influence macro-scale factors. For instance, large-scale development driven by urbanization induces regional warming and urban heat island effects, increasing the frequency of extreme weather events such as heat waves and intense precipitation, thereby producing multi-scale superimposed effects on climate at both global and regional levels [12]. At the micro level, adjustments to functional and land use structures can reduce negative feedback to macro-level systems. For example, land use optimization techniques targeting “urban–nature” nested configurations can effectively enhance climate adaptability, reducing waterlogging risk by 19% to 28%. It is therefore imperative to construct a multi-scale linked analytical framework spanning from national and regional to urban scales, and to strengthen cross-scale linked projection and simulation oriented toward the dual objectives of development and safety.

In recent years, AI models based on generative adversarial networks (GANs), reinforcement learning (RL), and deep reinforcement learning (DRL) have been widely applied in regional and urban planning [13]. These methods currently optimize predominantly for single-objective tasks, with insufficient attention to multi-objective tasks that reflect urban complexity. With the emergence of large language models (LLMs) and Generative AI, GeoAI is profoundly reshaping the spatial optimization technology paradigm [14,15]. Multi-agent reinforcement learning (MARL), for instance, has been introduced into planning decision-making processes, enabling spatial resource allocation to transition from Markovian evolution toward Nash equilibrium [16]. These artificial intelligence methods provide foundational support for integrating dynamic mechanisms across different spatial scales, enhancing the precision of urban spatial optimization techniques under cross-scale, multi-objective conditions, and enabling top-down and bottom-up transmission of key factors throughout the simulation process.

This paper addresses the lack of a cross-scale, coordinated optimization approach for regional and city spatial structures. The framework comprises three core components: a region–city linked adaptability assessment technique; a hierarchical constraint transmission technique operating across “zoning–structure–pattern” levels; and a city spatial adaptability optimization technique characterized by “multi-scenario anticipation, multi-scheme optimization, and multi-parameter feedback”. Leverage AI agents to drive the coordination of three sets of technologies, enabling the dynamic optimization of cities spaces and adaptation to changing scenarios (Fig.1).

The data employed in this study encompass 22 subtypes across seven categories—land use, socioeconomic, climate and ecology, road and transport, urban composite indices, flow data, and others—comprising nearly 40 individual datasets. Specifically, the land use, socioeconomic, climate and ecology, and road and transport subtypes are primarily utilized in Section 1 for the region–city spatial adaptability assessment technique, with data aggregated to kilometer-scale grids or 2136 “municipal district–county/county-level city” units. The urban composite indices, flow data, and adaptability assessment results are employed in Section 3.1 for the regional-level urban network optimization simulation, with data aggregated to prefecture-level and above city units. The land use, climate and ecology subtypes, together with the government work reports and five-year plan training corpus within the “others” category, are employed in Section 3.2 for the city-level land use scale and structural simulation, with data aggregated to “municipal district–county/county-level city” units.

1 Region–City Spatial Adaptability Assessment Technology

Guided by the “Precision Adaptability-Based Spatial Optimization Theory” [17,18], this study proposes an integrated multi-scale adaptability assessment technology spanning regional and city levels. At the regional scale, a livability–safety pattern adaptability assessment addresses the question of “where should be protected and where can growth be accommodated at the macro level”, establishing macro-level ecological carrying capacity and safety benchmarks to measure the rational adaptability between city spatial expansion and natural patterns. Concurrently, an efficiency–equity network adaptability assessment measures the synergistic performance between urban spatial agglomeration and industry, infrastructure, and services, thereby providing support for regional integration. At the city level, attention is directed toward spatial use efficiency and construction quality within urban development boundaries, including differentiated assessment of brownfield and underutilized land potential, as well as evaluation of climate adaptability indicators such as sponge city runoff coefficients and thermal environment regulation metrics [19].

1.1 Regional-level pattern adaptability assessment

The first component is the regional-level livability–safety pattern adaptability assessment. The comprehensive delineation of livability–safety-oriented construction mode zones and the assessment of livability levels and safety risk grades across different regions constitute essential prerequisites for city spatial development. First, livability conditions are determined using a Fuzzy Weighted Overlay (FWO) algorithm, which performs weighted calculations integrating parameters of temperature–humidity conditions, land resources, ecosystems, and water environments with population distribution, economic conditions, and transport network matrices. Second, particular emphasis is placed on safety risk projections under the influence of extreme weather and climate events—including heat waves, heavy rainfall, and typhoons—driven by global climate change [20]. Under emission scenarios ranging from SSP1-2.6 to SSP5-8.5 within the CMIP6 framework, kilometer-scale grid projections are constructed based on regional climate models coupled with urban canopy schemes, enabling spatiotemporal downscaling of precipitation and temperature data [21,22]. The downscaled projection results are subsequently transformed into spatiotemporally weighted graph networks, from which composite climate hazard risk coefficients are extracted for each node. Third, livability and safety factors are integrated through the reconstruction of high-dimensional distance metrics in the GDBSCAN algorithm. Given that meteorological disaster chains such as heat wave–drought and typhoon–storm surge–flooding exhibit cascading effects [23], and that the concurrent conditions of intensified extreme precipitation and high-temperature heat waves display strongly coupled risk characteristics, a time-varying covariance tensor is constructed under the Adaptive Bayesian Hierarchical Multi-source Fusion (ABHMF) framework¹. The generalized neighborhood predicate N_{\epsilon }\left(p_i\right) in GDBSCAN is jointly defined by a geospatial constraint \epsilon {}_{\mathrm{g}\mathrm{e}\mathrm{o}} and an attribute-space Mahalanobis distance threshold \epsilon {}_{\mathrm{a}\mathrm{t}\mathrm{t}\mathrm{r}}:

(1)N_{\epsilon }\left(p_i\right)=\left\{p_j\in D\hspace{0.33em}|\hspace{0.33em}{d}_{\text{geo}}\left(p_i,p_j\right)\le {\epsilon }_{\text{geo}}\text{and}{d}_{\sum _t}\left(x_i,x_j\right)\le {\epsilon }_{\mathrm{a}\mathrm{t}\mathrm{t}\mathrm{r}}\right\}

Where D is the set of all spatial analytical units, d_{\text{geo}}\left(p_i,p_j\right) is the Haversine spatial distance, and d_{\sum _t}\left(x_i,x_j\right)=\sqrt{{\left(x_i-x_j\right)}^T{\text{Σ}}_t^{-1}\left(x_i-x_j\right)} denotes the attribute-space Mahalanobis distance weighted by the time-varying covariance matrix \sum _t. Here, \sum _t is dynamically computed from multi-attribute hazard tensors (integrating precipitation downscaling, heatwave recurrence, and impacts of other disasters) under the ABHMF framework to capture multi-hazard compounding correlations.

In the GDBSCAN algorithm, the minimum number of points (MinPts) is set to 8, the spatial distance threshold {\epsilon }_{\text{geo}} is 120 km (with a moderate increase for areas west of the Hu Huanyong Line), and the attribute distance threshold {\epsilon }_{\text{attr}} is set to 1.65. The value of {\epsilon }_{\text{attr}}=1.65 was determined using the maximum curvature inflection point method applied to the k-distance graph (where k=\text{MinPts}=8).

Using “municipal district–county/county-level city” as the basic analytical unit, this study classifies the national terrestrial territory into 5 first-tier zones (livability conditions) and 18 second-tier zones [24], forming a nationwide base map of city construction mode zoning. Based on the SSP2-4.5 pathway, projections for 2050 indicate that the middle reaches of the Yellow River and southern Gansu will experience significant changes within their primary zones, with livability levels rising by 8% (95% Bayesian credible interval: 6%–10%); meanwhile, climate disaster risks in the Beijing–Tianjin–Hebei region, the Guangdong–Hong Kong–Macao Greater Bay Area, the Chengdu–Chongqing region, and the middle reaches of the Yangtze River are projected to increase by 18% (95% Bayesian credible interval: 16%–20%). These intervals were derived using an Adaptive Bayesian Hierarchical Multi-source Fusion (ABHMF) framework; by utilizing the uncertainty associated with downscaled CMIP6 model ensembles as a prior constraint, Markov Chain Monte Carlo (MCMC) sampling was employed to obtain the 95% Highest Posterior Density Interval (HPDI) for the time-varying posterior distributions of the composite indices for each region.

The second component is the regional-level efficiency–equity pattern adaptability assessment. At the regional scale, this assessment evaluates the adaptability between the spatial patterns of urban agglomeration and dispersion and four key dimensions: population distribution (P), industrial economy (E), transport networks (T), and facility services (F). Using the nationwide “municipal district–county/county-level city” units as the basic spatial unit enables continuous tracking of urban agglomeration and dispersion patterns. Methodologically, the entropy weight method is first applied to assign objective weights to multi-source heterogeneous indicators, yielding composite factor agglomeration scores for each spatial unit. Subsequently, the Jenks natural breaks classification method is employed to categorize basic spatial units into five tiers—“high, moderately high, moderate, moderately low, and low”—with further sub-classification achievable through key indicators. On this basis, spatial autocorrelation models (such as global and local Moran’s I indices) are introduced to quantitatively characterize the spillover effects of factor agglomeration within city clusters and metropolitan areas. This assessment framework enables the identification of structural bottlenecks including cross-boundary commuting resistance in metropolitan areas, homogeneous industrial competition, and ecological function conflicts.

Long-term time-series measurements (2000–2020) reveal a phased transition in China’s city spatial organization from “point-dispersed” to “network-coordinated” configurations. In 2000 (the initial dispersal phase), spatial organization was dominated by point agglomeration, with a pronounced urban–rural dual structure; highly coordinated areas were only sporadically distributed, and regional imbalances were prominent. By 2010 (the network development phase), with the extension of transport networks, city clusters and metropolitan areas began to take embryonic form. Regions such as Chengdu–Chongqing were among the first to exhibit moderately high coordination characteristics, and factor flows began to generate differentiated responses. By 2020 (the coordination acceleration phase), city clusters and metropolitan areas had become the dominant form of spatial organization, with a surge in the number of high/moderately high coordination units and substantial improvements in urban–rural integration. This evolutionary trajectory corroborates the necessity of multi-scale networked assessment and provides a baseline reference for land use structure optimization at the city level.

1.2 City-level construction pattern adaptability assessment

A construction pattern adaptability assessment is conducted for the built environment of each city, evaluating the constraints imposed on city development and construction by influencing factors including ecological environment, socioeconomic conditions, and humanistic environment. Considering the subsequent requirements for region–city cross-scale spatial optimization simulation, the aforementioned influencing factors can be synthesized into a “density–scale” dual-oriented indicator system to objectively reflect the characteristics and quality of urban development and construction modes.

Density-control indicators primarily reflect the intensive utilization efficiency of city space, encompassing metrics such as per capita construction land, building density, and land productivity. For instance, through nonlinear regression analysis of long-term historical data from over 300 prefecture-level and above cities nationwide, reasonable intervals for the relationship between urban population and construction land have been defined across different climate zones—including severe cold, cold, and hot-summer/warm-winter regions—with the elasticity regulation indicator set at 65–115 m² per capita.

Scale-control indicators primarily reflect the scale effects and performance of city space. In terms of measuring safety performance, indicators such as upper limits on floor area ratio, sponge city runoff coefficients, the proportion of blue-green open space within built-up areas, and green coverage ratios of built-up areas can be established. These indicators exhibit sensitivity to external macro-level ecological and environmental factors; for example, the proportion of blue-green open space within urban built-up areas is closely correlated with waterlogging mitigation, while the green coverage ratio of built-up areas is closely associated with the alleviation of heat island effects. In terms of measuring development performance, indicators such as land use mix ratios, the proportion of public service facility land, and the proportion of industrial and warehousing land can be established.

2 Cascading Linkage Technology under “Zoning–Structure–Pattern” Constraints

The aforementioned “region–city” multi-scale spatial adaptability assessment constitutes the scientific prerequisite for future urban spatial optimization simulation, providing a systematic evaluation from a holistic perspective for the site selection, structural optimization, and functional adjustment of city spaces. This study proposes a multi-constraint cascading linkage technology based on “zoning–structure–pattern”, which leverages AI agents, Language-based Multi-Agent Reinforcement Learning (LangMARL), and Residual Graph Attention Networks (ResGAT) to achieve bidirectional communication and dynamic feedback between macro and micro levels for spatial parameters of complex elements and flows, as well as policy intentions.

2.1 Top-down constraint transmission mechanism

The regional-scale assessment establishes the livability and safety “zoning” base map for city spatial expansion, while the clarification of inter-urban linkage relationships defines structural and mobility boundaries (the connection boundaries of urban networks), providing both development conditions and constraints for network optimization of urban systems under multiple scenarios. The city level assessment provides constraints on micro-level development and construction performance (density–scale), establishing the prerequisite conditions for construction modes across different regions and different types of cities. This study proposes a cross-scale constraint transmission mechanism of “zoning defines boundaries, structure defines the skeleton, and pattern defines parameters” elaborated as follows:

“Zoning” defines boundaries (macro-masking for spatial boundaries): based on the livability and safety zoning base map delineated through dynamic GDBSCAN, the zoning results can be transformed into a “Global Boolean mask” or a spatial probability decay field covering the entire domain. For instance, when global climate change triggers a rapid escalation of systemic disaster risk in a given region, the mask matrix for that region will be automatically activated. In micro-level multi-agent models, land expansion factors within such regions will be subject to high suppression, while seeking greater resilience-oriented spatial expansion drivers. Given that the top-down “Global Boolean mask” imposes overly rigid constraints on the continuity of the action space in microscopic LangMARL games, a decay function for the mask penalty can be incorporated into the reinforcement learning reward function.

“Structure” defines the skeleton (network topology as weighting multipliers): through the classification results of the efficiency–equity pattern adaptability assessment, a priori basis can be provided for the connection strength and efficiency of hierarchically stratified urban networks. For cities in which the number of basic spatial units assessed as moderate or above at the prefecture-level unit exceeds 60%, the network connection strength between pairwise adjacent cities should be reinforced, or additional connection corridors should be established². Such prior parameters enable the addition of 7.2%–13.6% new connections in the subsequent network topology simulation of inter-urban structures under different scenarios. Simultaneously, these parameters serve as nonlinear amplification multipliers (weighting multipliers) of local utility functions for corresponding spatial units in subsequent cities land use prediction.

“Pattern” defines parameters (performance thresholds as action space constraints): the construction mode adaptability assessment at the city micro-scale can be transformed into parameterized constraint conditions for functional zone or land use competition at the micro level. For instance, the upper limits on floor area ratio and lower limits on building density determined for specific hazard zones, as well as sponge city runoff coefficients determined for storm-flood risks, can serve as validation parameters for the action space in subsequent agent-based simulation models. When constraints are exceeded during the simulation process, the corresponding actions will be pruned during the trial-and-error exploration phase of Monte Carlo Tree Search (MCTS), thereby ensuring that micro-level development conforms to the requirements of land intensification and climate adaptability.

2.2 Bottom-up parameter aggregation and feedback mechanism

To ensure methodological transparency, while this study establishes the theoretical blueprint of bidirectional multi-scale coupling, the empirical simulation toward 2035 in Section 3 primarily operationalizes the top-down constraint transmission pathway (Zoning → Structure → Pattern). The bottom-up ResGAT feedback loop—updating regional network impedance from aggregated micro-level land changes in the underlying surface—is fully formulated algorithmically in Section 2.2 and serves as an extensible modular interface for subsequent iterative research.

Existing spatial optimization models often neglect the reverse influence of adaptive behaviors at the micro-agent level on the macro-system. The cascading perturbation effects of localized spatial adjustments on regional systems are increasingly significant. When deep-level stock function redistribution and land use structural adjustments are undertaken at the micro level—such as the large-scale installation of stormwater sponge corridors in response to extreme precipitation risks, or the provision of blue-green open space to mitigate urban heat island effects—although these constitute minor physical spatial modifications, they nonetheless alter the surface physical parameters of the corresponding urban spatial nodes to a certain extent (e.g., surface roughness, impervious surface impedance coefficients, and surface albedo).

Within the transmission system, these underlying micro-parameter changes can be inversely propagated through the node aggregation function of the Residual Graph Attention Network (ResGAT). Specifically, upon completion of structural optimization by micro-level agents, their updated the adjustment parameters for the resulting industrial, transportation, and blue-green open space land uses can be incorporated into the node feature vectors, which are subsequently fed into the base-layer parameters of the ResGAT model. The ResGAT recalculates the material–energy exchange efficiency and disaster resistance capacity between the subject node and surrounding cities via attention weight redistribution. This bottom-up feedback loop dynamically corrects the Joint Reward Function in the subsequent round of global optimization simulation, thereby reinforcing positive adaptive behaviors at the micro-spatial level based on disaster prevention and mitigation or ecological restoration. This paper proposes the aforementioned bottom-up parameter aggregation feedback mechanism from a methodological perspective.

2.3 Intelligent evolutionary closed loop of AI agents

When confronted with complex policy texts and massive heterogeneous datasets, spatial optimization simulations based on conventional random forests or heuristic rules are prone to computational bottlenecks. This study introduces and restructures the open-source autonomous evolutionary agent framework EvoScientist, constructing a decision-making closed loop termed “regional-city simulation agent” that spans from prior knowledge cognition to dynamic assessment feedback. This endows the spatial simulation with intelligent capabilities of self-correction, long-term evolution, and continuous adaptation to various parameter changes and transmissions. The agent framework transcends static AI workflows by deploying a Multi-Agent Team that collaboratively executes scientific discovery and simulation tasks. The agent team comprises the following three core roles:

Prior rule generator (researcher agent): leveraging the extensive knowledge injection capacity of Large Language Models (LLMs), this agent is responsible for ingesting and parsing complex natural language policy texts such as the “15th Five-Year Plan” spatial policies and dual-carbon strategies. Within the evolutionary closed loop, it functions as an “Action-Space Pruner”, precisely translating ambiguous macro-policy directives into high-confidence Prior Rules and probability flows, which constitute a critical foundation for multi-scenario simulation schemes.

Sandbox engineering executor (engineer agent): upon receiving prior rules or probability flows, this agent conducts trial-and-error simulation within an open-source sandbox simulation platform (integrating underlying graph computation, Markov decision processes, Nash equilibrium game-theoretic decision-making, etc.) using the UCT algorithm under Monte Carlo Tree Search (MCTS), while dynamically tuning relevant parameters.

Meta-Algorithm Optimizer and Memory Hub (Evolution Manager Agent, EMA): the EMA functional module utilizes a persistent memory module to monitor the simulation process in real time. First, it distills spatial configuration patterns that successfully achieve Pareto dominance and writes them into Persistent Ideation Memory (Ideation Memory, M_{\text{I}}). Second, it records parameters that trigger local conflicts or efficiency collapse into Persistent Experimentation Memory (Experimentation Memory, M_{\text{E}}). By continuously extracting reusable spatial governance patterns from historical iterations (for example, floor area ratio and green space allocation parameters validated as effective in coastal typhoon-prone areas), the EMA optimizer is capable of continuously optimizing the hyperparameter configuration of ResGAT or reconstructing reward function code. For instance, in cities land use scale and structural prediction, the EMA optimizer monitors the Global Reward trajectories of all “municipal district–county/county-level city” agents. When the global average reward growth rate remains below the threshold for consecutive steps, a gradient update of the dimensional weights and the temperature parameter τ is automatically triggered to achieve autonomous evolution.

In the subsequent regional-level optimization simulations, the multi-agent team establishes three scientific analysis pipelines for agent collaboration: first, the data-driven and simulation chain (loading raw data → Urban Spatial Deduction DB [simulation knowledge base] → Urban Network Builder [flow data fusion + three-layer network] →Code Sand box [MCTS Code Execution]); second, the AI reasoning chain (DeepSeek Client → Urban Master Agent [plan evaluation] → Urban Network Optimizer [network integration and optimization] → final simulation strategy recommendations); and third, the evolutionary optimization chain (EvoScientist-Optimizer [autonomous evolutionary iteration] → urban network de-densification → Urban Network Optimizer → final prediction results).

3 City Spatial Adaptation Optimization Technology Based on “Multi-Scenario Anticipation–Multi-Scheme Optimization–Multi-Parameter Feedback”

3.1 Regional-level urban network optimization simulation

In response to the strategic demands for optimizing the layout of the national urban center system and network system toward 2035 and beyond, this study proposes a hybrid simulation framework that integrates ResGAT with Urban Scaling Laws from spatial economics (Fig.2). This framework addresses the limitations of traditional gravity models in handling the nonlinear evolution of high-dimensional complex networks. In terms of specific methodology, a dual-layer graph attention network architecture of ResGAT is adopted, integrating a metadata fusion head for the urban feature matrix, a scaling law constraint regularizer, and a Speaker-Listener Label Propagation Algorithm (SLPA) overlapping network bias, thereby enabling a transfer learning strategy from present-state training to future-oriented simulation. The urban feature matrix encompasses the three-tier urban network together with centrality indicator values for urban economy, innovation, and transportation, as well as parameters related to livability-safety construction mode zoning. The principal pre-training parameters include: 4-head attention concatenation, 3 residual blocks, input dimension of 4, output dimension of 1, hidden dimension of 32, and a dropout rate of 0.1, an Adam optimizer with an initial learning rate of 0.005, 200 training iterations, Exponential Moving Average (EMA) decay (decay = 0.999), and a scaling law emergence constraint coefficient of 0.1.

First, a multi-tier urban topological network is constructed. Considering that the flows of socioeconomic elements, short- and long-distance population migration, and different transportation modes in China all exhibit hierarchically stratified spatial differentiation characteristics [25,26], three hierarchical levels of network topology are established as prior settings for the urban network to authentically reflect the actual spatial relationships among cities. The first-tier network constitutes the network among core cities, defined as those ranking in the top 20%, around which a fully connected matrix is constructed [n × (n − 1)/2 edges], forming the backbone of the national urban network. The second-tier network encompasses a defined hinterland area surrounding the aforementioned core cities, forming a core–hinterland network. The third-tier network comprises the interconnection network among cities in the remaining areas, forming a topological network among regional nodes. This hierarchically stratified urban topological network analysis differs fundamentally from traditional gravity-based network analysis models for inter-urban relationships. The multi-tier urban topological network serves as the foundational premise for urban spatial simulation; under multi-scenario simulation, variable factors such as the number of high-order central city nodes, and the distance scales between cities in the second- and third-tier networks, can be considered.

Second, the Comprehensive Urban Centrality Index for 2035 is predicted to provide node-level anticipation for subsequent urban network simulation. The national urban network is modeled as a heterogeneous topological graph, in which urban node feature vectors integrate measurement indicators including innovation indices, transportation network centrality, and comprehensive network centrality. The core algorithm employs the ResGAT, which quantifies the gravitational pull exerted by high-energy-level cities on surrounding cities, and adopts graph convolutional architecture to predict the latent potential of cities within the network topology. The present-state Comprehensive Centrality Index assigns differentiated measurement indicator weights according to cities of different scales, reflecting the heterogeneous driving mechanisms across cities of varying sizes. In predicting the Comprehensive Centrality Index, dynamic factors are considered, including the city’s network centrality position, economic growth potential, present-state path dependence, and neighborhood diffusion coefficients. Based on predictions of the comprehensive centrality index using 2019 training data, the ResGAT model—employing a three-layer network structure—achieved MAE, RMSE, R², and Spearman ρ values of 0.024, 0.035, 0.902, and 0.949, (Fig. 3)³; in contrast, the corresponding metrics for the gravity model were 0.274, 0.288, −5.67, and 0.589, demonstrating that ResGAT possesses superior predictive capability for centrality compared to the traditional gravity model. This approach also outperforms the results of Ridge training (without a graph structure).

To avoid the mean convergence problem commonly associated with deep learning, the allometric scaling law is introduced for scale correction, and Log-Logistic Mapping is applied for smoothing. This ensures that the Comprehensive Urban Centrality Index exhibits rapidly diminishing marginal gains once it surpasses a specific threshold during the prediction process. Finally, adjustment factors accounting for the impacts of future climate change on “livability-safety” zoning are incorporated to derive the corrected final Comprehensive Centrality Index. The regularization formula constrained by urban scaling laws is as follows:

(2){\mathcal{L}}_{\text{total}}={\mathcal{L}}_{\text{ResGAT}}\left(Y,\hat{Y}\right)+{\lambda }_{\text{scaling}}{\sum }_{i=1}^N{\left[{\log \hat{Y}}_i-\left({\log Y}_0+\beta {\log N}_i\right)\right]}^2

where {\mathcal{L}}_{\text{ResGAT}}\left(Y,\hat{Y}\right) denotes the base prediction fitting loss (Mean Squared Error) between the observed composite centrality vector Y and the predicted vector \hat{Y}; N is the total number of urban nodes (N=337 prefecture-level and above cities in China); {\hat{Y}}_i and N_i represent the predicted composite centrality index and the permanent population scale of city i, respectively; Y_0 is the baseline scaling intercept; \beta is the empirical allometric scaling exponent (calibrated as \beta \approx 0.86 in this study via nonlinear power-law regression across historical censuses); and {\lambda }_{\text{scaling}} is the regularization weight (Lagrange multiplier), set to 0.1 based on hyperparameter tuning to avoid mean convergence while preserving tier differentials.

To evaluate network performance before and after optimization, four topological metrics are quantitatively defined:

(1) Global Efficiency (E):

(3)E=\frac{1}{N\left(N-1\right)}{\sum }_{i\ne j}\frac{1}{d_{ij}}

where d_{ij} is the shortest path length between city i and city j weighted by flow impedance.

(2) Cascading Vulnerability (V):

Defined as the relative collapse of giant connected components under targeted removal of top-5% hub nodes:

(4)V=1-\frac{S{{}^{\prime }}_{LCC}}{S_{LCC}}

where S_{LCC} and S{{}^{\prime }}_{LCC} represent the sizes of the largest connected subgraph before and after cascade propagation under the Motter–Lai capacity-overload model.

(3) Small-world parameter (\sigma):

(5)\sigma =\frac{C/C_{\text{rand}}}{L/L_{\text{rand}}}

where C and L denote the clustering coefficient and characteristic path length of the empirical network, compared against equivalent Erdős–Rényi random graphs (C_{\text{rand}},L_{\text{rand}}).

The prediction methodology for the Comprehensive Urban Centrality Index, through multiple regularization and scaling re-projection, effectively preserves energy-level differentials while enhancing the anti-perturbation capacity of the network topology, achieving Pareto dominance under global anti-perturbation resilience at the system level. While Table 2 initializes the T1 backbone across a candidate set of cities ranking in the top 20% (approx. 67 cities), the message passing in ResGAT coupled with scaling-law regularizers automatically drives hierarchical stratification toward 2035. Specifically, Jenks natural breaks classification of the predicted index reveals that the upper tier further converges into two distinct sub-tiers: a primary national core sequence of 16 megacities (including Shanghai, Beijing, Shenzhen, Guangzhou, Chengdu, and Chongqing) with a mean composite index of 0.74, and a secondary sequence of 19 sub-national hubs with a mean index of 0.39, establishing an evident energy-level differential of 0.35. Cities such as Suzhou and Hefei exhibit the largest increases, reaching 0.20, with particularly notable improvements in innovation energy levels and transportation centrality (Fig.4).

Third, multi-scenario urban spatial network simulations are conducted. Three scenario modes are established: economic-driven (economic_driven), flow-oriented (flow_oriented), and balanced-network (balanced_network), all sharing an identical network topology structure (three-tier hierarchical network). Here, optimal weight allocation across different scenarios is determined based on Large Language Models (LLMs). For instance, under the economic-driven orientation, the weight parameter for inter-urban industrial mutual investment flows is set at 60%; under the flow-oriented mode, the weight parameter for inter-urban population flows is set at 60%; and under the balanced orientation, the weights for industrial mutual investment, patent transfers, population flows, and corridor weights (logistics) are relatively balanced. Based on the three-tier network parameters, this study deploys a dynamic tuning agent that further feeds back optimal adaptation parameters through the execution results of the three scenarios, to be utilized in subsequent comprehensive scheme simulations within the Sandbox Engineering Executor. Considering the strategic imperative of balanced development at the national level in the future, this study introduces a mechanism for industrial and innovation factor spillover and diversion organized under the guidance of urban transportation corridors⁴, whereby the intensity of connectivity between central cities is evaluated to determine whether multi-channel connections should be added, thereby enhancing the redundancy resilience of the urban network. On this basis, the UCT optimization algorithm under Monte Carlo Tree Search (MCTS) is employed to generate future urban network configurations, thereby achieving Nash equilibrium between new network corridor planning and the consolidation of existing strategic corridors.

Fourth, multi-round iterative simulation of the comprehensive scheme is executed through the Sandbox Engineering Executor of EvoScientist, the optimization was capped at 20 rounds and terminated at round 9 when the improvement in fitness fell below {10}^{-3} over three consecutive rounds. Simulation results indicate that to raise the overall efficiency—which measures the aggregate accessibility of cross-regional flows of population, capital, and innovation elements between cities—from the current 0.42 to 0.50, while the vulnerability index regarding the network’s resistance to cascading collapse following localized damage decreased from the current 0.28 to 0.15. The composite resilience score—reflecting the network’s ability to withstand shocks, adapt to adjustments, and restore connectivity—improved from 0.24 to over 0.37, and the small-world parameter σ of the urban network is 3.97, compared to the present-state baseline of 2.15), verifying that the optimized network preserves high local clustering alongside short path lengths. Even under scenarios where extreme weather events or major public health emergencies render a small number of core cities and critical corridors inoperative, the degradation of global network connectivity efficiency does not exceed 23% (compared to the current level of 40%–50%).

3.2 Simulation of land-use scale and structure at the city level

China’s urban development has entered a phase of stock renewal, wherein the redistribution of spatial resources exhibits characteristics of multi-stakeholder gaming, subject to strong policy-semantic constraints and value orientations [27]. Under the theoretical frameworks of spatial economics and complex urban system evolution, land-use scale expansion and structural adjustment are no longer reducible to simple Markovian evolution, but rather constitute a Nash Equilibrium optimization process encompassing the demands of multiple stakeholders [28].

Building upon the conventional Random Forest–Cellular Automata (RF + CA) land-use prediction model [29], this study employs LLMs to translate “development and safety” policy texts into micro-behavioral constraints and reward functions, subsequently integrating an LLM-coupled multi-agent reinforcement learning model (LLM-MARL) [30] to achieve Pareto optimality consistent with anticipated objectives. The methodology proceeds as follows: LLMs are trained on textual corpora to parse various macro-policy texts (e.g., “dual carbon” targets, non-grain conversion controls) and output prior guidance strategies; on this basis, different city functional zones or land-use types are converted into dynamic cellular weighted graph networks, and a joint reward function balancing economic development and safety baselines is constructed to conduct multi-agent game-theoretic simulations. Given that complex policy texts are difficult to mathematize directly, LLMs are utilized to parse policy semantics and output value tensor matrices, which are subsequently smoothly projected onto probabilistic strategies of spatial behavior via Boltzmann distributions.

To mitigate the inherent stochastic hallucinations and subjective biases of Large Language Models (LLMs), the LLM is strictly constrained to function as an auxiliary tool for semantic parsing. Structured prompts are designed to construct the pre-training dataset—incorporating role definitions, spatial contexts, policy excerpts, domain schema definitions, and requirements for controlled JSON output—while mandating the use of few-shot contextual examples and strict data type constraints. Furthermore, the model’s generation parameters are controlled by setting the inference temperature to τ = 0.0 (deterministic greedy decoding) and Top-p to 0.95; additionally, each policy document undergoes at least five inference iterations to assess semantic reproducibility. Building on this, a dual-loop filtering mechanism—comprising automated consistency verification and a Human-in-the-Loop (HITL) audit pipeline—is employed to solidify parameter mappings into the Urban Spatial DeductionDB, thereby ensuring that subsequent cellular automata and reinforcement learning simulations are grounded in verified and reproducible empirical data.

This study employs DeepSeek-v4-flash to pretrain on datasets comprising government work reports at the provincial, prefectural, and county levels from 2010 to 2023, as well as municipal five-year plans for national economic and social development since the Twelfth Five-Year Plan period. The pretraining dataset construction incorporates semantic classification across dimensions including climate change, ecological priority, innovation-driven development, industrial upgrading, and stock renewal. These probability matrices are incorporated as novel “policy factor” features and merged with the historical land-use and environmental feature vectors of the RF model. The training results are then converted into a reward function via an LLM-MARL model and passed to interactive agents representing “municipal district–county/county-level city”. Subsequently, each agent generates differentiated probabilities when executing various land-use conversion and adjustment strategies; finally, the integrated RF-CA algorithm is employed to conduct multi-scenario projections of urban land use across the country for the years 2035 and 2050. We introduce the Semantic Alignment and Task Decomposition (SA-TD) method from multi-agent reinforcement learning (MARL); this approach leverages pre-trained language models to suggest potential objectives, thereby facilitating appropriate task decomposition and sub-goal allocation [31]. To address computational efficiency in multi-agent game simulations, we construct 90 agent types based on a zoning scheme comprising five primary categories and 18 secondary categories of urban development patterns, while simultaneously simplifying the action strategies for the corresponding agents.

Specifically, the policy semantic parsing pipeline translates qualitative municipal guidelines into structured vector weights:

(6)R_{\text{policy}}={\sum }_k{w}_k\cdot S_k

Where S_k\in \left[-1,1\right] represents the semantic alignment score extracted by DeepSeek-v4-flash across five strategic dimensions (e.g., dual-carbon stringency, greening targets, and stock redevelopment), and w_k is the prompt-derived dimensional attention weight. The action space of the 90 functional agents across 2,136 county-level units is defined as a multi-discrete decision vector a=\left[\text{Δ}{L}_{\text{ind}},\text{Δ}{L}_{\text{res}}\text{, Δ}{L}_{\mathrm{c}\mathrm{o}\mathrm{m}}\text{, Δ}{L}_{\text{green}}\text{, ΔFAR}\right], representing the marginal reallocation of industrial, residential, commercial-office and blue-green land shares, alongside floor area ratio (FAR) intensity bounds under cellular spatial constraints. Prior to forward-looking projection, a historical back-casting test was conducted on the land-use transitions across the 2,136 county-level units from 2010 to 2020 to validate the predictive robustness of the LLM-MARL framework. The simulation achieved an overall classification accuracy of 89.49% ± 4.28% and an average Kappa coefficient of 0.81 ± 0.08, confirming the empirical reliability of the multi-agent decision model.

Driven by the multi-agent game-theoretic interactions under macro-policy semantics and safety baselines, the simulation results indicate that national urban construction land expansion will enter a plateau of slow growth during the 2035–2040 period. Regarding the regional distribution of land-use increments, new expansion in eastern coastal regions (the Yangtze River Delta, the Guangdong–Hong Kong–Macao Greater Bay Area, and Beijing–Tianjin–Hebei) will approach net zero around 2030, indicating that their primary growth drivers have decoupled from land-increment dependence and transitioned into functional adjustment and optimization of existing land stock. In central and western regions, particularly at comprehensive transportation hubs and along major development corridors, land-use scale retains moderate incremental capacity, supported by national regional coordination policies and localized employment of populations. Furthermore, future land-use increments in China’s major metropolitan areas are projected to exceed the national average by 15%–25%.

Regarding land-use structural adjustment, the proportion of blue-green open space in national urban areas will increase from 27.44% in 2020 to 28.67% in 2035. This increment of 1.23 percentage points will yield approximately 6300 square kilometers of additional blue-green open space, which will play a positive role in addressing future heavy precipitation events and mitigating urban heat island effects. Regionally, in Northeast, Northwest, and Central China, the proportion of blue-green open space increases by 2.5–3.0 percentage points, driven by industrial site greening and ecological corridor restoration; whereas in high-density development areas such as East and South China, constrained by land availability, the increase remains below 0.8 percentage points.

4 Discussion and Prospects

4.1 Discussion of principal findings

Based on a “region-city” nested spatial configuration, this study constructs a comprehensive spatial optimization framework designed to systematically balance a multi-objective matrix comprising development efficiency, regional equity, livability constraints, and disaster resilience. First, livability, ecological constraints, and safety risk benchmarks are mapped into global spatial decay masks to inform the regulation of urban expansion and the adjustment of functional layouts. Second, at the regional-level, a Residual Graph Attention Network (ResGAT)—regularized by urban scaling laws—is employed to harmonize development efficiency with regional equity. By integrating multi-source socioeconomic flow matrices, the ResGAT simulates the hierarchical diffusion of innovation and capital across a three-tier urban system. To prevent excessive divergence between the core and the periphery, spillover mechanisms linked to transportation corridors channel factor flows toward secondary hubs; this simultaneously enhances national circulation efficiency and narrows regional development disparities. Third, at the city-level, an LLM-MARL approach prioritizes the balance between economic development and safety risk mitigation, thereby ensuring the resilience and risk-coping capacity of the urban system. Furthermore, the study proposes a bottom-up aggregation mechanism to feed these micro-structural adjustments back into the urban network, thereby enhancing the network’s structural redundancy and its buffering capacity against cascading failures.

Grounded in the spatiotemporal evolutionary mechanisms of nested regional–city complex configurations, this study constructs a spatial adaptation optimization technical framework integrating Large Language Models, spatiotemporal graph attention networks, and multi-agent reinforcement learning. Through the application of this technical framework to simulate China’s 2035 urban network and predict land-use scale and structure, the findings reveal that future urban space at the national scale does not represent a simple zero-sum game between polarized agglomeration and territorial equilibrium, but should rather achieve a dynamic balance between global factor circulation efficiency from macro to micro scales and local regional disaster resilience. The future should witness the construction of urban spatial networks and structures characterized by “moderate redundancy resilience, globally efficient development, and hierarchically coupled systems”. The urban network resilience assessment and spatial adaptation optimization algorithm based on the ResGAT proposed in this study constitutes precisely an adaptive optimization methodology addressing China’s complex regional economy and urban networks superimposed with the high-dimensional “Disaster Multiplier Effect” of multi-hazard compounding, capable of responding promptly to the dual imperatives of development and safety. To this end, national macro-regional development strategies should, in the future, ensure global efficiency through cultivating high-level core hubs while simultaneously enhancing the redundancy of regional branch networks to withstand localized risks, thereby advancing the policy implementation of “national central cities + national corridors + regional hubs + emerging nodes”. This study also demonstrates that, for China’s vast western regions, greater investment in long-distance connectivity infrastructure (railways and aviation) is particularly imperative, and strengthening multi-channel connections with central and eastern cities constitutes a critical measure for mitigating national urban vulnerability.

4.2 Future iterative prospects for spatial simulation algorithms

This study has constructed a livability- and safety-oriented city spatial adaptation optimization framework, providing preliminary simulation and prediction results for China’s 2035 urban spatial network structure and city land-use structure. Confronted with the complex nested mechanisms of regions and cities and the intricate dynamic mechanisms of spatial evolution, further in-depth research is required in the following areas.

The spatial optimization at regional and urban levels discussed here represents a preferred solution identified under specific scenarios and constraints; state transitions are modeled via a Markov process, whereas Nash equilibrium reflects the stability of a given strategy—meaning the relationship between these two models is not one of simple substitution. In the future, the Hierarchical Markov Decision Process (H-MDP) model may be introduced to reconstruct bidirectional mechanisms, thereby optimizing the spatial simulation model for regional–city coordination. Within top-down constraint governance, macro-level national livability–safety zoning changes and regional-level efficiency–equilibrium network configurations are to be transformed into global tensors for micro-spatial units, enhancing differentiated constraints on diverse micro-spaces; concurrently, within bottom-up feedback, micro-level land-use adjustments at the city level automatically feed back to macro-level real-time adjustments of transportation networks and ecological corridors. At the regional scale, AI-Agent frameworks may be further deployed to enhance the accuracy of long-term predictions, particularly by structurally converting the regional and urban spatial governance methodologies and experiences accumulated during China’s forty-year urbanization process into persistent experiential rules within the EvoScientist framework, thereby facilitating a priori evaluation of future policies. At the city level, the LLM-MARL model should more precisely parse the “consensus information” [32] extracted by large language models in terms of Global Reward, further leveraging the natural language causal reasoning and inductive capabilities of LLMs to investigate spatial game convergence challenges under sparse reward environments, thereby providing interpretability for the Nash Equilibrium optimization process.

Further deepening of micro-scale spatial adaptation optimization technologies is warranted. In the future, attention should be directed toward confronting and overcoming the spatiotemporal simulation bottlenecks under coupled high-resolution urban climate disaster chains, in response to the highly non-stationary characteristics of global climate change. By integrating fuzzy Choquet integrals and nonlinear amplification mechanisms, further research may be conducted on the cascade propagation characteristics and city spatial impacts of typical multi-factor compound disaster chains—such as “typhoon–storm surge–surface waterlogging” in coastal cities and “extreme rainstorm–major river flooding–high-density city waterlogging” in inland cities—across different spatial scales. On this basis, high-dynamic spatiotemporal graph neural networks may be introduced to refine intra-city livability–safety zoning at granular scales and to propose high-precision resilience governance and layout schemes. Other critical safety dimensions—including biodiversity security, food security, and public health security—are not yet systematically integrated into the multi-scale assessment and simulation framework. These represent important directions for future extension of the technical framework.

5 Conclusions

Coordinating high-quality economic development with high-level systemic safety constitutes a pivotal challenge for China’s territorial spatial governance in the post-rapid urbanization era. In response to the limitations of conventional extrapolation models in cross-scale dynamics and nonlinear disaster disturbances, this study develops a multi-scale spatial optimization framework coupling regional networks with city-level land allocations. By integrating Spatiotemporal Graph Attention Networks (ResGAT) governed by urban scaling laws with Language-based Multi-Agent Reinforcement Learning (LLM-MARL), the proposed technical architecture establishes a coherent transmission pathway spanning “zoning–structure–pattern”. Demonstrative multi-scenario simulations toward 2035 validate the feasibility and efficacy of the framework:

(1) Regional Network Reshaping: by incorporating allometric scaling regularizers into topological optimization, the global network efficiency across Chinese cities improves by 19.0% (from 0.42 to 0.50), while cascade vulnerability under targeted disruption of critical hub nodes decreases by 46.4% (from 0.28 to 0.15). Concurrently, the small-world coefficient reaches 3.97, verifying enhanced system redundancy and localized anti-perturbation resilience.

(2) City-Level Structural Evolution: Macro-policy semantics parsed by LLMs reveal that urban construction land expansion nationwide will plateau between 2035 and 2040, with eastern coastal megacities fully decoupling from incremental land expansion by around 2030. Simultaneously, the proportion of functional blue-green resilience spaces nationwide is projected to increase by 1.23% (yielding approximately 6300 {\text{km}}^2), providing vital ecological buffering against heat islands and convective storm-runoff risks.

Overall, this research demonstrates that future territorial optimization must transcend polarized agglomeration to achieve dynamic equilibrium between macro-factor circulation efficiency and regional safety resilience. While the top-down constraint mechanism has been operationalized, achieving a fully coupled bidirectional loop through micro-parameter feedback remains a crucial frontier. The proposed technical framework offers an exploratory, data- and model-driven decision-support paradigm for resilient national spatial planning and sustainable regional coordination.

References

[1] Seto K C , Golden J S , Alberti M . et al. Sustainability in an urbanizing planet. Proceedings of the National Academy of Sciences of the United States of America, 2017, 114(34): 8935–8938

[2] Lu D D . Urbanization process and spatial sprawl in China. Urban Planning Forum, 2007, (4): 47–52

[3] Wang K , Chen M , Zhang D N . Optimisation of national urban spatial pattern: new thinking based on economic potential and security risk dimensions. Urban Planning International, 2023, 38(1): 1–9

[4] Wang K. 2035 Urban outlook for Chinese-style modernization. Available online (accessed 8 Jul 2024) (in Chinese)

[5] Kaiser S K , Rodrigues F , Azevedo C L . et al. Spatio-temporal graph neural network for urban spaces: interpolating citywide traffic volume. Expert Systems with Applications, 2026, 316: 131823

[6] Sheng H T , He C F , Hu W B . Intercity economic ties amplify natural disasters spillovers in China. Nature Cities, 2026, 3(3): 283–292

[7] Zhou W Q , Yu W J , Qian Y G . et al. Beyond city expansion: multi-scale environmental impacts of urban megaregion formation in China. National Science Review, 2022, 9(1): nwab107

[8] Zhang Y P , Lin J P , Huang Y M . et al. Delineation of urban growth boundary based on FLUS model under the perspective of land use evaluation in hilly mountainous areas. Journal of Mountain Science, 2024, 21(5): 1647–1662

[9] Liu Y , Zhen T , Niu T C . et al. Research on spatial suitability characteristics and multi-scenario simulation under urban land use evolution. Ecological Science, 2025, 44(5): 208–218

[10] Hu M M , Li Z T , Wang Y F . et al. Spatio-temporal changes in ecosystem service value in response to land-use/cover changes in the Pearl River Delta. Resources, Conservation and Recycling, 2019, 149: 106–114

[11] Bao S W , Cui W L , Yang F . Future land use prediction and optimization strategy of Zhejiang Greater Bay Area coupled with ecological security multi-scenario pattern. PLoS One, 2024, 19(4): e0291570

[12] Luo X Y , Chen M X . Research progress on the impact of urbanization on climate change. Advances in Earth Science, 2019, 34(9): 984–997

[13] Shen J , Zheng F H , Chen T Y . et al. Optimizing urban land-use through deep reinforcement learning: a case study in Hangzhou for reducing carbon emissions. Land, 2025, 14(12): 2368

[14] Wang C L , Tang H Y , Yang X Y . et al. Uncovering inequalities in new knowledge learning by large language models across different languages. Proceedings of the National Academy of Sciences of the United States of America, 2025, 122(51): e2514626122

[15] Zheng Y , Xu F L , Lin Y M . et al. Urban planning in the era of large language models. Nature Computational Science, 2025, 5(9): 727–736

[16] Qian K J, Mao L J, Liang X, et al. AI agent as urban planner: steering stakeholder dynamics in urban planning via consensus-based multi-agent reinforcement learning. arXiv preprint arXiv: 2310.16772, 2023

[17] Wang K , Xu H , Zhou Y J . The essence of “National Spatial Planning Theory”: methods and applications of precise analysis and adaptive technology. Urban Planning Forum, 2022, (6): 8–17

[18] Wang K , Xu H , Zhou Y J . Theoretical approach and practice of “National Spatial Planning Theory”: how Chinese cities and towns precisely adapt to nature and culture. Frontiers of Urban and Rural Planning, 2024, 2(1): 4

[19] Guo H D , Liang D , Sun Z C . et al. Measuring and evaluating SDG indicators with Big Earth Data. Science Bulletin, 2022, 67(17): 1792–1801

[20] Wang K , Kuang W H , Fang W H . et al. Spatially explicit adaptation characteristics of urban development and construction across China over the past three decades. Science China Earth Sciences, 2025, 68(5): 1624–1640

[21] Pei L , Miao S G , Huang X Y . et al. Assessing the added value of convection-permitting modeling for urban climate research: a case study in eastern China. Advances in Climate Change Research, 2025, 16(1): 1–11

[22] Pei L , Miao S G , Chen D L . et al. Rising rainfall variability amplifies hydroclimate extremes in eastern China. Environmental Research Letters, 2026, 21(8): 084014

[23] Ma S , Li J H , Deng H Z . et al. Construction and uncertainty analysis of urban climate models under multi-source data fusion. International Journal of GEOMATE, 2025, 29(134): 169–183

[24] Wang K , Xu H , Jia P F . et al. A precision adaptation resilience-planning methodology that integrates regions and cities. Strategic Study of CAE, 2025, 27(4): 43–54

[25] Chen W , Xiu C L , Ke W Q . et al. Hierarchical structures of China’s city network from the perspective of multiple traffic flows. Geographical Research, 2015, 34(11): 2073–2083

[26] Gao P , Qi W , Liu S H . et al. Comparison of spatial structure and organization mode of city networks from the perspective of long-term and short-term population flows. Human Geography, 2023, 38(6): 82–91

[27] Ni H, Wang Y Z, Liu H. Planning, living and judging: a multi-agent LLM-based framework for cyclical urban planning. arXiv preprint arXiv: 2412.20505v1, 2024

[28] Song H Y, Wang H. A study on benefit distribution of multi-agent urban residential land supply based on game theory. In: Lu X H, Zhang Z, Lu W S, et al, eds. Proceedings of the 25th International Symposium on Advancement of Construction Management and Real Estate. Singapore: Springer, 2021, 1209–1221

[29] Kuang W H . 70 years of urban expansion across China: trajectory, pattern, and national policies. Science Bulletin, 2020, 65(23): 1970–1974

[30] Pelcner L, Marcolino L S, do Carmo Alves M A, et al. Multi-objective quantile-based reinforcement learning for modern urban planning. In: Proceedings of the 34th International Joint Conference on Artificial Intelligence. Montreal, Canada, 2025

[31] Li W H, Qiao D, Wang B X, et al. Multi-agent credit assignment with pretrained language models. In: Proceedings of the 28th International Conference on Artificial Intelligence and Statistics. Mai Khao, Thailand, 2025, 1945–1953

[32] Yao H Y, Da L C, Liu X O, et al. LangMARL: natural language multi-agent reinforcement learning. arXiv preprint arXiv: 2604.00722, 2026

Rights & permissions

The Author(s) 2026. This article is published by Higher Education Press.