[Objectives] Cities have long been metaphorically described as living organisms, yet prior research has rarely advanced beyond rhetorical analogy or macro-scale statistical descriptions, leaving a persistent gap in formal modeling and actionable governance support. This study aims to close this theoretical and technical divide by proposing a rigorous, hierarchical, and computable urban organism framework that can underpin full life-cycle governance, cross-system cascade analysis, and intelligent decision-making, particularly for megacity resilience, livability, and sustainability challenges. [Methods] Grounded in complex systems science—including dissipative structure theory, Complex Adaptive Systems (CAS), scaling laws, and multilayer network dynamics—we operationalize the urban organism metaphor by constructing a heterogeneous cascading network representation. We encode urban elements (e.g., energy, buildings, resources, economy, governance, society, and ecology) as typed nodes with spatial coordinates, functional intensities, and temporal trajectories, while defining intra- and inter-subsystem couplings via spatial proximity and temporal flows. Symbolic network dynamic equations are introduced to abstract three fundamental interaction logics (competition, synergy, dependency) while preserving physical conservation constraints for material and energy transfers. To translate the framework into governance operations, we further integrate digital twin platforms, LLM-generated decision agents for micro-level behavioral emulation, and Reinforcement Learning (RL) for policy optimization, forming a closed-loop “simulation + decision” paradigm. [Results] The proposed framework models cities as a coupled system of three sustainability dimensions and seven interacting subsystems linked through cascading exchange pathways. This formal networked structure enables symbolic representation and numerical solution of urban dynamics, supporting stable-state inference, cross-scale emergence deduction, and critical threshold identification. In hypothetical emergency scenarios such as pluvial flooding, the model captures cascade propagation from drainage overload to energy disruption and transport blockage, generating optimized pump operation and evacuation routing strategies through RL search. In chronic urban disease contexts such as long-term spatial resource imbalance, the framework dynamically simulates feedback loops among population redistribution, infrastructure investment, transit accessibility, and public service coverage, allowing RL to discover element configuration strategies that reduce systemic commuting costs and enhance service equity more effectively than static, centralized planning heuristics. [Conclusions] By formalizing the metaphor of cities as living organisms into a heterogeneous cascading network model and embedding it into a digital twin environment with RL-based policy optimization, this study delivers a unified theoretical language, a cross-system dynamic modeling paradigm, and a novel closed-loop governance path. The framework shifts urban research from symbolic analogy toward computational governance, offering a scientifically grounded and decision-actionable foundation for megacity resilience management, long-term spatial equilibrium optimization, and AI-assisted urban policy coordination.

[Objectives] Cities are widely regarded as complex mega-systems, and the concept of the “urban organism” has been explicitly introduced China national polices, which emphasizes the need to treat cities as holistic systems. However, both academic and practical domains still lack a unified understanding of urban systems; their functional structures, constituent elements, and coupling relationships remain unclear. This study argues that the urban system is not an a priori objective entity, but rather a socio-technical assemblage continuously constructed through ideas, discourse, institutional practices, and social interactions. The functions of municipal governments essentially represent a core act of construction, defining not only "what a city is" but also "how a city should be governed", thereby delineating the boundaries of "urban problems". Based on this idea, this study adopts a constructivist perspective and proposes a government-function-based cognitive pathway for understanding urban systems, with the aim of achieving a systematic and holistic comprehension of the city as an integrated whole. [Methods] Specifically, drawing on both governmental functions and academic literature as dual knowledge foundations, and leveraging large language models and graph retrieval-augmented AI techniques, this research develops an interdisciplinary knowledge graph of urban systems to identify the complex interrelations among elements. [Results] The study extracts 19 major categories of government functions and their associated elements from 131 government functional documents covering Guangzhou, Shenzhen, and Shanghai, and aligns them with the Sustainable Development Goals. The constructed urban system knowledge graph comprehensively and effectively captures the interconnections among elements, supporting an in-depth understanding of the functional structure and operational mechanisms of urban systems. [Conclusions] Theoretically, this study contributes a new cognitive framework and methodological pathway for urban system research, expanding systematic comprehension from the perspective of the urban organism and laying a solid foundation for future urban governance practice and sustainable development studies.

[Significance] As a complex organism characterized by self-organization, metabolism, and adaptive evolution, a city exhibits highly systematic and dynamic behaviors in both its operation and long-term development. With rapid urbanization and increasing system complexity, contemporary cities are confronted with persistent and intertwined challenges, such as environmental degradation, traffic congestion, resource imbalance, and public safety risks. Traditional urban governance paradigms, which often rely on static planning and reactive interventions, are increasingly inadequate for addressing these non-linear and multi-scale urban problems. This situation calls for a fundamental shift toward a full life-cycle governance paradigm that treats the city as an integrated and evolving organism. Digital twin technology, serving as a bridge between the physical city and its virtual counterpart, offers a promising technical pathway for realizing such a paradigm. By enabling continuous interaction between real-world observations and computational models, digital twins provide new possibilities for fine-grained perception, system-level understanding, and proactive governance. Against this background, this paper focuses on digital twin technologies from the perspective of the urban organism, aiming to establish a methodological foundation that supports perception, understanding, inference, and decision-making throughout the entire life cycle of future urban organisms. [Analysis] This paper systematically analyzes the core functionalities and development pathways of digital twins for urban organisms, with a particular focus on urban modeling, operational simulation, and intelligent decision-making. First, acknowledging the inherent incompleteness and uncertainty of urban data, we explore data completion strategies based on Bayesian inference and graph diffusion models. These approaches enable the reconstruction of missing or sparse observations and support the construction of digital twin models that span multiple interconnected subsystems of the urban organism. Second, to capture the coupled dynamics of physical infrastructure and social behaviors, we integrate dynamic system models with agent-based simulation methods, and propose a co-simulation mechanism that enables coordinated modeling of physical and social urban elements. This mechanism is designed to be continuously updated through real-time monitoring data, ensuring consistency between simulated states and real-world urban conditions. Furthermore, to support the early warning and governance of urban diseases, we introduce a spatiotemporal foundational modeling approach for urban indicator prediction and counterfactual inference. This approach allows different governance strategies to be simulated, evaluated, and iteratively optimized within the digital twin environment before real-world implementation. [Prospect] From the perspective of the urban organism, this study establishes a systematic technical framework and research approach for urban digital twins that goes beyond static physical mapping. By emphasizing data completion, multi-system co-simulation, and predictive inference, the proposed framework aims to enhance the capability of digital twins to support perception, understanding, inference, and decision-making across the full life cycle of urban governance. This work provides methodological insights for developing next-generation urban digital twins that are better aligned with the complexity, dynamics, and evolutionary nature of real cities.

[Objectives] Urban challenges such as imbalances in the spatial allocation of public service facilities and persistent traffic congestion—arising from the obstruction of urban factor flows and the fragmentation among subsystems—are increasingly difficult to address through traditional governance models that focus on isolated subsystems. This study aims to construct a governance model that transforms urban governance from passive response to proactive intervention. [Methods] The Urban Organism Theory, which highlights systemic characteristics such as metabolism, self-renewal, and adaptability, provides a crucial theoretical foundation for proactive governance. Based on this, this study defines the characteristic of psychological resilience as an intrinsic learning mechanism for dynamic strategy optimization based on the accumulation of historical governance experience, which corresponds to the decision-making agent module in the reinforcement learning framework. Further, this paper proposes a simulation-decision synergistic governance concept grounded in the urban organism paradigm, systematically elucidates its core features and inherent governance logic, and constructs a comprehensive proactive urban governance framework integrating simulation deduction and reinforcement learning technologies. The framework establishes an asynchronous synergistic mechanism of "slow simulation" verification and "fast decision" response by decoupling training and decision processes, where the urban simulator completes high-fidelity strategy verification in non-real-time conditions, and the decision-making system achieves rapid emergency response in actual scenarios. Moreover, it proposes differentiated governance pathways for chronic diseases characterizing sub-health and acute diseases corresponding to sudden risks, with the simulator undertaking evolutionary law modeling and strategy optimization for the former and risk deduction for the latter. [Results] Experiments on the SUMO simulation platform demonstrate that in diverse stress scenarios, the framework effectively optimizes resource allocation and emergency schemes, outperforming traditional fixed-timing and adaptive control strategies while enhancing network efficiency and systemic resilience. [Conclusions] The synergistic framework proposed in this study provides a feasible technical path for the precise configuration and dynamic restoration of urban multi-factors, effectively breaking the limitations of traditional fragmented governance and promoting the transformation of urban governance from passive response to proactive prevention and dynamic collaborative restoration. Future research can further explore multi-source data fusion technologies to integrate Internet of Things data, population mobility data, and public service resource data, improving the accuracy of urban state perception. Additionally, it is necessary to focus on the balance between computational power and simulation accuracy, optimize the asynchronous update mechanism of the reward model, and expand the application scenarios of the framework to public service allocation, ecological environment governance, and other fields to enhance its practicality and universality.

[Objectives] This paper clarifies the disciplinary meaning and methodological scope of combining urban simulation with Deep Reinforcement Learning (DRL) for intelligent city planning. Rather than presenting a general survey of all AI-based planning methods or proposing a single new algorithm, it examines an emerging computational paradigm in which an urban simulator, a planning agent, and reward-based feedback form a closed loop for generating and evaluating planning alternatives. [Methods] In this paradigm, the urban simulator works as an interactive digital twin that represents land parcels, roads, facilities, population distribution, travel demand, and planning constraints. These heterogeneous urban elements are encoded as graph structures, and planning interventions are formulated as Markov Decision Processes. A DRL agent observes the simulated urban state, selects planning actions, receives reward signals that quantify accessibility, efficiency, cost, equity, or sustainability, and updates its policy through repeated interaction with the simulator. Graph neural networks further allow the agent to capture spatial adjacency, functional flows, and multi-scale dependencies in irregular urban systems. [Results] Existing studies show that this framework has been applied to several representative planning problems, including the spatial layout of 15-minute communities, road planning for urban villages and informal settlements, large-scale public facility location selection, and metro network expansion. These cases indicate that the framework can search a large solution space, respect domain constraints through action masking or knowledge-informed operations, and provide measurable improvements over conventional baselines in accessibility, service coverage, computational efficiency, and planning diversity. More importantly, the framework changes the role of geographic information science and planning support systems: GIS-based analysis is no longer limited to evaluating predefined alternatives, but can become a computational laboratory for actively producing, testing, and comparing alternative urban futures. For scholars in human geography, urban studies, and planning, it also provides a way to translate concepts such as accessibility, spatial interaction, service fairness, resilience, and behavioral constraints into computable states, transitions, and rewards while preserving space for expert deliberation. [Perspective] The integration of urban simulation and DRL therefore represents more than a technical combination of models. It suggests a shift from experience-driven and static plan comparison toward a dynamic "simulation-learning-evaluation-collaboration" mode of planning intelligence. Future research should strengthen data fusion across physical, social, mobility, and institutional sources; design reward functions that make public values explicit; improve interpretability and causal reasoning; incorporate long-term feedback between infrastructure investment and urban evolution; and develop human-agent collaborative interfaces that allow planners to guide, contest, and refine algorithmic proposals. By addressing these challenges, this paradigm may contribute to more transparent, adaptive, equitable, and sustainable urban planning.

[Background] With the continuous deepening of urban digital transformation, traditional data management models have increasingly revealed shortcomings, such as “information silos” and inefficient decision-making support, in complex urban governance scenarios. These issues hinder the ability to meet the practical needs of refined governance. [Method] This study proposes a new paradigm: urban digital twin data volume, which aims to systematically address data fragmentation and the challenges of intelligent decision-making in urban governance through data lifecycle management and technological innovation. [Process] The study begins by explaining the concept and characteristics of the digital twin data volume, emphasizing that it is not merely a data aggregation, but an organically structured entity with inherent logical connections. A four-layer architectural system is then constructed, centered on spatial integration, information expansion, application-driven design, and intelligent empowerment: Through the 3D integration of urban base space, industry components, and socio-economic information, a comprehensive digital mapping foundation is established. By employing a hierarchical design of core and domain attributes, along with an object-oriented attribute chain association mechanism, the model enhances data flexibility and adaptability across scenarios. Scenario-driven modular data segmentation and dynamic recombination allow for the transformation of global data into customized and precise applications. Leveraging artificial intelligence, the model enables intelligent integration and deeper cognitive processing of multi-source data, achieving a closed loop of predictive analytics and optimized decision-making. Based on this data architecture, the study develops a smart parking collaborative governance scenario. By integrating data from multiple departments, including planning, transportation, commerce, and emergency management, it addresses the mismatch of parking resource and data silos through intelligent scheduling, real-time cross-domain data linkage, and dual-mode (normal/emergency) collaborative management. This validates the theoretical framework’s effectiveness in complex urban governance. [Prospect] Finally, the study outlines future directions for enhancing data security, privacy, and trusted computing, providing both theoretical support and practical pathways for evolving from digital twins to smart symbiosis cities.

[Objectives] Spatial planning is an important policy tool for guiding the optimized and sustainable development of territorial space, and it is a crucial basis for the future development, construction, and management of cities. However, due to differences in compilation standards and design units across various spatial levels and types of planning, a diversity and complexity of planning information have been created, leading to "planning conflicts." [Methods] This paper leverages the advantages of knowledge graphs in data organization and semantic relationship expression. By integrating planning elements at different levels and types, it explores the construction of an urban planning collaborative knowledge graph. By semantically linking the spatiotemporal attribute information of spatial planning resources, it achieves the integration and association of extensive urban planning information within a unified spatiotemporal framework, enhancing the overall effectiveness of multi-level and multi-type urban planning collaboration. [Results] Taking Shenzhen as a case study, a planning collaborative knowledge graph was developed, and a dual-perspective detection framework of "conflict manifestation-conflict source"was proposed. Empirical research accurately identified spatial overlaps between medical and ecological zones in Bao'an District, as well as attribute conflicts between industrial and forestry planning indicators in Longgang District, achieving quantitative detection and policy traceability of planning conflicts. [Conclusions] This paper argues that the "planning conflict" issue of spatial planning—an important basis for territorial space development and urban management—can be resolved through knowledge graph technology. The construction and application of the urban planning collaborative knowledge graph fully verify the feasibility of integrating knowledge graphs with urban planning. It not only builds a comprehensive database and simplifies management collaboration via visualization, but also promotes the organization, association and reconstruction of different plans by integrating data sources, providing technical support for element collaboration and conflict detection, and ultimately effectively improving the overall effectiveness of planning collaboration.

In recent years, urban region representation learning, as a vital technical approach for understanding and modeling urban spatial structures and functions, has gradually become a hot research topic at the intersection of Geographic Information Science and Artificial Intelligence. This field focuses on embedding complex and multi-source urban information into a unified representation space to support downstream tasks such as urban function identification, social behavior understanding, and economic indicator prediction. With the growing volume of multi-source urban data and the increasing complexity of modeling requirements, urban region representation learning plays a crucial role in urban computing, spatial modeling, and intelligent analysis. [Objectives] Despite remarkable advances in its applications, there is still a lack of systematic review articles in this field. To fill this gap, this paper provides a comprehensive synthesis that helps researchers grasp the development trajectory of this field, identify key scientific questions, and outline promising directions for future work. [Methods] The study builds an integrated analytical framework that links spatial unit construction, multimodal fusion, and structural co-modeling. At the spatial unit level, it compares alternative partition designs and analyzes the mechanisms of how granularity and boundary choices affect representation quality and task performance. At the multimodal fusion level, it investigates feature alignment and integration strategies for heterogeneous data sources, with emphasis on spatial and temporal alignment, robustness to missing or noisy inputs, and principled aggregation that preserves urban structure. At the structural modeling level, it compares multi-graph, heterogeneous-graph, and hypergraph approaches, and discusses the technical traits of prompt learning, adversarial learning, and meta-learning as complementary paradigms for task conditioning, robustness, and rapid adaptation. [Results] By integrating existing studies, the review reconstructs the evolutionary pathways of major techniques and clarifies their interrelations. It summarizes the field's common challenges, such as sensitivity to partition schemes, instability in multimodal fusion under misalignment and noise, and performance degradation under domain shifts. Moreover, it categorizes application scenarios and methodological characteristics, offering a clear mapping from problem settings to appropriate techniques, along with practical guidance for method selection and evaluation. [Conclusions] This paper further proposes the developmental trajectories of this field in the future. By highlighting potential breakthrough areas from the three dimensions of theoretical foundation, technical architecture, and application paradigms, it offers theoretical support and methodological guidance for the establishment of a new-generation urban spatial intelligent analysis system.

[Objectives] With the rapid advancement of personalized tourism services, tourism Point of Interest (POI) recommendation has emerged as a crucial technology for enhancing user experience and optimizing travel decision-making. However, existing POI recommendation methods often struggle to accurately capture users’ dynamic and context-dependent travel preferences due to several challenges, including the high sparsity of user check-in data, the complex spatiotemporal dependencies among interactions, and the dynamically changing nature of multi-source features. These issues hinder the effective modeling of user behavior and limit recommendation performance in real-world scenarios. [Methods] To address the above challenges, this study proposes a novel Meta-learning-based Cross-view Contrastive POI Recommendation model (MCC-POI) that integrates dynamic feature fusion with cross-view contrastive learning to achieve adaptive and robust recommendation. Specifically, a dynamic feature fusion mechanism based on meta-learning is designed to automatically adjust the relative importance of different types of features according to the task context and temporal variations. This enables the model to capture users' evolving preferences and contextual dependencies more effectively. By doing so, the model not only captures short-term behavioral shifts but also learns long-term preference evolution patterns across different temporal and spatial contexts, achieving a more fine-grained understanding of user intentions. Furthermore, a cross-view contrastive learning module with a symmetric constraint is introduced to jointly model the latent correlations between POI semantic attributes and geographical spatial information. This module effectively aligns heterogeneous feature representations across different modalities, thereby reducing feature redundancy and improving representation coherence. Such design enhances the alignment between multiple feature views and improves the generalization and robustness of the recommendation model under varying data distributions. [Results] Extensive experiments are conducted on four real-world benchmark datasets to evaluate the effectiveness of the proposed approach. The results demonstrate that MCC-POI significantly outperforms both traditional baseline methods(such as Popularity, PersTour, and POIRank) and state-of-the-art deep learning models (including DeepTrip, SelfTrip, and BertTrip). Compared with the best-performing baseline model (AR-Trip), MCC-POI achieves improvements of 1.6% to 6.0% in terms of F1 and pairs-F1 scores, verifying its superiority in capturing complex user-POI interactions. [Conclusions] In conclusion, the proposed MCC-POI framework effectively overcomes the limitations of static feature fusion in existing POI recommendation models by leveraging meta-learning for dynamic adaptation and cross-view contrastive learning for multi-modal representation alignment. The findings of this study provide a new perspective and methodological foundation for building intelligent, adaptive, and personalized tourism recommendation systems in complex and dynamic real-world environments.

[Objectives] Metaphor is an important rhetorical device in linguistics to map the abstract indigestible noumenon object using the intuitive vivid metaphorical object. This kind of rhetorical device can be introduced into map representation which usually acts as the second language of geo-science research to result in a new map type - metaphor map. This is a map-like visual representation of non-spatial information (semantic information). Within an artificially constructed virtual space, the relational, hierarchical, causal, and process-based characteristics of semantic information are mapped onto this virtual space and displayed through spatial graphic symbols, thereby achieving a vivid and intuitive spatial visualization experience. Based on the combination of geographic characteristics in ubiquitous mapping, this study examines the generation background of metaphor map as well as the conceptual connotation and denotation. [Methods] Aiming at the representation of its self-properties and inter-properties with each other, the study develops three design methods from the perspective of cartography principles, namely the generalization of semantic information, the dimension increasing from semantic space to geographic space and the symbol simulation of map-like. ① The generalization of semantic features tries to summarize the semantic information to be expressed by means of abstraction thinking and scale transformation. For the generalization of association features, causal features and temporal features, it simplifies the connotation of interacting semantic information and extracts the main influencing factors. ② The dimension increasing of semantic information is a key process in the "spatialization" of semantic information. The representation of semantic space is usually at low-dimensional: discrete entities in a set space are 0-dimensional, or semantics with linear extension are 1-dimensional. When transforming semantic space into geographic space, a dimension increase is required, which is exactly the opposite of the dimension reduction process of map projection in conventional mapping. Geometrically, the dimension increase transformation from 1-dimensional to 2-dimensional is related to the problem of "space-filling curve". ③ The Imitation of map-like symbols is to present vivid spatial characteristics with graphic representation. It follows Bertin's symbol parameter rule for conventional map symbol design, and establishes correspondence between each branch feature of semantic information and symbol graphic parameters. Metaphorical map symbol design should consider the semantic, syntactic, and pragmatic expression rules of linguistics at the same time. The organization law and structural pattern of simulated graphics should consider the syntactic rules such as arrangement and association of the semantic information to be expressed. [Results] Using the visualization of computer directory file, an example is built to show how design a metaphor map by the simulation of administration region map. [Conclusions] The metaphorical map provides a visualization method for the spatial representation of semantic information, allowing audiences to comprehend abstract information content through vivid spatial perception, thereby expanding the application of maps in other visualization fields.

[Objectives] With the rapid advancement of urbanization, urban spatial functions have undergone profound restructuring, gradually transforming from traditional single-function spaces to highly integrated and complex functional zones. In recent years, extensive studies have been conducted on the identification of static street functions and the analysis of dynamic travel mobility. However, most existing research focuses on either the physical attributes of urban streets or the characteristics of resident travel flows separately, and insufficient attention has been paid to the interactive relationship and internal coupling mechanism between street functions and mobility patterns. Traditional approaches often ignore the complex connections between physical space and human activities, making it difficult to reveal the interactive mechanism behind urban spatial structure evolution and travel pattern dynamics. [Methods] To fill this critical research gap, this paper proposes a comprehensive framework that integrates street function classification and travel mobility laws to explore their potential interrelationship. Firstly, multi-source urban data, including taxi trajectory data and street view imagery, are employed to extract dynamic travel behaviour characteristics and static physical environment features at the street segment level. These features are mapped into a dual graph structure, and a deep graph attention embedding model is used to achieve adaptive street clustering. Then, points of interest(POI) data are introduced, and the term frequency-inverse document frequency (TF-IDF) model is applied to quantitatively identify and classify dominant street functions. Finally, the DBSCAN algorithm is used to cluster travel origin-destination (OD) flows, and a series of statistical indicators are constructed to characterize mobility patterns. Correlation analysis is further conducted to reveal the quantitative relationship between street functions and mobility indicators. [Results] The empirical results in Jinan City show that the proposed method can effectively identify five types of mixed functional streets: leisure-commercial, commercial-public service, administrative-medical, work-living, and mobility-living streets. Among them, the commercial-public service streets occupy a core position in the urban spatial structure, serving as the key hub for urban public activities and traffic connections, and the differentiated mobility patterns of other functional streets form an effective linkage with them, jointly promoting the orderly operation of the urban street system and the rational distribution of travel flows. Further spatial analysis demonstrates that streets in Jinan present an obvious hierarchical structure corresponding to different functional combinations and travel intensity distributions. [Conclusions] These conclusions can provide valuable insights and decision support for the optimization of street spatial structures, the scientific allocation of urban public facilities, and the high-efficiency governance of modern cities.

[Objectives] Marine environmental variables are commonly represented as three-dimensional scalar fields distributed over curved Earth space and sampled along strongly non-uniform vertical layers. Conventional volume rendering methods often rely on Cartesian proxy boxes, spherical approximations, or fixed-step ray marching. When applied to WGS84 ellipsoidal geometry and realistic ocean-depth structures, these methods may cause geometric mismatch, invalid sampling, distortion from depth linearization, loss of shallow-layer details, and redundant computation in deep waters. To improve geographic consistency, vertical structural fidelity, and rendering efficiency, this paper proposes a geometry-aware and layer-adaptive volume rendering method for oceanic three-dimensional scalar fields under WGS84 ellipsoid constraints. [Methods] The proposed framework integrates ellipsoidal proxy construction, non-uniform depth indexing, and adaptive ray marching into a WebGL2-based rendering pipeline. First, the original longitude-latitude-depth data are organized as three-dimensional textures while preserving the native vertical layers of the CMEMS dataset. A physical-depth-to-texture-coordinate mapping strategy based on depth-layer lookup and intra-layer interpolation is designed to avoid structural distortion caused by forced equal-depth resampling. Second, a dual-layer ellipsoidal shell proxy geometry is generated using WGS84 parameters to define the effective sampling interval of ocean volume data in Earth-centered Cartesian space. This proxy constrains the ray marching domain within the real geographic extent of the ocean volume, reducing curvature mismatch and invalid sampling while maintaining spatial correspondence between rendered structures and real geographic locations. Third, a layer-aware ray marching strategy dynamically adjusts the local sampling step according to normalized vertical layer spacing, increasing sampling density in shallow dense layers and reducing redundant computation in sparse deep layers. Thus, the method balances shallow-layer feature preservation with interactive visualization efficiency. [Results] Experiments using the CMEMS temperature field over the China Seas show that the proposed method effectively preserves the curved geographic structure of the study area and avoids vertical distortion caused by depth linearization. The experimental grid contains 451×313×50 samples, corresponding to approximately 7.06 million voxels. Representative depth-layer reconstruction results present low RMSE values, indicating that the original scalar-field structure can be maintained during texture organization, coordinate mapping, and rendering. Compared with fixed-step ray marching, the adaptive strategy achieves similar visual quality while increasing the average frame rate by approximately 24.9%. Web-based preprocessing, transmission, parsing, and initialization tests further indicate that the workflow can support million-level to nearly ten-million-level voxel data volumes with second-level initialization performance, demonstrating its feasibility for browser-based ocean volume visualization. [Conclusions] By coupling WGS84 ellipsoidal geometry with ocean-specific non-uniform vertical stratification, the proposed method improves spatial consistency, vertical structural fidelity, and interactive rendering efficiency. It provides a scalable solution for accurate visualization of large-scale non-uniform marine scalar fields and can support digital twin ocean applications.

[Objective] Few-shot object detection in remote sensing images is of great significance for disaster monitoring and military reconnaissance. However, due to complex backgrounds, large scale variations of targets, high inter-class similarity, and the scarcity of annotated samples, existing models are prone to overfitting and often achieve low detection accuracy on novel classes. Therefore, this paper proposes a feature-optimization-based few-shot object detection method for remote sensing images. [Methods] First, an Adaptive Dense Fusion Pyramid Network (ADFPN) is designed in the neck of the detector. Through feature preprocessing and dense feature fusion, this module enables sufficient interaction among multi-scale features while preserving information integrity during cross-scale fusion, thereby providing more accurate feature representations for the subsequent detection head. Second, a Deformable Region Proposal Network (D-RPN) is proposed. By introducing a convolutional kernel offset prediction module, it achieves dynamic and adaptive adjustment of sampling points, enabling the network to flexibly adjust the receptive field according to the geometric characteristics of the target object, thereby enhancing the model’s localization capability for irregularly shaped objects. Finally, a Cross-Hierarchical Fusion Region of Interest extractor (CHF-RoI) is proposed, which encourages the model to jointly attend to both the target itself and its contextual cues during training. By leveraging contextual information, this module effectively alleviates overfitting and improves detection performance. [Results] Comparative and ablation experiments are conducted on the NWPU VHR-10.v2 and DIOR datasets. On NWPU VHR-10.v2, compared with the baseline Meta-RCNN, the proposed method improves the novel-class mAP (mean Average Precision) by 2.1%, 1.5%, 6.4%, and 6.2% under the 2-shot, 3-shot, 5-shot, and 10-shot settings, respectively. Under split-1 of DIOR, the proposed method achieves higher novel-class mAP than Meta-RCNN across all few-shot settings; under split-2, it also obtains competitive novel-class mAP performance. Moreover, comparisons with classic methods such as FSCE, TFA, and P-CNN demonstrate that the proposed approach achieves overall superior novel-class mAP and exhibits clear advantages over these existing methods. Ablation studies further verify that the three proposed modules consistently improve the model's capability of detecting novel classes in few-shot remote sensing object detection, confirming the necessity and effectiveness of each component. [Conclusions] The proposed modules effectively improve novel-class detection accuracy under few-shot settings, demonstrating that the proposed method can alleviate the performance degradation caused by limited samples in complex scenes and achieve strong detection accuracy and generalization capability. Therefore, this method provides an effective solution for few-shot object detection in remote sensing imagery, with promising potential for practical applications in disaster monitoring and military reconnaissance.

[Objectives] High-resolution remote sensing imagery has become an important data source for urban planning, land-use monitoring, environmental assessment, and disaster management. However, the significant variations in object scales, complex spatial layouts, and blurred object boundaries in such imagery pose considerable challenges to semantic segmentation. Although existing lightweight semantic segmentation methods achieve favorable computational efficiency, most of them simplify the decoding stage as a progressive resolution recovery process and pay insufficient attention to the structural relationships among multi-level features. As a result, they often suffer from structural fragmentation in complex scenes, insufficient representation of small objects, and degraded boundary quality. To address these limitations, this study introduces a structure-recovery perspective into the decoding process and formulates semantic decoding as a hierarchical structure reconstruction procedure. [Methods] A lightweight semantic segmentation network, termed Selective Alignment Fusion Network (SAF-Net), is proposed for high-resolution remote sensing image interpretation. The proposed framework follows a structure-recovery-driven progressive decoding paradigm and consists of three key stages. First, a Spatial Reconstruction Module (SRM) is designed to reconstruct the spatial continuity of high-level semantic representations and enhance global structural perception. Second, a Selective Alignment Fusion Module (SAFM) is introduced to alleviate semantic inconsistencies between adjacent feature levels through adaptive semantic alignment and selective information interaction, thereby improving cross-level feature aggregation. Third, a Multi-Scale Feature Recovery Module (MSFRM) is employed in shallow decoding stages to capture fine-grained details at multiple receptive fields, while a boundary-constrained refinement strategy is incorporated to preserve object contours and improve boundary quality. Through these components, the decoder evolves from conventional scale restoration into a structure-aware optimization framework that jointly preserves semantic consistency, structural integrity, fine-grained details, and boundary continuity. [Results] Extensive experiments were conducted on three publicly available benchmark datasets, including ISPRS Vaihingen, ISPRS Potsdam, and LoveDA. The proposed SAF-Net achieved mIoU scores of 84.57%, 86.89%, and 51.04%, respectively. Compared with the lightweight baseline LightFormer, SAF-Net improved mIoU by 1.65%, 0.59%, and 3.02% on the Vaihingen, Potsdam, and LoveDA datasets, respectively. Furthermore, comparative evaluations against representative lightweight networks, Transformer-based methods, and recent Mamba-based approaches demonstrate the strong competitiveness of SAF-Net in terms of both segmentation performance and model efficiency. Ablation studies further reveal the contribution of each proposed component, while visualization analyses verify that the structure-recovery mechanism effectively enhances cross-level semantic coordination, restores complex spatial structures, improves small-object representation, and preserves boundary continuity. [Conclusions] By reformulating semantic decoding as a hierarchical structure recovery process, this study provides a new perspective for lightweight decoder design in high-resolution remote sensing semantic segmentation. The proposed SAF-Net effectively balances accuracy and efficiency while achieving superior structural representation capability. The experimental results demonstrate that structure-recovery-driven decoding provides an effective and generalizable paradigm for lightweight semantic segmentation, offering new insights into decoder design for high-resolution remote sensing imagery.

[Objectives] Remote Sensing Change Detection (RS-CD) aims to identify land-cover changes by comparing remote sensing images acquired over the same area at different times. It is widely used in urban expansion monitoring, disaster assessment, land-use investigation, and ecological environment monitoring. In complex remote sensing scenes, changed objects often show large scale variations, irregular boundaries, and diverse spatial distributions. At the same time, illumination differences, shadows, seasonal variations, and local registration errors may introduce false changes. Existing methods still have limitations in bi-temporal difference representation and cross-level feature fusion, which can lead to missed detection of small changed targets, incomplete extraction of large changed regions, inaccurate boundary localization, and false alarms in background areas. To address these problems, this paper proposes a remote sensing image change detection method for complex scenes. [Methods] A Hierarchical Difference Fusion Network (HDFNet) is proposed. In the encoding stage, a Multi-Scale Difference Enhancement Module (MSDE) is introduced to strengthen difference representation at multiple feature levels. According to the representation characteristics of shallow and deep features, the Efficient Multi-scale Difference Enhancement module (EMDE) and the Coordinate-aware Difference Enhancement module (CADE) are adopted respectively. EMDE is used to enhance fine-grained change responses and preserve local boundary details in shallow features, while CADE is used to improve semantic structural consistency and spatial localization in deep features. In the decoding stage, a Dual Feature Fusion module (DFF) is designed to improve cross-level feature interaction. Specifically, Channel Self-Attention Gating (CSAG) is used to select effective semantic channels and suppress redundant information, and the Multi-Scale Spatial Mixer (MSSM) is used to enhance spatial alignment and multi-scale spatial information interaction. In addition, a weighted multi-loss joint optimization strategy is adopted to improve the adaptability of the model to complex backgrounds and the discriminative ability of changed regions. [Results] Experiments were conducted on four public datasets, including LEVIR-CD, WHU-CD, CLCD, and SYSU-CD. The F1 of HDFNet reached 91.31%, 94.05%, 78.73%, and 83.46%, respectively, and the IoU values reached 84.01%, 88.77%, 64.92%, and 71.62%, respectively. Compared with the recent representative method HA2F, HDFNet improved the F1 by 0.69%, 1.76%, 1.96%, and 0.89% on the four datasets, and improved the IoU by 1.16%, 3.09%, 2.63%, and 1.31%, respectively. Visualization results show that HDFNet performs better in suppressing complex background interference, recovering boundary details, and preserving the integrity of large-scale changed regions. Ablation experiments further verify the effectiveness of the multi-scale difference enhancement strategy and the dual feature fusion strategy. [Conclusions] The results indicate that HDFNet can effectively improve the accuracy and robustness of remote sensing change detection in complex scenes.

[Objectives] Cloud detection is a critical preprocessing step in optical remote sensing image processing, and generating cloud-free base maps holds substantial significance for land-use classification, change information extraction, and long-term time-series land-cover monitoring. In recent years, deep learning-based cloud detection methods have demonstrated pronounced improvements in accuracy across complex regions with mixed high-brightness backgrounds compared to traditional feature-based approaches. With the rapid evolution of frontier applications such as satellite on-orbit computing, onboard cloud detection has emerged as a prominent research hotspot. However, constrained by the limited computing power of onboard devices, existing deep learning models exhibit limitations regarding parameter volume and computational resource consumption, thereby hindering their further deployment on resource-constrained platforms. [Methods] To address these limitations, this paper constructs a lightweight cloud detection network for optical remote sensing imagery, termed LWCloudNet, based on the general vision foundation model DINOv3 in conjunction with model optimization techniques like knowledge distillation. By decoupling the training and inference stages, the proposed model simultaneously assimilates general semantic knowledge from DINOv3 and cloud-specific features during the training phase, achieving high-efficiency inference and high-precision cloud detection for the lightweight model during inference. Concurrently, the network integrates the local feature extraction capability of Convolutional Neural Networks (CNNs) with the long-range dependency modeling advantages of Transformers, theoretically enhancing the deep learning model's capability to detect morphologically diverse cloud targets. Verification experiments were conducted using remote sensing datasets from domestic optical satellites, including Gaofen-1 and SuperView-1. Furthermore, mobile-end experimental testing was carried out on resource-constrained devices, such as the Jetson platform. [Results] The experimental results demonstrate that the proposed lightweight cloud detection model can precisely segment cloud-covered regions. In complex areas such as cloud pixel boundaries, it achieves superior accuracy compared to existing models, yielding an Intersection over Union (IoU) improvement of 2.86% over the lightweight model STCNet. In terms of inference efficiency on the Jetson mobile device, the proposed model completes the processing of a single-scene Gaofen-1 PMS image within 1.34 seconds, marking a significant efficiency increase of 17.62% compared to FastViT. Regarding distillation strategies, the multi-scale knowledge distillation framework proposed in this paper achieves a 2.08% increase in IoU compared to conventional guided feature distillation methods. [Conclusions] The lightweight cloud detection model developed in this study is expected to provide essential preprocessing technical support for intelligent on-orbit satellite information interpretation, while offering a theoretical reference for high-precision cloud detection executed on low-computing-power devices.

[Objectives] Sea ice monitoring is of great significance for ensuring winter shipping safety and marine engineering maintenance. [Methods] Aiming at the problem of feature extraction difficulties caused by the diversification of sea ice scales, as well as the low ice-water discriminability caused by complex background interference, this paper proposes a SegFormer-based sea ice extraction method integrating multi-source physical features and dual attention. In order to solve the problem of feature extraction difficulties caused by the diversification of sea ice scales, an SCSA-Inception module is designed in the encoder. By combining the Inception multi-scale convolution and the Spatial and Channel Synergistic Attention (SCSA) mechanism, it effectively balances the global context information of large-area ice sheets with the local details of fragmented drift ice, thereby enhancing the feature extraction capability for sea ice at different scales. To address the issue of low ice-water discriminability caused by complex background interference, this paper optimizes the model from three aspects: data input enhancement, network feature decoding, and loss function constraint. First, by combining the physical characteristics of visible light bands, a physical information-enhanced input—including intensity, a variant of the Normalized Difference Water Index (NDWI), and the value component—is constructed to reduce ice-water spectral confusion and improve the surface optical contrast. Subsequently, a CPCA-guided attention module is designed in the decoder, utilizing Channel Prior Convolutional Attention (CPCA) to adaptively allocate feature weights, amplify sea ice feature responses, and weaken coastal land texture noise, effectively suppressing complex background interference. Finally, a boundary-aware compound loss function is designed. By explicitly extracting sea ice boundary information using the Sobel operator and combining it with regional semantic loss, it prompts the model to optimize the extraction details of sea ice edges while paying attention to overall classification accuracy, thereby synergistically enhancing ice-water discriminability under complex backgrounds. [Results] Taking Liaodong Bay as the experimental area, validation is conducted on the constructed HY-1C/D satellite sea ice dataset. The experimental results demonstrate that the method proposed in this paper achieves 81.51%, 90.28%, and 90.98% in the mIoU, mPA, and Accuracy metrics, respectively. Particularly in the core evaluation metric mIoU, compared with current advanced models such as SAM-RS, SCTNet, CGGLNet, and SegFormer, it improves by 14.08%, 13.50%, 7.20%, and 5.88%, respectively, demonstrating good extraction accuracy and robustness. [Conclusions] This multi-module method effectively overcomes the challenges of diverse ice scales and complex backgrounds, providing an efficient and precise monitoring tool for winter maritime safety and engineering maintenance.

[Objectives] To address the limitations of single-source remote sensing data in distinguishing coniferous and broadleaf forests under complex mountainous environments—where phenological variability and topographic heterogeneity significantly affect classification accuracy—this study proposes a fine-scale forest type classification framework based on the integration of multi-source remote sensing features and a Random Forest (RF) algorithm. By incorporating a staged feature selection strategy that balances multi-source information synergy and high-dimensional feature optimization, this study aims to achieve high-precision mapping of coniferous and broadleaf forests and to further characterize their spatial distribution patterns across heterogeneous terrain. [Methods] The Qinling Mountains of China were selected as the study area due to their pronounced environmental gradients and diverse forest composition. Multi-dimensional feature sets were constructed by integrating Landsat 8 optical imagery and PALSAR-2 polarimetric synthetic aperture radar (SAR) data acquired during 2020-2022, together with digital elevation model (DEM) data. These features encompass basic spectral bands, multi-temporal spectral features, basic vegetation indices, multi-temporal vegetation indices, texture metrics derived from gray-level co-occurrence matrices, polarimetric decomposition features, and topographic factors including elevation, slope, and aspect. The RF model was employed for feature selection and classification modeling, while a staged feature selection scheme incorporating feature importance ranking, accuracy response analysis, and correlation constraints was developed to achieve dimensionality reduction and key variable optimization. Permutation importance was utilized to quantify the relative contributions of different feature types. In addition, terrain gradient analysis was conducted to explore the spatial distribution characteristics of coniferous and broadleaf forests in relation to topographic conditions. [Results] (1) Spectral bands and vegetation indices play a dominant role in forest type discrimination. The Difference Vegetation Index (DVI), the Normalized Difference Vegetation Index (NDVI), and the near-infrared band exhibit the highest importance. The incorporation of multi-temporal features and topographic variables significantly enhances classification robustness under complex mountainous conditions, highlighting the complementary advantages of multi-source data, forming a synergistic mechanism characterized by spectral dominance, temporal enhancement, structural supplementation, and topographic constraint. (2) Accuracy assessment indicates that the overall accuracies of forest type classification for the years 2020, 2021, and 2022 reach 0.976, 0.942, and 0.928, respectively, with corresponding Kappa coefficients all exceeding 0.90. The producer's accuracy for coniferous forest identification in this study was consistently higher than 0.94, while the user's accuracy was consistently above 0.90. In contrast, both the producer's accuracy and user’s accuracy of coniferous forests in the MODIS MCD12Q1 product were 0, whereas the highest producer's accuracy and user's accuracy of coniferous and broadleaf forests in the GLC_FCS30 product were 0.924 and 0.864, respectively. Overall, compared with the other two products, this study achieved higher classification accuracy and stronger spatial detail representation in complex mountainous environments, effectively reducing the omission of fragmented forest patches in mountainous areas. (3) The spatial distribution of forest types exhibits clear vertical zonation. Broadleaf forests are predominantly distributed in mid-mountain zones below 2 000 m, while coniferous forests increasingly dominate above 2 500 m with rising elevation. Moreover, coniferous forests exhibit a higher distribution proportion on shady and semi-shady slopes. [Conclusions] The integration of multi-source remote sensing features with a staged feature selection strategy provides a robust and reliable methodological basis for forest resource monitoring, ecological assessment, and sustainable management in the Qinling Mountains and similar environments worldwide.

[Objectives] To address the insufficient identification of temporal turning points, spatial reorganization, and hotspot activation of motor vehicle theft risk under structural change, this study takes City X as the study area to reveal the spatiotemporal evolution process and environmental associations of motor vehicle theft risk after phased change. [Methods] Based on motor vehicle theft data in City X from 2016 to 2024, segmented interrupted time-series analysis was first used to identify the structural breakpoint of the monthly motor vehicle theft series. Then, using 500 m × 500 m grids as the analytical units, grid-level increment contribution comparison and local Getis-Ord Gi* statistics were combined to characterize the spatial concentration of risk growth and the evolution of hotspots before and after the breakpoint. Finally, a binary Logit model was constructed within high-accessibility areas to examine the relationships between pre-breakpoint traffic volume, point-of-interest density, functional mixedness, network centrality, distance to the nearest police station, baseline crime level, and post-breakpoint hotspot activation. [Results] Taking City X as the experimental area, the empirical analysis based on motor vehicle theft data from 2016 to 2024 shows that motor vehicle theft in City X experienced a clear structural turning point around August 2022. After the breakpoint, risk growth was mainly concentrated in a small number of spatial units. The top 10% and top 20% of grids in terms of average monthly increment contributed 50% and 72% of the citywide average monthly increase, respectively. The hotspot pattern showed spatial reorganization characterized by the continued reinforcement of existing high-risk areas, expansion into adjacent areas, and activation of some potential areas. Within high-accessibility areas, 118 grids changed from non-significant hotspots to significant hotspots, accounting for 19.87% of high-accessibility grids. The model results indicate that point-of-interest density and functional mixedness were significantly positively associated with hotspot activation probability, whereas traffic volume showed no significant effect. [Conclusions] The study shows that motor vehicle theft risk after structural change did not increase uniformly across the whole city, but rather manifested as a spatially selective process of local reinforcement. High accessibility is closer to a background condition for the release of motor vehicle theft risk, while point-of-interest concentration and functional mixedness better explain the transformation of potential risk into manifest hotspots. The analytical path constructed in this study, consisting of structural breakpoint identification, hotspot state transition, and activation condition testing, can provide methodological reference for potential hotspot identification and prospective risk assessment in crime research in China. For domestic prevention and control practice, this framework can be combined with local case data, parking organization, community management, video surveillance, property governance, and police patrol conditions to identify key places within high-accessibility areas that are more likely to undergo risk activation, thereby providing support for refined prevention and control of motor vehicle theft and similar opportunity-based crimes.

[Objectives] Under the dual background of population aging and the urban heat island effect, the elderly face increasingly prominent health risks of heat exposure. Existing studies typically treat the elderly as a generalized whole, overlooking the spatial coupling pattern formed by the concentration of older populations in high-density aged neighborhoods, while environmental attribution research on heat exposure across blocks of different aging degrees remains insufficient. This study aims to reveal the spatial differentiation of heat exposure across blocks of varying aging degrees and to identify the differential driving mechanisms of the built environment. [Methods] Taking 4 795 blocks within Beijing's Fifth Ring Road as the research unit, we integrated remotely sensed land surface temperature, the ASPECT elderly population dataset, and multi-source built environment data. The surface urban heat island intensity was calculated and coupled with the elderly population to construct a block-scale Heat Exposure Index (HEI). Blocks were classified into mild, moderate, and severe aging categories based on the aging rate. Bivariate spatial autocorrelation was applied to reveal the spatial coupling between aging and heat exposure. An indicator system of 11 variables across four dimensions—street form, building form, green space, and facility accessibility—was constructed, and a random forest model integrated with SHAP interpretability was employed to quantify the nonlinear contributions of built environment factors. [Results] (1) HEI increased significantly with the aging degree, with mean values of 8.80, 11.76, and 14.12 for mild, moderate, and severe blocks, respectively, and a bivariate Moran's I of 0.345 9 indicating significant spatial coupling. (2) Environmental attribution exhibited gradient differentiation along built environment intensity. The green space dimension contributed 36.7% in mild blocks (ranking first), green space and building height became equally important in moderate blocks, while the building morphology dimension rose to 41.4% in severe blocks—approximately 12% higher than the 29.1% contribution of the green space dimension. (3) The random forest model outperformed traditional regression in all three categories. SHAP analysis revealed that the unit cooling benefit of NDVI was more pronounced in severe blocks but constrained by overall low vegetation levels; the positive effect of block area weakened in severe blocks; and building coverage exhibited a distinct threshold transition. [Conclusions] Aging and heat exposure show significant spatial coupling within Beijing's Fifth Ring Road, with built environment driving mechanisms exhibiting gradient differentiation across aging degrees. These findings provide scientific support for the thermal environment optimization of age-friendly cities, while the integration of classification modeling with the RF-SHAP framework offers a new analytical paradigm for urban thermal environment attribution analysis.

[Objectives] Academic publications contain rich semantic information and multidimensional spatiotemporal characteristics that reflect the evolution of a discipline. Traditional bibliometric methods, however, often rely on structured indicators such as keywords, publication counts, and author information, which limits their ability to deeply mine the semantic content of unstructured texts. In addition, they have difficulty quantitatively revealing how external attributes, such as time and region, influence topic evolution. To address these limitations, this paper aims to construct a multidimensional semantic mining framework for journal literature based on the Structural Topic Model (STM). [Methods] Taking the Journal of Geo-information Science as a case study, this paper selected the abstracts of 3,523 papers published from 1996 to 2024 as empirical samples to verify the effectiveness and applicability of the proposed framework. After data collection and preprocessing, STM was used to identify latent topics from the abstracts, while external covariates were introduced to analyze the influence of time and space on topic distribution. The optimal number of topics was determined by combining semantic coherence, exclusivity, and topic correlation. On this basis, year was incorporated as a covariate to explore topic evolution over time, while regional and institutional information was used to examine the spatial heterogeneity of research themes. In addition, the thematic structure and development trajectories of the journal were compared with those of mainstream international journals in the field. [Results] The empirical results show that the framework has strong capabilities in temporal, spatial, and comparative analysis. First, it effectively extracted 15 core latent themes and accurately captured the rapid growth of emerging hotspots such as "trajectory data analysis" and "spatial network analysis", as well as the maturity characteristics of foundational topics such as "spatial data management". At the same time, the discipline has undergone a paradigm shift from GIS engineering to GIScience and then to GeoAI. Second, by quantifying the spatial differences in topic distributions, the framework reveals the research preferences of different universities and research institutes, as well as the dual driving forces of research resource concentration and region-oriented practical demands. Third, through a comparison between Chinese and international development paths, the study finds that the Journal of Geo-information Science emphasizes an application-oriented and integrative innovation path closely related to major national strategies. [Conclusions] The analytical framework proposed in this paper breaks through the limitations of one-dimensional bibliometric analysis. It can systematically deconstruct disciplinary development from three dimensions: temporal topic evolution, spatial patterns, and the comparison between Chinese and international journals. It provides a scientific and reusable quantitative paradigm for academic evaluation and discipline planning in geo-information science and other scientific fields.