Spatio-temporal information serves as a fundamental base and knowledge repository in digital economic activities such as logistics distribution and navigation services. However, 2D to 2.5D spatio-temporal information has progressively proven insufficient to cater to emerging digital economy scenarios, such as the low-altitude economy, the integration of culture and tourism, and embodied intelligence. To promote the transformation of the digital economy from a 2D paradigm to a 3D framework, it is essential to integrate 3D realistic geospatial scene data with digital economy scenarios, thereby forging a novel form of digital economy — the 3D Economy. This study pioneers the introduction of the 3D Economy concept, systematically dissecting its core technical implications, principal development trajectories, immediate priority tasks, and implementation strategies. To be specific, the 3D Economy positions 3D realistic geospatial scene data as its essential productive factors, spatio-temporal computing platforms as its primary infrastructure, spatio-temporal intelligent algorithms as its core enabler, and diverse 3D economic application scenarios as its principal substance. In fact, the 3D Economy represents a new digital economy paradigm distinguished by 3D spatial dimensions. Its core implications encompass the digitization of the physical 3D world, the assetization of spatio-temporal data assets, the intelligentization of computing platforms, and the contextualization of empowering applications. The major development directions of the 3D Economy are summarized as cultivating new-quality productive forces anchored in 3D realistic scenes, reshaping business paradigms, and pioneering development pathways for the 3D Economy. The immediate priority tasks involve constructing a trustworthy 3D data space, developing spatio-temporal intelligent computing platforms, and fostering an open and innovative technological ecosystem.
The development of the 3D Economy holds substantial promise yet demands a protracted and arduous endeavor. It will not only unlock the full value potential of 3D data but also secure a first-mover advantage and forge core competitiveness in domains including the intelligent economy and geospatial intelligence. To this end, it is imperative to enhance scientific insight, reinforce technological innovation, foster exchange and cooperation, and propel the profound integration of scientific and technological innovation with industrial development.
[Objective] With the rapid acceleration of urbanization in China, a substantial number of high-rise buildings have entered their “mid-life phase”, during which exterior wall hollowing and spalling pose significant safety hazards. Conventional manual inspection methods are characterized by low efficiency, high operational risks, and an inability to meet the demands of large-scale, routine monitoring. This study aims to develop an efficient and automated detection technology for building exterior wall cracks by integrating unmanned aerial vehicle (UAV) imagery with deep learning. The core objective is to systematically evaluate the applicability and performance limitations of various deep learning architectures for crack segmentation in UAV-acquired images, specifically addressing the inherent challenge of balancing the preservation of detailed features with the modeling of long-range structural context. Ultimately, this paper endeavors to provide empirical evidence to guide model selection in practical engineering applications.
[Method] A hierarchical comparative framework for model evaluation was established, with the standard U-Net serving as the baseline. To progressively enhance contextual modeling capabilities, a series of improved networks was developed by sequentially incorporating residual connections, attention mechanisms, and Transformer modules into the U-Net architecture. Furthermore, the pure Transformer-based SegFormer and the atrous convolution-based DeepLabV3+ were introduced as representative mainstream advanced paradigms for external benchmarking. All models were trained and evaluated under a unified experimental protocol utilizing a self-constructed, professional crack dataset, thereby ensuring a fair comparison of their efficacy. Additionally, an ablation study on loss functions was conducted to analyze the performance of various loss functions—including BCE, WCE, Focal loss, F1 loss, and combined losses (BCE+Dice loss, Focal+Dice loss)—in scenarios characterized by extreme class imbalance.
[Result] Regarding model architecture, the proposed TransUNet achieved the optimal overall segmentation performance, registering an IoU of 81.06% and an F1 score of 89.48%. Its success is attributed to the effective synergy between the Transformer module’s global contextual modeling and the U-Net’s local detail preservation via skip connections. Residual U-Net demonstrated a balanced trade-off between performance and robustness, rendering it a suitable candidate for practical deployment. In contrast, SegFormer performed significantly inferiorly (IoU of only 4.76%), indicating a fundamental limitation of pure Transformer architectures in extracting features of fine, linear structures such as cracks without the inherent local inductive bias of convolutions. This result underscores the necessity of hybrid architectures for this specific task. In the loss function experiments, F1 loss delivered the best-balanced performance under class imbalance (IoU 77.59%). Notably, while Focal Loss achieved the lowest training loss (0.0117), it yielded the poorest segmentation performance (IoU 53.79%), clearly demonstrating that a reduction in training loss does not equate to an improvement in model segmentation capability, thus necessitating evaluation based on task-specific metrics such as IoU and F1 score.
[Conclusion] Through a systematic and hierarchical comparative experiment, this paper clarifies the performance characteristics and applicable scenarios of different deep learning architectures for UAV-based crack detection. The results confirm the superiority of hybrid architectures that fuse global modeling with local perception, such as TransUNet, for this task, while revealing the inherent limitations of pure Transformer architectures in recognizing subtle, linear structures. The study provides practical guidelines for engineering model selection: TransUNet is recommended for scenarios prioritizing high precision; Residual U-Net is suitable for routine inspections requiring a balance of stability and efficiency; the application of SegFormer necessitates careful architectural adaptation; and DeepLabV3+ requires targeted adjustments to training strategies. Future work will focus on constructing more diverse and extensive crack datasets, developing lightweight yet efficient hybrid models, enhancing cross-scene generalization capabilities, and promoting the practical application and technological transformation for intelligent structural health monitoring.
[Objective] In the field of remote sensing change detection, a fundamental contradiction persists within datasets: high-quality datasets are constrained in size due to the substantial human and temporal resources required for annotation, whereas large-scale datasets frequently suffer from mislabeling and omissions, resulting in inconsistent sample quality. This paper aims to optimize the training process to amplify the role of high-quality samples, mitigate the interference of low-quality samples, and enhance model detection accuracy. Furthermore, it addresses the inadequate screening of mislabeled and omitted samples in large-scale, low-quality datasets within existing research.
[Method] This study proposes a method that integrates a data dynamic evaluation model (data_evaluator) with a dynamic sample weight adjustment mechanism grounded in forgetting events. The data_evaluator shares the same structural architecture as the main network but operates with 1/4 of the channels and 1/16 of the parameters. Every 10 epochs, all training samples are evaluated, sorted by mIoU, and assigned corresponding loss weights. The bottom-scoring 40% of samples are excluded, while the remaining samples are categorized into easy-to-learn and difficult-to-learn groups. The forgetting event mechanism identifies a sample as undergoing a forgetting event if it was correctly predicted in epoch e–1 but incorrectly predicted in epoch e, recording the number of such occurrences, k. Based on this, a forgetting weight is calculated: the weight of samples that have never been forgotten is reduced, while the weight of those occasionally forgotten is increased. The final sample loss weight is derived by proportionally fusing the evaluation weight and the forgetting weight. Additionally, a hybrid loss function combining Focal Loss and Dice Loss is employed to alleviate the issue of class imbalance. Recall, precision, F1 score, and IoU are adopted as evaluation metrics. Experiments are conducted using FC-Siam-conc, FC-Siam-diff, and FC-EF as base models, with the SYSU-CD and LEVIR-CD+ datasets partitioned into training, validation, and test sets in a 7︰1︰2 ratio. The experimental settings include 100 epochs, a batch size of 8, the AdamW optimizer, an initial learning rate of 0.005, and sample weight adjustments every 10 epochs.
[Result] The experimental results demonstrate that, on both the SYSU-CD and LEVIR-CD+ datasets, the FC-Siam-conc model employing the proposed method exhibits higher accuracy than the original FC-Siam-conc model. Visualization results indicate that the original model suffers from both missed detections and false detections, with the latter being more prevalent. The proposed method effectively reduces false detections and accurately identifies a greater number of genuine change areas, thereby achieving a notable improvement in accuracy.
[Conclusion] The dynamic sample weight adjustment method proposed in this paper, through the data evaluation model for sample selection and the forgetting event mechanism for weight adjustment, effectively enhances the accuracy of remote sensing change detection models. It attains IoUs of 70.47% and 82.58% on the SYSU-CD and LEVIR-CD+ datasets, respectively. However, the method possesses certain limitations. Firstly, the experiments are conducted using high-quality public datasets, and the accuracy improvement is limited, failing to fully validate the method’s effectiveness on large-scale, low-quality datasets. Secondly, the method exhibits low sensitivity to samples with a minor proportion of errors, making it difficult to identify such samples. Future research should focus on larger-scale, low-quality datasets and optimize the method to improve its capability to identify samples with a low proportion of errors.
[Objective] Pedestrian navigation networks serve as essential supplements to vehicular road systems, facilitating walking, hiking, and other slow-mobility activities in outdoor environments. However, the large-scale construction of such networks faces significant challenges: professional surveying techniques are costly and susceptible to environmental occlusions (e.g., tree canopies), while existing trajectory-based methods predominantly focus on extracting 2D geometric structures, neglecting critical 3D attributes such as elevation and slope. These attributes are fundamental for accurate outdoor path planning, particularly in terrains characterized by complex topography. To address this gap, this paper proposes a hierarchical method for constructing 3D pedestrian navigation road networks from crowdsourced outdoor trajectory data, aiming to enhance both geometric accuracy and topological completeness.
[Method] The pedestrian network is modeled as a graph comprising walking areas, intersections, and connecting segments. The methodology encompasses three principal steps. First, raw trajectories undergo preprocessing, including data cleaning, resampling, and DEM-assisted elevation correction. Specifically, anomalous points are removed via mean filtering; trajectories are aggregated using a gravity-repulsion model to mitigate positioning errors; and a DEM with 12.5 m resolution is utilized to correct elevation drift by calculating average deviations and filtering outliers. Second, a direction-constrained density-based clustering algorithm is introduced to identify walking areas and intersections. This algorithm extends DBSCAN by integrating both spatial proximity and trajectory direction similarity, thereby classifying grid cells based on the number and orientation of trajectory clusters. Walking areas are characterized by chaotic direction distributions, intersections by multiple dominant directions, and road segments by one or two consistent directions. Third, 3D road topology is reconstructed through hierarchical extraction: intersection entry/exit points are identified via trajectory clipping and clustering; road centerlines are extracted by clustering trajectories based on start-end region similarity and Hausdorff distance, followed by 3D B-spline fitting; finally, topological optimization connects road segments with intersections and walking areas to form a complete network.
[Result] Experiments were conducted using 3406 hiking trajectories collected from the Yuelu Mountain Scenic Area in Changsha, China, between January 2015 and June 2023. The trajectory data exhibited an average elevation deviation of 14.5 m relative to DEM data, which was effectively reduced through the proposed preprocessing. The method identified 264 walking area grids and 324 intersection grids, achieving a precision of 92.2%, a recall of 67.8%, and an F1-score of 0.781 for intersection detection. These results outperform baseline methods, including MAMC (precision 63.4%, F1=0.705) and KDE (precision 46.8%, F1=0.574). Qualitative analysis demonstrates that the proposed method effectively distinguishes true intersections from curved road segments and maintains robustness in areas with uneven trajectory density. The reconstructed 3D network captures elevation variations and slope continuity, avoiding erroneous connections prevalent in 2D-then-3D approaches. The average positional deviation of the extracted network is 4.4 m. Visualization overlays with remote sensing imagery and commercial maps confirm the topological correctness and fine-grained detail of the results.
[Conclusion] This paper presents a hierarchical framework for constructing 3D pedestrian navigation networks from crowdsourced trajectory data. By synergizing direction-aware clustering with DEM-assisted elevation correction, the method overcomes key limitations of existing approaches, achieving high precision in intersection detection, accurate slope representation, and reliable topological connectivity in complex outdoor environments. The output network directly enables intelligent outdoor navigation applications, such as personalized route planning that considers elevation gain or accessibility requirements. Future work will focus on integrating multi-source data (e.g., remote sensing imagery, social media check-ins) to improve coverage in trajectory-sparse regions and enrich semantic attributes such as surface type and path width.
[Objective] To cater to diverse and dynamic travel demands, it is imperative to develop hierarchical and networked public transport systems, for which the hierarchical configuration of transit stations constitutes a critical prerequisite. Existing studies on multi-level station configuration typically proceed from the assumed functions of stations within the transit network and identify hierarchies based on predefined classification systems, thereby constraining methodological flexibility across varying urban contexts. Inspired by the capacity of multi-scale subgraphs to characterize network hierarchy from a structural perspective, this study proposes a transit station configuration method grounded in multi-scale subgraph generation.
[Method] The proposed method is implemented on a transit-supporting road network derived from the urban road network. Leveraging a percolation model, it simulates the continuous evolutionary process of the transit network, progressing from local connectivity to global integration. During this evolution, subgraphs generated at each scale serve as the fundamental units for station configuration. The specific transit station within each subgraph is identified by selecting the node exhibiting the highest travel demand intensity, which is estimated based on the total building floor area within the node’s vicinity. The configured station sets at different scales are subsequently organized from bottom to top in alignment with the percolation process, thereby forming a multi-scale transit station system. This framework enables station hierarchies to emerge from the structural evolution of the network, rather than being imposed by predefined hierarchical labels.
[Result] The case study conducted in Shenzhen demonstrates that, compared with degree-centrality ranking, the recursive minimum cut method, and the Leiden method, the proposed method achieves a more balanced spatial distribution of stations across scales and delivers superior service performance under identical station quantity constraints. The resulting station sets also exhibit clearer hierarchical characteristics in spatial organization. Specifically, the proposed method effectively distinguishes coverage-oriented stations at finer scales from structurally significant stations at coarser scales, thereby reflecting the hierarchical differentiation of transit stations in a more consistent manner. These findings indicate that the proposed method not only enhances the performance of station configuration but also more effectively reveals the hierarchical organization inherent in the transit system.
[Conclusion] This study proposes a transit station configuration method based on multi-scale subgraph generation. By integrating percolation-based network evolution with demand-oriented station selection, the method offers a flexible alternative to conventional approaches that rely on predefined station hierarchies. The results confirm that it can generate a multi-scale transit station system characterized by more balanced spatial organization, improved service performance, and a clearer hierarchical structure. Consequently, this study provides a scientific basis for the optimal configuration of multi-level transit stations.
[Objective] The extraction of cadastral parcel boundaries serves as a cornerstone for land administration, spatial planning, and urban governance. Conventional approaches, however, rely heavily on manual delineation or rule-based algorithms; these methods are not only labor-intensive and prone to human error but also struggle to adapt to the complexities of real-world geospatial data, such as road network discontinuities and multi-line representations. To address these limitations, this study proposes an automated and precise boundary extraction framework that synergizes large language model (LLM) agents with dynamic topological optimization. The core objective is to establish an end-to-end intelligent workflow that translates natural language user intent directly into accurate, topologically closed four-boundaries, thereby significantly enhancing automation and adaptability.
[Method] The methodological framework comprises three core components. First, a multi-level semantic parsing module, powered by LLM, is developed to autonomously identify geographic entities (e.g., road names) and their spatial relationships (e.g., “bounded by”) from unstructured user queries. Second, to bridge the semantic gap to structured databases, a specialized Text-to-SQL converter equipped with a bidirectional verification mechanism is implemented. This component not only generates executable spatial SQL queries but also validates the retrieved datasets against the original intent, ensuring retrieval reliability. Finally, a novel dynamic topological optimization algorithm is designed to process the retrieved, often imperfect, linear road data. Utilizing a multi-scale iterative buffering strategy coupled with adaptive closed-loop detection, this algorithm automatically repairs data breaks, fuses disjoint segments, and constructs a coherent boundary polygon that adheres to predefined geometric and topological constraints.
[Result] Experimental validation was conducted using a real-world road network dataset from Shanghai, comprising over 3000 samples covering both closed and non-closed topological scenarios. The results demonstrate that the proposed framework achieves substantial improvements in extraction accuracy compared to the standard minimum bounding rectangle (MBR) baseline. The system effectively handles complex cases involving road fractures and dual-line representations, scenarios that typically cause traditional methods to fail or generate significant errors. Crucially, the requirement for manual correction and post-processing is markedly reduced, signifying a significant leap in operational efficiency.
[Conclusion] In conclusion, this research presents a robust and intelligent framework for four-boundary extraction. Through the deep integration of natural language understanding, reliable spatial data retrieval, and adaptive geometric processing, the proposed method effectively addresses the key challenges of automation and data imperfection inherent in traditional approaches. The framework’s capacity for intuitive text-based interaction positions it as a practical and scalable tool for cadastral surveying, land resource management, and smart city planning, offering a clear pathway to reduce human dependency and improving processing consistency within geospatial data production pipelines.
As a fundamental and strategic national resource, geographic information confronts an increasingly pronounced tension between its high precision and sensitivity, and the market-driven circulation and sharing of data elements. Ensuring the secure, compliant, and efficient circulation and utilization of geographic data, while safeguarding national security and data sovereignty, has emerged as a critical issue requiring urgent resolution. This paper reviews recent advancements in geographic information security technologies, focusing on core challenges pertaining to copyright attribution and traceability, security control, public application, trusted transactions, and privacy protection.
Digital watermarking serves as a pivotal mechanism for copyright protection and traceability. Lossy watermarking enhances robustness against complex attacks through embedding techniques in the spatial or transform domain, making it suitable for practical scenarios such as infringement traceability. Lossless watermarking encompasses reversible watermarking, feature-based lossless watermarking, and zero watermarking; each approach enables the construction and authentication of copyright identifiers without compromising data accuracy. Currently, digital watermarking is evolving from a static authentication tool into a dynamic rights confirmation mechanism. Security control technology facilitates the proactive protection of geographic information data. Traditional models, including discretionary access control (DAC), mandatory access control (MAC), and role-based access control (RBAC), have been implemented for 2D data. However, these models are characterized by coarse granularity and insufficient dynamic adaptability. Recent research has shifted toward fine-grained and dynamic control strategies. For 3D reality models, security control technologies are progressively extending to the object and attribute levels, thereby enabling local encryption and integrity verification. Nevertheless, balancing control granularity against system performance remains a persistent challenge.
Confidentiality processing technology plays a crucial role in the public application of geographic information data. For 2D data, techniques such as radial basis functions and differential privacy have been introduced to achieve reduced accuracy while enhancing security. For 3D data, approaches combining nonlinear transformations, deep learning-based object detection, and generative adversarial networks have been proposed to address challenges including elevation accuracy reduction, feature height processing, and texture desensitization. These technologies strike a balance between data usability and security, enabling geographic information data to enter circulation and transaction channels lawfully, in accordance with national classification and grading standards.
Blockchain technology addresses the issues of trust deficits and unclear ownership in geographic data circulation. Existing research focuses on establishing mapping mechanisms between on-chain indexes and off-chain data entities. By integrating digital watermarking technology, collaborative ownership verification is achieved within transaction scenarios. The introduction of smart contracts generates an immutable transaction log for activities such as data upload, authorization, trading, and re-authorization, significantly enhancing the transparency and security of data transactions.
Privacy computing technology demonstrates broad application prospects in scenarios requiring unobservable yet usable geographic information data. Methods such as differential privacy, homomorphic encryption, multi-party secure computation, and inner product encryption have been widely applied in areas including geographic data desensitization, ubiquitous spatial positioning, and trajectory data matching. These technologies provide effective solutions for resolving the conflict between geographic information confidentiality and circulation.
In terms of technological engineering and application, digital watermarking and access control technologies have been fully deployed within national and provincial units of the natural resources system. Relevant software has been implemented across multiple provinces and cities nationwide, facilitating the issuance of policy documents at the provincial level. Technologies such as 3D model desensitization, blockchain, and privacy-preserving computation are also being progressively explored in localized scenarios, reflecting an overall trend of orderly advancement.
Finally, this paper delineates three major developmental trends in geographic information security technologies: the synergy and integration of security technologies, technological innovation driven by artificial intelligence, and the construction of a trusted geographic information data space oriented toward data elements. This paper aims to provide theoretical references and technical support for the continuous innovation and high-quality development of geographic information security technologies.
Vector maps are distinguished by high production costs, stringent accuracy requirements, and substantial application value. Serving as indispensable tools in sectors such as autonomous driving, urban planning, and watershed management, they are regarded as valuable national strategic resources. As digital assets, vector maps are easily copied, modified, and distributed. While this facilitates data sharing and broad application, it simultaneously renders them susceptible to piracy, tampering, and unauthorized use, resulting in frequent incidents of copyright infringement and data misuse.
Vector map data often encompass a wealth of sensitive, or even classified, geographic information. Detailed representations of critical infrastructure, military installations, and other strategically significant sites must be rigorously controlled to mitigate potential security threats. Unauthorized or improper use of such data not only impedes the healthy development of the geospatial information industry but also bears direct implications for national defense and strategic security. Consequently, as informatization and networking continue to advance, securing geospatial data—including vector maps—within an open and shared environment has emerged as a critical issue demanding urgent attention. Digital watermarking technology offers a promising solution by embedding user copyright information into vector map content in an imperceptible manner, thereby creating an inseparable entity. The embedded watermark information can be extracted when necessary to verify data ownership and trace the source of leaks, providing a robust means of securing vector maps. This technology must address unique challenges posed by vector map data, including various geometric transformations, coordinate system conversions, and map projection changes that could potentially degrade or destroy embedded watermarks.
This study systematically reviews recent advancements in digital watermarking technologies for vector maps. These include robust watermarking technologies for copyright identification and authentication, fragile watermarking approaches for tamper localization and integrity verification, and zero-watermarking frameworks for high-fidelity applications and trusted transactions.
Robust watermarking technology is designed to resist various attacks and data manipulations while preserving the integrity of embedded watermarks. These methods commonly employ frequency-domain transformation, quantization index modulation, or feature-based embedding to ensure resilience against common operations and attacks. Fragile watermarking technology serves as a key technique for verifying data integrity. It detects unauthorized modifications, localizes tampered regions, and assesses the severity of alterations. By generating feature-based watermarks derived from map content and embedding them into the data structure, the method enables the effective identification and evaluation of changes during verification. Zero-watermarking technology is particularly suitable for high-fidelity applications and trusted transactions that require the preservation of original data integrity. By constructing watermarks from inherent data characteristics without altering the host data, this technology achieves optimal compatibility with fields demanding absolute data fidelity.
Finally, this study outlines prospects regarding future development directions and key unresolved issues for vector map digital watermarking in the era of artificial intelligence, aiming to provide valuable references and insights for related research and practical applications. In the era of big data and artificial intelligence, data products such as “small data” vector maps and generative AI synthetic maps are being widely utilized. Traditional digital watermarking technologies for vector maps face challenges in being directly applied to their security protection. Meanwhile, as geographic information has become a new type of production factor, it is imperative to integrate multiple data security technologies to establish a comprehensive technical framework for security assurance, covering the entire data lifecycle from production to distribution and application.
[Objectives] Vector maps play a crucial role in various fields, including military, economic, and transportation applications. Their information security is directly related to national security and the development of the geographic information industry. However, existing encryption methods for vector maps have not sufficiently considered effective integration with compression processes. This means that the targets for encryption and transmission remain relatively bulky raw data without compression. In resource-constrained scenarios where computational resources and network bandwidth are limited—such as autonomous driving, mobile navigation, and IoT terminals—a critical challenge is to ensure encryption security while improving efficiency and reducing computational and network transmission overhead.
[Methods] We propose a novel efficient encryption method with compression that takes into account the characteristics of vector map data structures, combining vector map compression and encryption. Firstly, we introduce a slope-based Douglas-Peuker (SDP) algorithm to categorize coordinates into key points and non-critical points. Grid-based mapping methods are applied to key points, while delta algorithms are used for differential processing of non-critical points. Secondly, we present a directional run-length encoding (DRLE) algorithm to compress integer sequences while securing their information. Additionally, a dynamic key generation scheme using SHA-512 and a unified chaotic system is proposed, and coordinate data is encrypted using XOR and scrambling algorithms, which are secure and efficient. Finally, a new Lossless Compression Structure is introduced to store the processed coordinate sequences, and LZMA2 (Lempel-Ziv-Markov chain-algorithm 2) is employed to compress them, thereby obtaining an encrypted map with lossless compression. In this study, publicly available data from the open-source geographic database OpenStreetMap (OSM) was utilized for experimentation. Vector map data in the shapefile format, varying in size, region, and object types, were adopted as the experimental datasets.
[Results] Experimental results demonstrate that the proposed method outperforms existing approaches. In terms of encryption, the information entropy of the proposed method is higher than that of the methods proposed by Ren et al. (2020), Jang et al. (2014), and Wang et al. (2021). Furthermore, the encryption time of the proposed method is only about [insert ratio here] times that of the methods by Ren et al. (2020), Jang et al. (2014), and Wang et al. (2021), respectively. Regarding compression performance, compared with classical compression algorithms such as 7 z, zip, and rar, the proposed method achieves an improvement in compression ratio of approximately 56.68%, 124.81%, and 75.77%, respectively. On the same dataset, the proposed method not only provides higher security but also achieves a higher compression ratio and faster encryption speed.
[Conclusions] The proposed method ensures security while significantly reducing computational and communication overhead, offering an effective solution for secure vector map data transmission in resource-constrained environments. In the future, the focus will be on enhancing the versatility of the method and further optimizing its performance on embedded platforms.
[Objective] A persistent challenge in digital watermarking is the inherent trade-off between imperceptibility and robustness. Conventional strategies typically rely on random or uniform embedding locations combined with fixed embedding strengths, which constrains their adaptability to the diverse content characteristics of different images. This paper aims to develop an optimization model that simultaneously and adaptively determines both optimal embedding locations and optimal embedding strength, thereby achieving a superior equilibrium and enhancing the overall performance of various watermarking algorithms applied to both remote sensing and natural digital images.
[Method] This paper proposes a dual-adaptive watermarking optimization model that integrates the gray-level co-occurrence matrix (GLCM) with particle swarm optimization (PSO). The methodology comprises two key steps. First, the host image is partitioned into 8×8 sub-blocks. The GLCM is then utilized to analyze the texture complexity of each sub-block by computing four critical statistical features: contrast, energy, entropy, and correlation. The sub-block exhibiting the highest complexity is subsequently selected as the optimal region for watermark embedding, thereby improving resilience against attacks such as cropping. Second, the PSO algorithm is employed to dynamically determine the optimal embedding strength. A pivotal aspect of this optimization is the formulation of a fitness function that synthesizes the peak signal-to-noise ratio (PSNR)—which quantifies imperceptibility—and the average normalized correlation (NC) coefficient extracted following the simulation of multiple common attacks (e.g., rotation, scaling, JPEG compression, salt & pepper noise). This function guides the PSO swarm to identify the strength value that optimally balances visual quality and robustness. The PSO parameters are configured with a population size of 20 and a maximum of 100 iterations to ensure efficient convergence.
[Result] Extensive experiments were conducted on four remote sensing images and four standard digital images, applying the proposed model to four distinct watermarking algorithms from the literature. The results demonstrate significant performance enhancements. Regarding imperceptibility, the optimized algorithms exhibited a consistent increase in PSNR values across all test images compared to their original implementations, confirming improved visual quality. More importantly, in terms of robustness, the NC values for watermark extraction were markedly elevated after optimization under a broad spectrum of attacks. These attacks encompassed geometric transformations (rotation up to 10°, scaling by 0.5× and 2×, and translation), various cropping patterns, filtering operations (Gaussian low-pass, median, and Wiener), noise additions (salt & pepper, speckle, and Gaussian), and JPEG compression. For instance, under severe cropping or high-degree rotation, the extracted watermarks were significantly clearer and more intact following optimization. The model demonstrated excellent adaptability and efficacy across different image types and underlying watermarking techniques.
[Conclusion] The proposed dual-adaptive model effectively addresses the core trade-off in digital watermarking by intelligently selecting both the embedding location based on texture complexity (via GLCM) and the embedding strength based on a holistic performance metric (via PSO). Experimental evidence validates that the model can be generically applied to enhance existing watermarking algorithms, endowing them with simultaneously heightened robustness against diverse disturbances and superior imperceptibility. This work provides a practical and effective optimization framework for securing multimedia data, including specialized imagery such as remote sensing data, where copyright protection and content authentication are of paramount importance.
[Objective] The imperceptible embedding of digital watermarks into vector maps necessitates rigorous evaluation to ensure the utility and integrity of geographic data are preserved. However, existing fidelity evaluation frameworks for watermarked vector maps are often constrained by two critical limitations: a reliance on a narrow scope of evaluation metrics and a fundamental disregard for the uncertainties inherent in both the watermarking process and the evaluation parameters. To address these issues, this paper proposes a novel soft evaluation method that integrates a comprehensive assessment model with uncertainty analysis, thereby achieving a more robust and realistic fidelity evaluation.
[Method] The proposed methodology is structured into two principal phases. First, we construct a multi-dimensional fidelity evaluation index system to transcend single-metric assessments. This system comprehensively encompasses three critical dimensions of vector data quality: topological consistency (e.g., the preservation of spatial relationships), geometric feature quality (e.g., shape distortion and vertex displacement), and data error (e.g., alterations in coordinate precision). To objectively synthesize these diverse indicators, we employ the fuzzy analytic hierarchy process (FAHP) to determine their respective weights, thereby effectively addressing the ambiguities associated with subjective judgment. Crucially, we incorporate uncertainty quantification at this stage by utilizing random distributions and triangular fuzzy numbers to characterize the imprecision in key evaluation parameters. Second, we introduce a Monte Carlo stochastic simulation to propagate these uncertainties throughout the entire evaluation model. This technique conducts large-sample statistical computations, simulating numerous potential evaluation scenarios across varying confidence levels. This process enables the fitting of a probability distribution for the overall fidelity index, rather than yielding a single, potentially misleading deterministic value. Consequently, this approach provides a robust mechanism for analyzing uncertainty propagation and visually representing the potential range of fidelity outcomes.
[Result] Experimental results derived from multiple sets of typical vector map data (e.g., large-scale cadastral maps and small-scale topographic maps) and diverse watermarking algorithms (e.g., spatial domain and transform domain methods) validate the efficacy of our model. The results confirm that our method effectively characterizes the random disturbances and fuzziness induced by watermark embedding. The fidelity assessment outcomes demonstrate superior stability and interpretability, as they comprehensively account for the full spectrum of potential impacts.
[Conclusion] The proposed method shows strong adaptability and flexibility across data of varying scales and application contexts, providing a universal and reliable framework for assessing the true fidelity of watermarked vector geographic data. This work bridges a significant gap in digital watermarking evaluation and provides a principled approach for quality assurance in geospatial data security.
