[Objectives] Geographic Knowledge Graph (GeoKG) employs knowledge graph techniques to represent geographic knowledge as a computer-interpretable, reusable, and inferable knowledge network. However, due to the sparsity of geographic information distribution and outdated updates, GeoKGs are often incomplete, which restricts their breadth and depth of application. Geographic knowledge graph completion techniques are needed to address this incompleteness. Nevertheless, existing knowledge graph completion methods fail to fully account for the semantic information within GeoKGs and the distance-decaying effect governing interactions among geographic entities, resulting in an embedding space that inadequately captures the true distribution of geographic entities and relations, thereby limiting completion performance. [Methods] To address this issue, this study proposes a Distance-Decaying Effect-Aware Geographic Knowledge Graph Completion method (DDGKGC). The method first captures semantic information and distance-related features between entities and relations through a semantic information aggregation module and a distance-decaying effect-aware module. Then, a dual-attention mechanism-based representation learning module adaptively learns neighborhood information of entities and relations to derive their embeddings. Finally, the ConvE scoring function is used for prediction, and the results are applied to complete the GeoKGs. [Results] To comprehensively evaluate model performance,this study conducts comparative experiments, ablation studies, and multi-dimensional validation analyses on the self-constructed datasets Multi-Geo, CityDirection, and CountyDistance, as well as the public dataset Countries-S3. Experimental results demonstrate that DDGKGC achieves outstanding performance across multiple metrics including MRR, Hits@1, Hits@3, and Hits@10. Particularly in terms of MRR, which comprehensively reflects model performance, DDGKGC outperforms the baseline methods by 4%, 3.1%, 1.8%, and 5.2% on the four datasets, respectively. Moreover, through multi-dimensional validation and analysis, it is proven that DDGKGC can more effectively model the spatial and semantic relationships among geographic entities, thereby enhancing the accuracy and geographic plausibility of completion results. [Conclusions] The results demonstrate that the proposed method not only effectively enhances the performance of the geographic knowledge graph completion task but also exhibits strong generalization capability and application potential. Furthermore, it provides reliable support for the advanced application of GeoKGs.

[Objectives] Traditionalmaps are characterized by a typical discrete nature, abstractly depicting geospatial information through point, line, and area symbols. However, this discrete expression mode has inherent limitations in capturing dynamic geographical processes and conveying multi-dimensional spatial information, which can no longer fully meet the growing demand for intelligent and dynamic geospatial visualization. With the rapid advancement of information technologies, the emergence of Pan-maps has opened up new avenues to overcome the discretization constraints of traditional maps and achieve continuous representation of geospatial information. From the perspective of continuous representation, this study aims to explore and reveal the potential association rules and intrinsic continuity mechanisms among various Pan-map types, and further construct a comprehensive dimensional model for the continuous expression of Pan-maps, thereby providing theoretical and methodological support for the innovation of geospatial visualization paradigms. [Analysis] To achieve the aforementioned research objectives, a systematic analytical framework was established in this study. Firstly, a scientific and hierarchical classification system for Pan-maps was constructed based on the differences in data sources, expression purposes, and spatial feature types of Pan-maps. On this basis, similarity calculation methods were employed to deeply explore the continuous variation rules of map symbols across multiple dimensions, including geometric shape, color attributes, and spatial relationships. Secondly, aiming at the continuity transformation between different map types, the FP-Growth (Frequent Pattern Growth) algorithmwas adopted to excavate the implicit continuous transformation rules among diverse Pan-map elements. Based on the mined rules, a multi-dimensional model for the continuous expression of Pan-maps was constructed, covering five core dimensions: map space, map base, spatial location, map symbols, and spatial relationships. This model integrates the synergistic changes of various elements to realize the unified description of the continuous transformation process between different Pan-map types. Finally, to verify the effectiveness and applicability of the constructed dimensional model, a series of continuous transformation experiments were conducted targeting point-type, line-type, and area-type maps, respectively. [Conclusions] This study has significant theoretical and practical implications. The constructed Pan-map continuous expression dimensional model helps break the discrete expression mode of traditional maps and establish a continuous expression thinking for Pan-maps. It enables multi-perspective continuous presentation of diversified geospatial information, enhancing map information transmission efficiency and users' perception of dynamic geographical processes. Additionally, the findings provide a new technical path for transforming geospatial visualization from static discretization to dynamic continuity, laying a foundation for intelligent and personalized mapping, and enriching the Pan-map theoretical system to support subsequent related research.

[Objectives] With the rapid development of Volunteered Geographic Information (VGI), high-timeliness crowdsourced road networks have become an important data source for smart city applications, where the efficiency and rationality of network selection are critical to multi-scale data services. Existing methods rely heavily on attribute information such as road class, name, or traffic flow, which often suffer from incompleteness and inconsistency, limiting their applicability. To address this issue, this study proposes an automatic modeling and selection method for urban road networks under attribute-deficient conditions, based on spatial syntax analysis. [Methods] Using OpenStreetMap (OSM) centerline data, the method adopts a programmatic workflow to automatically perform geometric simplification, topological correction, and pseudo-node removal to generate segment maps and compute syntactic indicators such as integration and choice. Strokes are then constructed to extract geometric features, and two composite indicators—Stroke-based Normalized Angular Integration (SNAIN) and Stroke-based Normalized Angular Choice (SNACH)—are introduced to jointly characterize topological accessibility and geometric continuity. A combined weighting strategy integrating the Entropy Weight method and the Analytic Hierarchy Process (EW-AHP) determines road importance, while a dead-end identification and grid-density repair mechanism improves network connectivity. [Results] Experiments in Lanzhou ( linear network ) and Chengdu (ring-radial network) show that under attribute-deficient conditions, the proposed method still effectively extracts main urban roads, achieving OSM-based accuracies of 0.942 1 and 0.971 1, respectively. Connectivity improved from 1.058 2 to 1.086 4 in Lanzhou and from 1.108 6 to 1.119 8 in Chengdu, where dead ends were fully eliminated. Ablation analysis further confirms that SNAIN enhances global connectivity, while SNACH strengthens geometric continuity. [Conclusions] Overall, the proposed method provides new theoretical support and a practical technical pathway for large-scale urban road network generalization in situations where attribute information is incomplete or unreliable.

[Objectives] To address the issue of the missing hierarchical structure of overpasses in OpenStreetMap (OSM) that hinders the development of high-precision maps and intelligent navigation, this paper aims to break through the strong reliance of traditional methods on elevation or LiDAR point cloud data and proposes a method for automatically identifying the hierarchical structure of overpasses that only fuses remote sensing images and vehicle trajectory data without the need for elevation information. [Methods] A framework for identifying the hierarchical structure of overpasses in OSM road networks by fusing remote sensing images and vehicle trajectory data is proposed. Firstly, based on the spatial topological relationship between remote sensing images and OSM road networks, the overlapping areas of roads are detected; linear features are extracted through the Hough transform and combined with a slope comparison strategy to initially determine the spatial relationship of upper and lower layers of overlapping roads. Secondly, a Gaussian mixture model is constructed using vehicle trajectory data to extract speed distribution features, and a random forest classifier is used to accurately identify parallel overlapping roads. Finally, a local-global reasoning algorithm is introduced to assign hierarchical attributes to OSM road network nodes and edges based on spatial geometric constraints and trajectory behavior patterns, and to achieve structural visualization output. [Results] The experiments were carried out in multiple typical overpass areas in Beijing. The results show that the accuracy rate of this method in the task of discriminating overlapping road levels is 99%, the recall rate is 89%, and the F1 score is 94%; in the task of identifying overlapping roads, the accuracy rate is 100%, the recall rate is 86.96%, and the F1 score is 93.02%. Compared with the existing methods that rely on airborne LiDAR or GPS trajectory elevation, this method does not use elevation information at all, significantly reducing the cost and threshold of data acquisition, and has higher overall recognition accuracy, demonstrating stronger practicality and scalability. [Conclusions] The multi-source data fusion framework proposed in this study effectively realizes the fine-grained identification of the hierarchical structure of OSM overpasses, breaks the dependence on elevation data, and provides a reliable technical path for improving the quality of open-source map data. It can be widely applied in intelligent navigation, high-precision map construction for autonomous driving, and urban traffic modeling.

[Objectives] This study aims to address the challenge of determining the optimal hyperparameter combinations in landslide susceptibility evaluation models. [Methods] This paper introduces a novel Spider Wasp Optimizer (SWO) algorithm designed to identify optimal hyperparameter combinations for machine learning models. Utilizing the SWO, hyperparameter optimization was performed on Random Forest (RF), Light Gradient Boosting Machine (LightGBM), and Categorical Boosting (CatBoost) models, resulting in the identification of optimal hyperparameter values for these models, which were subsequently used to construct a landslide susceptibility assessment model. Building on this foundation, the machine learning models optimized by SWO were integrated using a stacking method. By comparing the evaluation results of each model, the optimal landslide susceptibility model was selected. Furthermore, the SHAP (SHapley Additive exPlanations) algorithm was employed to conduct an interpretability analysis of the evaluation results for the optimal model. [Results] This paper uses the slopes along the Yaxue Highway in Heilongjiang Province as a case study and employs the SWO optimization algorithm to optimize the hyperparameter combinations of the aforementioned machine learning models. The results demonstrate that SWO-LightGBM, SWO-CatBoost, and SWO-RF improve the AUC (Area Under the Curve) values of the models by 2.4%, 1.6%, and 2.2%, respectively, compared to the models optimized prior to the application of SWO. This indicates that the SWO algorithm effectively enhances the overall performance of the machine learning models, particularly in predicting landslide susceptibility. Notably, the SWO-LightGBM model exhibits the best performance, achieving an AUC value of 0.939. The evaluation results for the four stacking models reveal AUC values ranging from 0.924 to 0.935, all of which are lower than those of the SWO-LightGBM model. Finally, an interpretability analysis of the SWO-LightGBM model indicates that slope, distance to road, average annual rainfall, and distance to river significantly contribute to landslide susceptibility. [Conclusions] This study employs the SWO Algorithm to identify the optimal combination of hyperparameters, thereby significantly enhancing the model's predictive accuracy and precision of results.

[Objectives] The thermal environment is a critical factor affecting urban residents' lives, and its spatiotemporal structure holds significant implications for optimizing green space planning, building layout, and promoting public health. Existing studies predominantly rely directly on Land Surface Temperature (LST) data to identify thermal spatial aggregation patterns, overlooking the influence of underlying heat sources and sinks on spatial structures. This results in insufficiently detailed revelations of thermal environmental spatial patterns. [Methods] This paper proposes a method for analyzing the local cold and hot spot structure of the urban thermal environment by integrating spatial blind source separation algorithms in response to the complex cold and hot patterns formed by temperature diffusion and superposition at different spatial locations. Based on the daily LST data of Xi'an in 2021, the potential cold and hot source components are separated. Combined with the Getis-Ord Gi* analysis and the calculation of spatial transfer rates, the spatio-temporal distribution and evolution characteristics of cold and hot spots are identified and quantified. [Results] ① The introduction of the Bayesian spatial blind source separation algorithm enables the comprehensive deconstruction of urban thermal environments based on multi-temporal LST data, isolating latent heat source components with distinct spatio-temporal characteristics. This provides a critical data foundation for subsequent identification of hot and cold spot structures. ②The distribution of hot and cold spots in Xi'an's surface temperature exhibits significant spatial heterogeneity and seasonal dynamics: the southeast and northwest sections of the Ring Expressway form persistent hotspot clusters, while cold spots are scattered across the northern, central, and southwestern suburban areas. During winter, cold spots expand in coverage and exhibit active migration; in summer, the proportion of hotspots rises markedly (reaching 53.2% by July). During the transitional periods of spring and autumn, the alternation between cold and hot spots is frequent, as exemplified by a spatial transfer rate as high as 73.2% in May. ③Compared with the traditional Gi^* hotspot analysis which can only identify significant high/low temperature aggregation areas, this study, by integrating spatial blind source separation and hotspot analysis, can parse the spatial distribution of each independent component, revealing the internal component evolution and heterogeneous characteristics of the thermal environment. The comparison shows that the integrated method can construct a more comprehensive and rich spatial structure of the thermal environment than the traditional method.[Conclusions] This research helps to reveal the more detailed hot and cold spot spatial structure within cities. It not only provides a new method for analyzing the spatiotemporal dynamics of urban thermal environments but also offers a scientific basis for ecological protection, green space planning, and precise prevention and control of heat risks.

[Objectives] Cross-industry co-agglomeration is a typical spatial organization pattern in the marine economy. However, due to a lack of rigorous measurement indicators, directional relationships—such as dependency, dominance, or dissociation—between industries have not been empirically verified at the micro level. This study aims to introduce and develop machine learning techniques to extract spatiotemporal big data related to marine enterprises, thereby overcoming methodological bottlenecks in measuring co-agglomeration relationships. [Methods] First, four models relying on different machine learning frameworks such as artificial neural networks are introduced. An integrated ensemble learning algorithm incorporating Natural Language Processing (NLP) was proposed for text classification, enabling the construction of a classifier to identify the industry categories of marine-related enterprises. Second, using the industry labels and geographic coordinates of marine-related enterprises, an improved Earth Mover's Distance (EMD) was adopted. This enhancement incorporated an entropy-regularization constraint and the Sinkhorn fixed-point iteration algorithm, enabling the measurement of directed Wasserstein distances between two industry distributions. The Monte Carlo simulation method was then used to generate counterfactual samples, repeated 1 000 times to compute the EMD. By comparing actual results with counterfactual simulations, a statistically significant vector-based industrial co-agglomeration index was constructed. [Results] Five Chinese coastal cities—Dalian, Qingdao, Ningbo, Xiamen, and Guangzhou—were selected as case study areas. Business registration information for marine-related enterprises was retrieved from enterprise information platforms, and coordinates were obtained via map service platforms, enabling the construction of a dynamic geographic database of marine industries. The performance of the four NLP models was evaluated using accuracy and F₁-score. The validity of the co-agglomeration index was verified through convergence speed and the distribution of counterfactual samples from Monte Carlo simulations. Spatial analyses were also conducted using the measured EMD and co-agglomeration index, demonstrating the method's practical analytical value. The classifiers achieved accuracy rates of 84.8%, 84.7%, 92.1%, and 92.2%, respectively. The ensemble soft-voting approach further improved classification reliability. Monte Carlo simulations generally converged after 200 iterations, and no more than 400 iterations were required in specific cases. Counterfactual samples indicating significant co-agglomeration closely approximated the overall distribution. The measured co-agglomeration index was statistically significant, and the indicators demonstrated strong economic relevance for marine industry research. [Conclusions] The integration of artificial neural networks, a machine learning-optimized EMD, and Monte Carlo simulation forms a comprehensive methodological framework for analyzing marine industry co-agglomeration. This approach provides robust technical support for investigating industry co-location based on micro-level data.

[Objectives] Holidays are pivotal periods for the prosperity and development of the regional tourism industry. Understanding the behavioral characteristics and patterns of non-local tourists during these times is crucial for enhancing the quality of regional tourism. However, limited by the availability of sensing data, existing research lacks analysis of non-local tourists' holiday behaviors based on individual-level data, as well as a targeted analytical framework and methodology. [Methods] To address this gap, this study proposes a data-driven analytical framework—"Identity Recognition-Behavior Analysis-Pattern Mining"—for analyzing the behaviors of non-local tourists during holidays, utilizing large-scale anonymized mobile location data. Based on mobile location data from Xiamen, a typical tourist city, during the 2023 Labor Day and Dragon Boat Festival holidays, this study analyzes the behavioral characteristics and patterns of non-local tourists. [Results] The results reveal that: (1) Significant differences exist in the temporal, spatial, behavioral intensity, and diversity patterns of non-local tourists in Xiamen between the Labor Day and Dragon Boat Festival holidays. During Labor Day, tourists tended to adopt a "breadth exploration" strategy, whereas during the Dragon Boat Festival, they preferred a "depth immersion" strategy. (2) Tourist behaviors exhibited clear sociodemographic heterogeneity: during Labor Day, female tourists spent 10.7% more time on activities than male tourists, and young tourists engaged 33.8% longer than older tourists. Conversely, during the Dragon Boat Festival, older tourists demonstrated 34.7% higher travel efficiency than younger tourists. (3) Cluster analysis identified six behavioral patterns:Roaming Experience, Relaxed Sightseeing, All-area Deep Travel, Systematic Touring, Fixed-point Deep Cultivation, and Efficiency-oriented Exploration. The distribution of these patterns exhibited notable holiday-specific tendencies: All-area Deep Travel, Systematic Touring, and Efficiency-oriented Exploration were predominantly observed during Labor Day, while Fixed-point Deep Cultivation was more frequent during the Dragon Boat Festival.This reflects the diversity of travel itineraries and the mechanism of holiday-based behavioral preferences. [Conclusions] This study extends the methodology for analyzing and identifying tourism behavior patterns of non-local tourists based on individual location data. The proposed methods and analytical conclusions enhance the understanding of tourist behaviors during holidays, providing a scientific basis and practical reference for improving tourism services in cities like Xiamen.

[Objectives] In response to the problem of difficulty in accurately docking patch data from natural resource monitoring with business management requirements, this paper aims to propose a method for effectively linking monitoring patches with business management, in order to achieve the goal of "one-time monitoring", to fulfill "multiple natural resource management requirements" and enhance the efficiency and precision of natural resource monitoring and governance. [Methods] Starting from the cognition and expression of the conceptual system of natural resources, the categorization of business rules, and the construction of business knowledge graphs, a set of business-related methods for natural resource monitoring patches oriented towards management requirements is proposed based on the business knowledge graph. A monitoring patch automatic push model based on business association rules was designed to achieve systematic correlation and accurate docking of monitoring patch information with business management requirements. [Results] This paper presents a study on the business association method for natural resource monitoring patches based on knowledge graphs. The research results indicate that: ① The cognition method of the conceptual system of natural resource monitoring objects, oriented toward management needs, systematically sorts out the implicit logic between complex business rules and numerous monitoring objects, thereby providing a theoretical foundation for the classification management and monitoring applications of natural resources; ② The construction of the business knowledge graph for natural resource monitoring facilitates the multi-dimensional dynamic correlation among monitoring patch attributes, business rules, and management requirements. This is achieved through visual mapping and interactive analysis, which subsequently provides graph-based decision support for the refined governance of natural resources; ③ Through an automatic monitoring patch push model based on the business knowledge graph, a mapping mechanism between patch knowledge and management requirements has been established, and a hierarchical and categorized intelligent push system has been constructed. This integrated approach successfully attains the core objective of leveraging a single monitoring event ("1") to service and underpin a multitude ("N") of distinct natural resource management business functions, thereby drastically improving the timeliness, targeting accuracy, and overall utility of monitoring outputs. [Conclusions] The proposed methodology, which is demand-oriented for management and links monitoring patches with business operations, in conjunction with the knowledge graph model, effectively addresses the identified shortcomings of the synergy between natural resource monitoring data and business requirements. Specifically, it mitigates issues related to an underdeveloped collaborative mechanism and inefficient, imprecise data dissemination. This research provides robust technical support for advancing the refined supervision of natural resources and significantly contributes to the intelligent management and control of territorial space.

[Objectives] This paper tackles the challenges of small target detection, complex background handling, and dense target distribution in remote sensing imagery, proposing an advanced solution to enhance detection performance. The research aims to improve the accuracy and robustness of target detection in high-resolution remote sensing images by addressing limitations in existing methods, particularly in scenarios with intricate backgrounds and densely packed or small-sized targets. [Methods] Based on the YOLOv11 framework, this paper proposes an advanced remote sensing object detection method that effectively integrates multi-scale feature collaboration and scenario-aware mechanisms. To achieve superior performance, three novel modules are specifically designed: the Parallel Kernel Feature Fusion Module (PKFFM), which performs cross-scale feature integration through parallel convolution kernels to significantly enhance feature representation capability; the Cascaded Dual-Branch Attention Module (CDBAM), which sequentially emphasizes critical spatial and channel-wise information to refine feature extraction; and the Scenario-Aware Module (SAM), which enables the network to better capture and utilize global contextual information in complex remote sensing scenes. Furthermore, the RS-WIoU (Remote Sensing Wise Intersection over Union) loss function is introduced to address the challenges of high-resolution imagery and varying object scales, leading to more accurate bounding box regression and substantially improved overall detection performance. [Results] To comprehensively validate the effectiveness of the proposed method, extensive experiments are conducted on three widely recognized high-resolution remote sensing datasets: TGRS-HRRSD, NWPU VHR-10, and DOTA-v1.0. The experimental results demonstrate that the proposed approach achieves outstanding mean Precision (mP) of 97.3%, 87.3%, and 84.3% on the respective datasets, significantly surpassing the baseline YOLOv11 model with relative improvements of 2.1%, 3.8%, and 2.9%. In terms of the more comprehensive metric mAP_50-95, the proposed method further delivers gains of 3.0%, 1.2%, and 1.5% across the three datasets. Beyond superior accuracy, the model exhibits remarkable lightweight characteristics and strong robustness against complex backgrounds, varying scales, and dense object distributions, consistently outperforming other state-of-the-art remote sensing object detection algorithms. [Conclusions] The proposed method dramatically improves both precision and robustness in high-resolution remote sensing image object detection by synergistically combining the PKFFM, CDBAM, SAM, and RS-WIoU loss function, delivering a highly efficient and effective solution for real-world remote sensing applications. This collaborative framework enables better multi-scale feature fusion, enhanced channel and spatial attention, adaptive scenario understanding, and more accurate localization, leading to state-of-the-art results. Future work will focus on validating these modules on additional datasets and downstream tasks to further strengthen generalization performance and drive continued innovation in remote sensing technology.

[Objectives] Oriented object detection is a cornerstone of modern remote sensing interpretation, directly enabling the efficient extraction of geographic information, precise identification of critical targets (such as aircraft, vehicles, and ships), and large-scale dynamic environmental monitoring. Despite significant advances, mainstream detectors continue to grapple with persistent challenges, primarily the boundary discontinuity problem inherent in angle regression and their limited capacity for effectively representing objects with highly diverse geometries, ranging from extremely slender structures to near-square shapes. [Methods] In response to these limitations, this paper introduces a novel Frequency-domain. Geometric and Angular Fusion RCNN (FGAF-RCNN) framework. The proposed FGAF-RCNN architecture consists of two dedicated modules for frequency-domain manipulation. The former, the Shape-Adaptive Amplitude Modeling (SA2M) module, enhances shape perception by leveraging frequency-domain amplitude as a robust and indirect representation of an object's morphology. It adaptively encodes fundamental geometric priors, most notably the aspect ratio, into specific amplitude patterns within the spectrum.Furthermore, a Rotation-Adaptive Spectrum Optimization Module (RASOM) is designed to integrate orientation information by mapping the object's angle to the phase component of the spectrum, thereby coupling shape and orientation into a unified frequency representation. In the complex frequency domain, amplitude and phase are intrinsically linked; thus, this operation effectively couples shape and angle into a single, cohesive frequency representation. The synthesized spectrum is subsequently refined through Fourier denoising, effectively resolving the feature boundary discontinuity problem, leading to a more stable regression target. [Results] To validate the efficacy of our approach, we conducted extensive experiments on the large-scale and publicly available DOTA-v1.0 and DIOR-R dataset. The experimental results are compelling: our FGAF-RCNN model achieves a state-of-the-art mean Average Precision (mAP) of 77.4% and 97.2% on the test set,while reducing the mean Average Orientation Error (mAOE) to 7.37. This represents a significant improvement of 2.3% and 0.7% over our strong baseline model and demonstrates competitive performance against other leading methods.To further evaluate the model's generalization capability, validation on the DIOR-R dataset demonstrated that our method achieves 64.3% mAP, showcasing its superior detection performance in complex scenarios. A thorough comparative analysis confirms that our method delivers dual advantages: it not only achieves superior accuracy in angle prediction but also exhibits remarkable robustness in detecting objects of varying shapes. [Conclusions] In conclusion, this work makes a substantive contribution to the field of oriented object detection by pioneering a frequency-domain perspective. The proposed FGAF-RCNN framework, through its synergistic SA2M and RASOM modules, provides a novel and effective solution to the long-standing challenges of shape adaptability and boundary discontinuity.

[Objectives] The fusion of high-resolution optical imagery and Synthetic Aperture Radar (SAR) imagery can effectively improve the accuracy and robustness of building extraction. However, in practical applications of remote sensing image analysis, the performance of multimodal fusion models often degrades significantly due to data missing or incompleteness in one modality during the inference stage. To address this issue, this study aims to develop a lightweight building extraction method that requires only single-modal input while achieving performance comparable to that of dual-modal fusion models. [Methods] We design a cross-modal fusion framework based on online knowledge distillation, which comprises a dual-modal teacher network and a single-modal student network. The core innovations include: (1) An Adaptive Gated Attention Fusion Module (AGAFM) in the teacher network, which dynamically weights and integrates optical and SAR features to achieve complementary fusion, effectively suppressing redundant information while highlighting salient building features; (2) A Pseudo-Feature Generation Module (LDAF/ESAR) in the student network, which simulates the missing modality's characteristics by learning latent distributions from the teacher's fused representations, enabling the student to operate independently with single-modal input; and (3) A multi-level knowledge distillation loss applied at both feature and output layers, which forces the student network to mimic the teacher's fusion capabilities by aligning intermediate features and final predictions. Additionally, to better capture the geometric properties of buildings, we introduce Deformable Convolution Modules (DCM) to adaptively adjust receptive fields for irregular building shapes, and a boundary-aware enhancement module (MAC-BEM) to refine contour details by leveraging gradient-aware constraints. [Results] Experiments were conducted on the Shandong and the Republic of Korea subsets of the DDHRNet_DATA dataset. The results demonstrate that when the SAR modality is missing, the student network achieves IoU scores of 83.68% and 77.24% on the two subsets, surpassing the sub-optimal algorithms by 3.06% and 2.66%, respectively. When the optical modality is missing, the student network achieves IoU scores of 77.78% and 77.20%, outperforming the sub-optimal algorithms by 4.01% and 1.31%, respectively. The proposed method shows significant advantages over single-modal comparison models such as SegNet, Deeplabv3, Deeplabv3+, UNetFormer, MFFDeeplabV3+, and SC_Deep. Ablation studies further verify the effectiveness of each core module. [Conclusions] In conclusion, the proposed method effectively addresses the practical bottleneck of modality missing during the testing phase. It provides a reliable and efficient solution for the real-world deployment of multi-modal remote sensing building extraction technology, offering strong generalization and robustness in scenarios with incomplete data. This work contributes to advancing the practical application of multimodal remote sensing analysis and supports the development of lightweight, high-performance models for geospatial information extraction.

[Objectives] Building information plays a crucial role in applications such as urban planning, environmental monitoring, and disaster emergency response. With the widespread adoption of high-resolution remote sensing imagery, how to achieve efficient and accurate automatic building extraction has become a key research focus in the field of remote sensing. While deep learning approaches have enhanced the efficiency and precision of building extraction from remote sensing images, existing algorithms continue to encounter challenges pertaining to contour integrity, anti-interference capability, and model lightweighting. The aim of this paper is to construct a building extraction network that balances high accuracy and low complexity by fusing the advantages of convolution and self-attention mechanisms, in order to promote the practical application of this method in resource-constrained environments. [Methods] To tackle these issues, this study integrates the U²-Net network with Transformer modules, thereby proposing a novel lightweight global attention network model termed U²-former. This method introduces three improvements based on the U²-Net network: ① the incorporation of a channel attention mechanism in the encoding segment to strengthen the capacity for capturing local features; ② the reconstruction of the decoder into a Transformer module, with global spatial dependencies established via an optimized multi-head attention mechanism and a channel-enhanced multi-layer perceptron; ③ the adoption of a multi-level feature fusion strategy to integrate outputs from different decoding layers, so as to improve boundary integrity. [Results] Experiments on three benchmark datasets, namely WHU aerial images, Massachusetts aerial images, and Inria aerial images, show that the U²-former model has only 6 M parameters, yet its IoU indices still reach 91.69%, 74.96%, and 80.13% respectively, which outperforms those of current popular algorithms; The ablation experiments further demonstrate the effectiveness of the three improvements: relative to U²-Net, the proposed U²-former model improves the IoU values on the three datasets by 1.33%, 3.54%, and 2.24%, respectively. In addition, visualization results confirm that U²-former effectively maintains building contour integrity in complex scenarios, reduces false detections and omissions, and exhibits strong robustness against shadow occlusion and structurally complex buildings. [Conclusions] By combining the local feature extraction capability of CNN with the global contextual modeling ability of Transformer, the U²-former model achieves high-precision building extraction with very low parameter complexity. The proposed method effectively addresses the key trade-offs among model lightness, feature integrity, local detail preservation, global dependency modeling, and anti-interference capability. It not only achieves outstanding performance in building extraction from high-resolution remote sensing images but also offers a promising technical pathway for intelligent interpretation of remote sensing imagery in resource-constrained scenarios such as edge computing.

[Objectives] The Qinghai-Tibet Plateau, the world's highest and China's largest plateau, plays a vital role in regulating regional climate, ecosystem processes, and hydrological cycles. Accurately characterizing the spatial and temporal distribution of its Land Surface Temperature Lapse Rate (LTLR) is therefore essential. However, existing studies often fail to capture the fine-scale spatiotemporal variability of near-surface temperature lapse rates over complex mountainous terrain, limiting their ability to represent detailed patterns and variations under such conditions. To address this gap, this study uses a Diurnal Temperature cCycle (DTC) model to estimate hourly land surface temperature across the Qinghai-Tibet Plateau and calculates hourly mean monthly LTLR to obtain a high-spatiotemporal-resolution LTLR dataset for the region. [Methods] Based on the TRIMS daily 1 km all-weather land surface temperature dataset for western China in 2022, this study adopted a DTC model to estimate hourly temperature across the Plateau. Hourly mean monthly LTLR was then computed using a sliding-window method, and the spatiotemporal characteristics of LTLR at the seasonal scale were systematically analyzed. This work fills an important gap by providing high-spatiotemporal-resolution LTLR estimates for the Qinghai-Tibet Plateau. [Results] The results show that: (1) The seasonal mean LTLRs are -6.12 ℃/km, -7.63 ℃/km, -5.89 ℃/km, and -3.23 ℃/km for spring, summer, autumn, and winter, respectively. LTLRs in spring and summer are generally steeper than those in autumn and winter. However, in the Hengduan Mountains, the mean winter LTLR is about 0.57 ℃/km higher than that in summer. (2) The maximum LTLRs in spring and summer (-14.45 °C/km and -13.92 °C/km) are steeper than those in autumn (-13.60 °C/km) and winter (-11.61 °C/km). Due to high elevation and dry, clear atmospheric conditions, the Qiangtang Plateau exhibits large seasonal differences in maximum LTLR. The winter maximum LTLR is the smallest, at approximately -13.67 ℃/km. (3) The minimum LTLR in summer is the steepest, about 3.05 ℃/km higher than in other seasons. Minimum LTLRs in the Hengduan Mountains remain relatively high year-round. For instance, the spring minimum LTLR (-1.16 ℃/km) is steeper than in other seasons. The lowest minimum LTLR occurs in autumn (0.03 ℃/km). Across all seasons, the Qiangtang Plateau exhibits the smallest minimum LTLR values. (4) Diurnal variations show that in spring, autumn, and winter, the steepest LTLRs occur between 11:00 and 14:00. The daily minimum in spring appears between 20:00 and 23:00, occurring about one hour earlier in autumn. In contrast, summer shows two steepest-LTLR periods, around 4:00—7:00 and 15:00—18:00, with the daily minimum occurring between 21:00 and 23:00. [Conclusions] This study provides new insights into the spatiotemporal variation characteristics and underlying mechanisms of land surface temperature lapse rates across the Qinghai-Tibet Plateau at seasonal scales.

[Objectives] Under the dual pressures of pervasive global urbanization and accelerating climate change, the synergistic effects of extreme heat and ground-level ozone pollution collectively pose an increasingly severe and multifaceted challenge to urban public health, significantly compromising human well-being, and fundamentally undermining sustainable urban development. However, existing scientific research predominantly focuses on individual environmental stressors in isolation, leading to a critical and pronounced lack of systemic understanding regarding the complex spatiotemporal evolution patterns and intricate underlying formation mechanisms of compound heat-ozone risks, which severely constrains the scientific development and effective implementation of robust and evidence-based urban climate adaptation strategies. [Methods] In this study, we investigate 45 major China's cities by integrating multi-source data encompassing air temperature, ozone concentration, population distribution, and housing prices to meticulously delineate the spatiotemporal patterns of urban exposure to compound heat-ozone risks between 2003 and 2022. Furthermore, leveraging advanced analytical methods such as the Gini coefficient and Theil index, we comprehensively explore the equity characteristics of this risk exposure from the dual perspectives of spatial imbalance and social fairness, while concurrently employing a random forest model to rigorously dissect its complex underlying driving mechanisms. [Results] (1) From 2003 to 2022, urban exposure to extreme heat and compound heat-ozone risks exhibited dynamic temporal fluctuations, whereas ozone pollution showed a clear and pronounced upward trend over time. High-exposure areas for these compound risks were primarily concentrated in core metropolitan regions such as Beijing, Shanghai, and Guangzhou. (2) Urban risk exposure exhibits pronounced environmental inequities, with ozone pollution disproportionately concentrated in areas with low housing prices, and low-income populations in many cities bearing a disproportionately high burden of compound risks. (3) Impervious surfaces were identified as the primary driving factor for various urban risk exposures, consistently followed by environmental factors such as wind speed and precipitation. All influencing factors demonstrated significant nonlinear threshold effects on environmental risk exposure, notably pronounced in the contexts of ozone pollution and compound risks. [Conclusions] This study substantially advances the scientific understanding of the evolution of urban compound heat and ozone risks by elucidating their spatiotemporal dynamics, patterns of environmental inequity, and complex underlying driving mechanisms. It thereby provides a robust and actionable evidence base to inform the development of integrated urban spatial governance frameworks that simultaneously enhance climate resilience and actively promote environmental equity.

[Objectives] Under the context of global climate change, hydrological drought events in the Yangtze River Basin occur frequently with increasing intensity, threatening water security, food security, and economic development. [Methods] Previous studies have primarily considered climate factors such as precipitation and temperature, often neglecting the substantial impacts of human activities—particularly large-scale reservoir operations. To address this limitation, this study comprehensively incorporated both climatic and anthropogenic influences on hydrological droughts. Temperature and precipitation were selected as representative indicators of climate change, while reservoir regulation was used to characterize human activities. Based on the Generalized Additive Models for Location, Scale, and Shape (GAMLSS) framework, a Nonstationary Standardized Runoff Index (NSRI) was developed, incorporating temperature, precipitation, and a reservoir index as covariates. Run theory was applied to identify and evaluate drought events, enabling analysis of the temporal and spatial evolution patterns and characteristics of nonstationary hydrological droughts in the Yangtze River Basin. Five representative hydrological stations—Cuntan, Hankou, Datong, Chenglingji, and Hukou—were selected to represent different sub-basins of the Yangtze River system. [Results] Results show that: ① During 1961—2020, the runoff at five hydrological stations in the Yangtze River Basin showed a significant increase from January to March (non-flood season) and a marked decrease from September to October (flood season). Significant abrupt changes in runoff occurred from January to April and September to November, though the timing of these shifts varied by month. ② The temporal evolution patterns of NSRI and the traditional Standardized Runoff Index (SRI) were generally consistent, yet notable differences were observed in characterizing drought intensity—particularly for extreme events, demonstrating the rationality of the GAMLSS model. ③ There was significant spatial heterogeneity in the evolution of hydrological droughts in the basin, with intensifying drought in the middle-upper reaches and a wetting trend in the lower reaches. ④ Compared to SRI, NSRI identified a slightly higher drought frequency, shorter average duration (by 2.80%~15.25%), and lower drought severity, indicating that NSRI was more sensitive to short-term drought events. Spatially, the response to non-stationarity varied noticeably among stations and their corresponding catchments. [Conclusions] Overall, the long-term hydrological series in the Yangtze River Basin exhibited significant non-stationarity. The NSRI, constructed using temperature, precipitation, and reservoir index as covariates, has greater advantages in revealing the spatiotemporal variation trends of non-stationary hydrological droughts. This study provides theoretical references for drought prevention and mitigation policies in the Yangtze River Basin.

[Objectives] Exploring the evolution mechanism and regulatory paths of the Urban Heat Island (UHI) effect against the backdrop of rapid urbanization, so as to provide scientific basis and methodological support for the management and control of urban thermal environment. [Methods] To address the deficiency of the traditional "urban-rural dichotomy" in characterizing the spatial differentiation of the heat island effect, this research takes Zhengzhou as the study area and constructs a multi-assessment framework for urban thermal environment by integrating the Local Climate Zone (LCZ) theory, source-sink theory, and morphological spatial pattern analysis. Landsat 8-9 remote sensing images from 2016, 2020, and 2024 were used for LCZ classification and Land Surface Temperature (LST) retrieval. Heat sources and heat sinks were divided using distribution indices, and core heat source and heat sink areas were identified through morphological spatial pattern analysis and connectivity analysis. The weights of natural and architectural resistance factors were determined by the Stacking ensemble learning model, and a cumulative resistance surface was constructed. Combined with circuit theory, key corridors and barrier points were identified. [Results] ① From 2016 to 2024, in Zhengzhou, the proportion of the heat source area increased from 29.82% to 34.95%, and the area of core heat source regions expanded from 460.81 km² to 666.75 km², exhibiting a spatial agglomeration trend toward the main urban area. In contrast, the area of core heat sink regions decreased from 3 025.04 km² to 2 672.38 km², with fragmentation significantly intensifying. ② For the core heat source regions, the total length of first-grade corridors increased from 3.30 km to 7.68 km. Conversely, the total length of first-grade corridors in core heat sink regions decreased from 5.27 km to 3.10 km, leading to a decline in the connectivity of heat sinks due to the shrinkage of core heat sink areas. ③ Heat source barriers are concentrated in regions covered by dense trees (LCZA), while heat sink barriers exhibit a trend of widespread diffusion across the entire study area. [Conclusions] Protecting and optimizing key thermal source and sink corridors, as well as reducing thermal sink obstruction points, can effectively mitigate the urban heat island effect. The technical framework developed in this study can serve as a methodological reference for regulating thermal environments in similar cities.