[Objectives] The trajectory knowledge graph effectively captures the deep semantic relationships between trajectories and geospatial entities, offering significant advantages in revealing complex associated information. However, traditional methods for constructing knowledge graphs from domain-specific data sources rely heavily on expert knowledge, involve extensive data preprocessing and entity-relationship extraction, and require high levels of professional expertise. [Methods] To address these challenges, this paper proposes a trajectory knowledge graph construction method that supports natural language-driven task execution through prompt learning with large language models. First, a prompt strategy for the preprocessing task is designed to guide large language models in automatically generating data processing code for cleaning abnormal trajectories. Second, a two-level system prompt strategy is developed to enable tool invocation by matching and calling the trajectory knowledge extraction tool. This strategy allows non-expert users to complete the graph construction process using simple natural language instructions, significantly reducing reliance on programming skills and deep semantic understanding. [Results] To evaluate the feasibility and effectiveness of the proposed prompt strategies, a set of test sentences was created for trajectory preprocessing and entity-relation extraction tasks. Real-world ship and vehicle trajectory datasets were used to support knowledge graph construction. Experiments conducted on two representative large language models, Tongyi Qianwen and Baidu Qianfan, achieved average accuracy rates exceeding 75% and 80%, respectively, demonstrating strong generalization ability and practical value. [Conclusions] This study verifies the effectiveness of combining large language models with prompt learning in constructing trajectory knowledge graphs with low technical barriers, demonstrating the strong generalization and application value of the proposed prompt strategy.
[Objectives] The deep integration of Graph Neural Network (GNN) and Building Information Modeling (BIM) has significantly advanced the Architecture, Engineering, and Construction (AEC) industry. In this paper, we propose a self-supervised GNN model that enhances node features to address the spatial identification and classification problem in BIM, thereby closing the gap between high‑fidelity BIM data and graph‑based representation learning through a task‑oriented framework that captures both geometric and functional semantics. [Methods] The model represents each building space as a node and fuses edge features based on node characteristics. To maximize the descriptive power of the features, topologically meaningful relationships—shared‑wall, passageway, corner, and open‑space connections—are encoded as weighted edges and injected into the node embeddings through a learnable fusion layer before graph propagation. The model is trained on an autonomously constructed knowledge base that includes 12 space types and 4 relationship features across 3 major types of building-space layout graphs, enabling the automatic recognition of functional space types in buildings. [Results] The experimental results show that (1) on the BuildingGraph benchmark, which spans school, apartment, and office typologies and comprises 300 large‑scale architectural layout subgraphs, the enhanced GNN with fused edge features achieved accurate building‑space node classification, attaining 96.83% accuracy while maintaining stable convergence of the training loss on graphs containing thousands of nodes. (2) Compared with existing GNN models under identical hyper‑parameter settings in ablation experiments, BGFEF significantly improved performance in terms of accuracy (97.08%) and F1 score (96.75%), surpassing Graph‑BERT and SAGE‑E by 3%~12% (3) In real‑world applications, nine BIM models across school, apartment, and office buildings were evaluated by extracting room connectivity topology graphs, and the method's spatial classification accuracy and interoperability were verified, with the best single school instance reaching 96.87% accuracy. [Conclusions] This study improves the efficiency and automation of spatial analysis and provides practical guidance for implementation in real AEC projects. By producing high‑confidence functional labels that map seamlessly to Industry Foundation Classes (IFC), the proposed model pipeline accelerates rule‑based design review, facility management, and evacuation‑route optimization, offering a scalable foundation for future ontology‑grounded graph reasoning across diverse building typologies.
[Objectives] Current vector tile maps often use a construction method that divides vector data into a series of fixed-size tiles. However, this method fails to account for the agglomeration and heterogeneity of geospatial data. As a result, it leads to significant imbalances in data distribution across tiles, particularly causing prolonged loading times in data-intensive map areas and consequently reducing visualization efficiency and user experience. To systematically address these limitations, we propose a combined grid division methodology for generating variable-size vector tiles that dynamically responds to spatial data density variations. [Methods] This study presents an innovative framework for large-scale vector tile processing, implemented through three key technical phases with enhanced computational efficiency. First, the method constructs a hierarchical vector tile pyramid multi-scale information model by precisely calibrating scale parameters at each level to perform multi-level vector feature simplification and selection. Simplification thresholds for points, lines, and polygons are determined based on optimal cartographic representation requirements, ensuring accurate representation of geographic features across different zoom levels. Second, an adaptive grid division method integrating quadtree and k-d tree (k-dimensional tree) strategies is employed. By dynamically adjusting the grid structure based on a vector tile data volume threshold, this approach enables global dynamic partitioning and local load balancing. Vector tiles are then generated in parallel across all levels. Finally, through hierarchical segregation of tile storage across pyramid levels and the implementation of Geohash encoding to establish a variable-size multi-resolution indexing framework, the system enables rapid vector tile retrieval, delivering seamless map services and optimized resource utilization. [Results] We tested the method in three cities with distinct spatial characteristics—Beijing, Chongqing, and Lanzhou—using vector datasets including points of interest, road networks, and building footprints. We conducted a comparative evaluation across four methods: uniform grids, quadtree, k-d tree, and the proposed combined grid division approach. Key metrics such as tile generation efficiency, data balance, and rendering response time were analyzed. The results demonstrate clear advantages of the proposed method.It achieves the highest tile generation efficiency, reduces redundant tile production, and maintains balanced data distribution, with a coefficient of variation consistently below 0.28, ensuring higher-quality tiles. In data-intensive viewports(e.g., Chongqing's 9th-level range),the proposed method improves loading efficiency by 43.03%, 26.23%, and 19.17% compared to the uniform grid, quadtree, and k-d tree methods, respectively. [Conclusions] The combined grid division method fundamentally resolves the core deficiencies of traditional uniform grid approaches in data-dense map regions. It effectively mitigates excessively large tile sizes, severe load imbalance, and associated rendering delays.
[Objectives] Fine-grained object recognition in remote sensing is a fundamental yet highly challenging task within both Earth observation and computer vision. It involves the accurate localization and detailed classification of objects in High-Spatial-Resolution (HSR) imagery, which often features highly complex backgrounds, inter-class similarities, and intra-class variations. In recent years, notable progress has been driven by algorithms that jointly exploit pixel-level, object-level, and neighborhood-level information. These approaches combine semantic features, texture characteristics, and spatial contextual relationships to form multi-source and multi-scale feature representations. Despite these advances, existing methods remain inadequate for directly utilizing higher-level fine-grained knowledge such as scene composition, entity semantics, attribute descriptions, and temporal dynamics. The core limitation lies in the absence of a formalized knowledge organization and representation paradigm capable of systematically bridging low-level visual perception and higher-order semantic reasoning. [Methods] To address these limitations, this study proposes a multi-level knowledge graph-based organization and representation framework specifically designed for fine-grained remote sensing object recognition. The framework adopts a four-layer hierarchical structure encompassing scene, entity, feature, and change dimensions, enabling dynamic and semantically rich descriptions of remote sensing targets. In this structure, scene nodes provide contextual constraints, entity nodes capture essential connotations of objects, feature nodes encode visual and semantic attributes, and change nodes represent temporal evolution. [Results] By incorporating spatiotemporal references, spatial morphology, and inter-object relationships, the proposed approach enables knowledge organization under multiple constraints, including scene, entity, feature, and temporal conditions. In doing so, it moves beyond purely data-driven perception and establishes a mechanism for knowledge-driven reasoning in remote sensing interpretation. Extensive experiments were conducted to validate the effectiveness of the proposed framework. When integrated into the baseline model STD, the knowledge graph yielded an improvement of approximately 3.82% in mean Average Precision (mAP) and 3.92% in recall, demonstrating its ability to enhance detection accuracy. Beyond this single case, the universality and robustness of the framework were confirmed by consistent performance improvements across several representative neural networks, including Oriented R-CNN, Oriented RepPoints, LSKNet, and STD. These results indicate that the proposed method not only improves recognition performance but also enhances interpretability and adaptability across heterogeneous architectures and datasets. [Conclusions] Overall, this study demonstrates that a multi-level knowledge graph provides an effective pathway for advancing fine-grained object recognition in remote sensing, transitioning from feature perception to knowledge reasoning. The method not only increases recognition accuracy but also enhances semantic interpretability and dynamic adaptability, offering a scalable solution for intelligent remote sensing analysis. Importantly, it provides new theoretical and practical insights for applications in geospatial information extraction, environmental and urban monitoring, disaster assessment, and military intelligence analysis. By systematically integrating structured knowledge with data-driven models, the proposed framework enriches the semantic depth of remote sensing interpretation and demonstrates strong potential for future developments in intelligent Earth observation systems.
[Objectives] The weight adjustment of outlier samples is a key factor influencing the accuracy of spatial interpolation of target attribute, and the existing methods cannot balance the geographic evenness and feature space representativeness of samples. [Methods] Taking the soil samples of soil organic matter in Horqin Left Wing Middle Banner as an example, a geographical-feature space weight adjustment method of outlier samples for spatial interpolation was proposed. Firstly, division methods of sample types in geographical space and feature space were developed to divide sample types, respectively. Secondly, global outlier samples and local outlier samples were respectively detected using the quartile method and local Moran's I. Given the sample types, the adjustment amount of sample weights and the adjustment rules were determined, and the weights of outlier samples were corresponding adjusted. Finally, different weight adjustment schemes for outlier samples and different spatial interpolation models were set up for comparative experiments, respectively, and their spatial interpolation accuracies were analyzed to evaluate the advantages and disadvantages of different weight adjustment schemes and spatial interpolation models. [Results] The results showed that 34 clustered samples, 380 evenly distributed samples and 19 sparse samples were obtained based on the geographical space division method, and 172 samples with low representativeness and 261 samples with high representativeness were obtained based on the feature space division method. Through the detection of outlier samples, 35 local outlier samples and 3 global outlier samples were obtained. According to the weight adjustment rules of outlier samples, a geographical-feature space weight adjustment method was constructed. Compared with the original samples, the standard deviation and variation coefficient of the samples after weight adjustment were reduced, and the data was closer to the normal distribution. The root mean square error and the mean absolute error of spatial interpolation using geographical-feature space weight adjustment method were reduced by 11.25% and 11.07%, respectively, and the corresponding accuracy was increased by 11.82%. The correlation between the predicted values and the measured values was significantly higher than that of the original data. Compared with other comparison schemes, the weight adjustment sample number was the smallest, and the adjustment optimization effect was the best. The interpolation results of the method developed in this study were better than the results of random forest model for the original samples, but worse than spatial interpolation results using random forest regression kriging model. [Conclusions] The geographical-feature space weight adjustment method developed in this study can balance the geographic evenness and feature space representativeness of outlier samples, and significantly improve the spatial interpolation accuracy of soil organic matter under the scenario of the minimum number of outlier samples in weight adjustment.
[Objectives] Accurate nowcasting of rainstorms remains challenging due to the chaotic and rapidly evolving nature of precipitation systems. In response to the complex characteristics of rainstorms, recent research has shifted from relying solely on deterministic or probabilistic methods to hybrid frameworks that combine both. However, many existing hybrid models remain purely data-driven and lack explicit physics-informed priors, resulting in limited physical consistency and generalization, especially under extreme rainfall conditions. [Methods] To address this, we propose a physics-informed prior-guided framework for rainstorm nowcasting that integrates physically-derived priors from partial differential equations into both a deterministic forecasting module and a diffusion-based spatiotemporal refinement module. The deterministic branch uses a multi-layer decoder with PDE-guided gating units to generate physically consistent forecasts, while the diffusion branch models residual uncertainties conditioned on physics-aware hidden states to refine fine-scale structures. This hybrid design captures both large-scale precipitation evolution and localized stochastic variations in a physically grounded manner. [Results] To evaluate our framework, we compared it against three representative state-of-the-art models covering different paradigms: the physics-guided deterministic model PhyDNet, the conditional GAN-based generative model DGMR, and the cascaded hybrid framework DiffCast that integrates deterministic prediction with diffusion-based refinement. Experimental results on a radar-precipitation dataset show that our method achieves the highest overall Critical Success Index (CSI) of 0.235 for predicting rainfall above 16 mm/h within a 3-hour forecast horizon, surpassing DiffCast (0.223) by 5.4%, PhyDNet (0.213) by 10.3%, and DGMR (0.140) by 67.9%. Frame-wise evaluations indicate that our model maintains stable advantages across all forecast lead times, with CSI improvements of 8.3%, 8.6%, and 6.2% over the best baseline in the first, second, and third forecast hours, respectively. Visual comparisons further reveal that our method provides sharper and more physically plausible rainfall structures, effectively mitigating over-smoothing issues in deterministic models and unrealistic spatial patterns in purely generative models. [Conclusions] These results underscore the importance of incorporating physics-informed priors for improving deep learning-based nowcasting. By leveraging PDE-based representations and diffusion-based residual correction, our framework achieves more accurate and physically consistent forecasts, particularly for extreme rainfall events where conventional models often struggle. This demonstrates that hybrid architectures guided by explicit physical knowledge offer a promising direction for enhancing both forecast reliability and interpretability in operational precipitation nowcasting.
[Objectives] This study presents a comprehensive optimization and enhancement of the spatio-temporal simulation framework for land use change by integrating two key components: a multi-source data transformation rule mining module and a dynamic improvement mechanism based on GBDT. The proposed approach significantly improves the capacity to capture complex, long-term patterns of land use evolution, enabling high-precision simulations through advanced machine learning techniques and adaptive modeling strategies. [Methods] This study takes Guangzhou City, Guangdong Province as a case and innovatively proposes a Land Use Simulation model (GDLUS) coupled with the GBDT multi-source data conversion rule mining and dynamic improvement module. This model builds a multi-source dataset based on multi-year land use data and driving factors, and uses Gradient Boosting Decision Tree (GBDT) to mine the mechanism of land use evolution. Secondly, a dynamic improvement module was introduced, which includes a parameter extraction module to construct a transition probability matrix and its range based on long-term land use evolution data. The target decomposition module solves and refines the future land use requirements in the time dimension based on the range of transfer probability. In the spatial dimension, it conducts multi-scale spatial evaluations such as neighborhood effect and development probability based on the results of neighborhood effect and GBDT mining, and divides the space into different types according to the transfer intensity of land use types. The weighted random sampling module regulates based on the differences between micro-layout and macro-demand during the evolution process, controlling the transformation between land use types. Finally, the GDLUS model was applied to the multi-scenario simulation of Shared Socio-economic Pathways (SSPs) to simulate the land use pattern in Guangzhou under multiple scenarios in 2050. [Results] The GDLUS method significantly improved the accuracy rate and FOM indicators by 18% and 33% respectively compared with the mainstream model PLUS. In particular, the dynamic improvement module has slightly enhanced the capture and simulation accuracy of local land use changes, significantly increasing the FOM index of the model by 14.5%. The multi-scenario simulation based on the Shared Socio-economic Pathways (SSPs) shows that under the SSP5 scenario, the water bodies and green Spaces in the study area are at risk of large-scale occupation of urban land, and the problem of ecological space degradation is extremely prominent. In the SSP1 scenario, the regional ecological space demonstrates strong resilience characteristics. [Conclusions] The GDLUS model provides more accurate and efficient technical support for the simulation of land use changes and its application in urban planning decisions.
[Objectives] Buildings and roads constitute fundamental components of urban spatial structures, forming the physical foundation and functional skeleton of modern cities. Their information is of critical importance for a wide range of advanced applications, including urban planning, disaster response, autonomous driving, and intelligent transportation management. The collaborative extraction of buildings and roads has emerged as an key research topic in interpretation of high-resolution remote sensing images based on the strong spatial correlation between roads and buildings. However, existing methods have not fully utilized the spatial structure between the two types of objects and their individual morphological characteristics, resulting in issues such as unclear contours and inability to detect some small-sized objects. [Methods] To address this issue, this paper presents a Structure-Aware Collaborative Extraction Network (SACE-Net) to enhance the collaborative extraction of buildings and roads from high-resolution imagery. SACE-Net is designed to comprehensively leverage both the spatial structural dependencies and the morphological diversity of the two object types, thereby improving their feature representation and mutual understanding within a unified framework. Specifically, the proposed network introduces a Feature Space Interaction Module (FSIM) by employing a query-guided cross-attention mechanism, enabling dynamically learn complementary spatial structures between buildings and roads across multiple feature scales. This design allows contextual information from one object type to guide the perception of the other. In addition, SACE-Net incorporates a Dual-Branch Decoding Module (DBDM), in which each decoding branch is specialized for one object type. By integrating attention-guided and direction-aware mechanisms, the decoder adaptively captures elongated road geometries and complex building contours. [Results] Experiments conducted on two public datasets, Massachusetts and AIOI, demonstrate that SACE-Net outperforms seven baseline deep learning methods in both quantitative and visual evaluations. Specifically, SACE-Net achieves an average Intersection over Union (IoU) score of 75.82% and 64.14% on the Massachusetts and AIOI datasets, respectively, surpassing the state-of-the-art baseline extraction methods by 10.70% and 4.90%, respectively. Visualization results further confirm that the proposed model effectively preserves building contour integrity and road connectivity and significantly reduces issues such as boundary ambiguity, omissions, and false detections. Ablation experiments show that incorporating the FSIM and DBDM into the baseline methods improves the average IoU by 18.96% and 14.01% on the two datasets, respectively, demonstrating the effectiveness and necessity of the proposed modules. [Conclusions] By enhancing the perception of spatial and morphological structural features of buildings and roads, SACE-Net significantly improves the accuracy of collaborative feature extraction. This work provides a practical framework and methodological reference for multi-object collaborative extraction in high-resolution remote sensing imagery.
[Objectives] Remote sensing assessments in complex mountainous regions are frequently undermined by topography-induced biases and inconsistent multiyear weighting. Specifically, illumination and slope effects distort NDVI retrievals, while year-by-year Principal Component Analysis (PCA) produces fluctuating component weights that limit the Remote Sensing Ecological Index's (RSEI) temporal stability and comparability. These problems are acute in high-relief landscapes such as the upper Jinsha River Basin, where steep slopes, varied aspect, and heterogeneous vegetation cover pose substantial challenges for long-term ecological monitoring. To overcome these limitations and improve both temporal consistency and spatial responsiveness, we propose a Global-Weighted Modified Remote Sensing Ecological Index (GW_MRSEI). [Methods] The GW_MRSEI replaces the conventional NDVI with a slope-corrected Normalized Difference Vegetation Index (SNDVI) to mitigate topographic illumination artifacts and better represent vegetation reflectance under varying slope and aspect conditions. To stabilize interannual weighting while preserving local sensitivity, we embed a local weight correction coefficient within a Global Principal Component Analysis (GPCA) framework; this reconciles globally consistent weighting with regional adaptability and enhances detection of abrupt ecological changes. We applied the method to the Yunnan-Tibet segment of the Jinsha River Gorge. Model validation and temporal analysis employed Theil-Sen trend estimation and the Hurst exponent to characterize trend persistence, while an XGBoost model identified and quantified principal driving factors. [Results] Compared to traditional approaches, SNDVI more faithfully captured vegetation dynamics across complex terrain, reducing spurious signals associated with illumination and slope. Quantitatively, its discriminative capacity improved by 19.31% over NDVI, indicating a more accurate representation of vegetation variability under rugged topography. GPCA produced markedly more stable interannual component weights than entropy-based or random-forest weighting schemes, ensuring consistent dimensional contributions to the composite index. GW_MRSEI showed improved trend coherence and heightened local sensitivity, suppressing year-to-year noise and effectively revealing ecosystem responses during extreme climatic years. Between 1990 and 2024, 54.3% of the study area exhibited net ecological improvement—concentrated in the Yunling Mountains and northern basin—whereas dry-hot valley zones accounted for the principal degradation. Projections indicate 31.6% of the area is likely to continue improving while 29.5% may remain degraded. Spatial variation was chiefly driven by vegetation cover, temperature, and elevation; temperatures exceeding 28 ℃ negatively affected vegetation, and elevations of 2 500~4 000 m provided comparatively favorable conditions. [Conclusions] GW_MRSEI integrates topographic correction, stable global weighting, and localized adaptivity to deliver enhanced temporal robustness and spatial responsiveness. The index supports long-term ecological tracking, anomaly detection, and transferable ecosystem assessment in mountainous regions, informing targeted restoration and conservation strategies.
[Objectives] Under the combined influence of factors such as mining activity, hydrology, and geology, potential landslides in open-pit mines exhibit characteristics such as widespread distribution and clustering along mining slopes. Furthermore, their InSAR deformation signals often resemble those of the surrounding background environment, leading to misclassification or missed detection by existing detection algorithms. At the same time, the contradiction between the multi-source triggering mechanism of potential landslides and the "black box"nature of detection models undermines the interpretability of detection results and limits the applicability of current models in mining areas. [Methods] To address these problems, this study proposes an augmented detection framework that integrates InSAR and interpretable deep learning, using YOLOv5n as the baseline model, to improve the accuracy and efficiency of detecting potential landslides in mining areas. First, the annual mean deformation velocity and time series are calculated using the SBAS-InSAR technique to construct an InSAR dataset for landslide hazard detection, and a geographic environment similarity criterion is introduced to optimize its spatial distribution. Next, a comprehensive identification method combining InSAR deformation information and morphological features from optical imagery is developed to rapidly identify potential landslides. Subsequently, SHAP is employed to identify key disaster-causing factors, and a SHAP-RF/SVM model is constructed to evaluate landslide susceptibility, thereby characterizing the clustering patterns of potential landslides along mine slopes. Based on this, a mechanism-driven spatial prior term incorporating susceptible zones and a gradient adjustment strategy is proposed to construct a risk-aware supervised penalty mechanism that guides the model to enhance feature attention in areas with high concentrations of potential hazards. [Results] Using an open-pit mine in Xilingol League, Inner Mongolia Autonomous Region, as the study area, this research employs Sentinel-1A SAR imagery from January 2020 to January 2022 for interpretable modeling and enhanced detection of potential landslides. The results demonstrate high spatial consistency and strong interpretability between the detection outcomes and susceptibility assessments, along with significantly improved detection accuracy. The mAP50 and F1-score reach 95.9% and 91.7%, respectively, exceeding the performance of other representative models, such as YOLOv5 and Attention U-Net, by at least 8.76% and 10.11%. [Conclusions] The proposed framework also shows high accuracy and practical applicability in open-pit mines with varying surface conditions. It provides important technical guidance for applying InSAR and deep learning methods to the detection of potential landslides in open-pit mining environments.
[Significance] The development of centralized photovoltaic power stations in desertified regions serves dual purposes of energy transition and ecological restoration. This study addresses three primary challenges in current research: (1) Limited accuracy in identifying photovoltaic arrays against highly reflective desert backgrounds; (2) Lack of reliable mechanisms for quantitatively assessing carbon sequestration benefits from vegetation beneath PV panels; (3) Absence of a carbon benefit accounting system that simultaneously incorporates "emission reductions above panels" and "carbon sequestration beneath panels." To address these gaps, we propose an integrated extraction and carbon benefit accounting method that combines deep learning with multi-source remote sensing data. [Methods] First, a sample database for desertified areas was constructed using high-resolution remote sensing images, and a DeepLabv3+ semantic segmentation model with ResNet50 as the backbone network was applied to achieve high-precision extraction of photovoltaic power plants under high-reflectance backgrounds. Second, carbon emission factors were calculated through life cycle assessment, and emission reduction benefits were estimated by combining grid emission factors with power generation data. Furthermore, multi-temporal NPP remote sensing data were used to compare vegetation productivity between photovoltaic and background areas, thereby quantifying the additional carbon sequestration resulting from vegetation recovery beneath panels. Finally, emission reduction and carbon sequestration benefits were monetized using carbon trading prices, and their spatial distribution patterns were systematically analyzed. [Results] Taking the Hexi Corridor region in Gansu Province as a case study, a total of 469.63 km2 of centralized photovoltaic installations were identified. Over a 25-year lifecycle, these systems achieved cumulative emission reductions of 48.61 million tCO2e and additional carbon sequestration of 2.77 million tCO2e, resulting in total carbon benefits of CNY 4.11 billion. Among the study areas, Wuwei City showed the strongest sequestration capacity, Jiayuguan City exhibited the highest emission reduction density per unit area, and Jiuquan City ranked first in total emission reductions but demonstrated relatively limited ecological improvement. [Conclusions] The method presented in this study marks significant advances in both the accuracy of identifying photovoltaic power plants and the framework for assessing carbon benefits. Compared with traditional machine learning methods such as Random Forest, our approach reduces misclassification between PV arrays and high-reflectance backgrounds (e.g., bare rock) by 47%. It also demonstrates greater robustness and overall accuracy than the U-Net model, especially in complex environments. In addition, this study introduces a quantitative method for evaluating vegetation-based carbon sequestration in areas beneath PVpanels, confirming the ecological restoration effects associated with PV installations. For the first time, the integrated carbon accounting system developed here combines "emission reduction from power generation" and "carbon sequestration by panel-side vegetation" into a unified assessment framework, improving both the systematization and spatial accuracy of evaluation. This work provides scalable methods and scientific support for advancing the coordinated development of renewable energy and ecological restoration through PV projects in desertified regions.
[Objectives] Farmland distributes widely in Loess Plateau region with abundant, scattered, and fragmented features. Satellite images are ideal data resources for investigating and monitoring farmland due to their plentiful sources and various spatial resolutions. Currently, there is a lack of relevant research on the identification capability and accuracy assessment of satellite data for cultivated land in different geomorphic units of the Loess Plateau, Scientific and accurate assessment on capability of these satellite data to identify farmland and the accuracy of identification is vital for farmland monitoring. [Methods] Taking Yuzhong County of Lanzhou City as the study area, this study performed an assessment on the ability of images from GF-1, Sentinel-2, GF01-WFV, and Landsat-8 OLI to recognize farmland in different geomorphic units of loess, and analyzed the influencing factors. In quantitative assessment, indicators including positional accuracy, area accuracy, omission error rate, commission error rate, and minimum identifiable plot area were constructed, and the spatial statistical analysis methods were used. [Results] ① The accuracy of farmland identification is highly correlated with the spatial resolution of imagery, but significant differences exist among various assessment indicators. The positional accuracy of farmland extraction is less affected by variations in spatial resolution, the positional accuracies of GF-1, Sentinel-2, GF01-WFV, and Landsat-8 OLI are 95.02%, 94.81%, 91.96%, and 91.66%, respectively. The area accuracy, commission and omission error rate, plot boundary accuracy and minmum identifiable plot exhibit significant variation trends with decrease of spatial resolution of images. The area accuracy decreased from 94.88% to 85.94%, the commission error increased from 3.91% to 16.57%, and the ommission error increased from 7.62% to 19.22%. The average accuracy of plot boundaries for GF-1, Sentinel-2, GF01-WFV, and Landsat-8 OLI are 89.69%, 83.93%, 79.11% and 74.74%. ② Accuracy of farmland identification is significantly influenced by geomorphology, and the higher the spatial resolution, the more susceptible it is to terrain influence. Overall, the accuracy in the Loess Plateau region is 3% to 5% higher than that in the Loess Ravine region, and 5% to 10% higher than in the Loess Hilly region. The ommission error is relatively high in the Loess Plateau region. Significant misclassification is observed in the Loess Ravine region, with an average ommission error of 15.94%, and Sentinel-2 and GF01-WFV both exceeding 16%. In the Loess Hilly region, the difference in accuracy between different image data is reduced. ③ In the Loess Plateau area, fragmented topography and crisscrossing gullies create significant challenges for farmland identification. The spectral characteristics of exposed loess usually overlap with those of fallow and plowed farmland, leading to frequent misclassification. The mapping results show that the cultivated land area extracted from satellite data with a spatial resolution of 2~15 meters is insufficient, while that extracted from 30-meter resolution data is excessive. The accuracy of cultivated land extraction using GF-1 data is over 95%, which is the closest to the actual cultivated land area. ④ Remote sensing imagery with spatial resolution finer than 10 meters can fundamentally satisfy the requirements for monitoring area, location, and cultivation status of farmland in the Loess Plateau region as well as farmland mapping. Selection of satellite imagery should comprehensively consider data availability, monitoring objectives, and accuracy requirements. [Conclusions] The findings of this study provide valuable reference for selecting appropriate remote sensing data sources for farmland monitoring and evaluating accuracy of farmland identification in the Loess Plateau region.
[Objectives] Accurate monitoring of the dynamic changes in burned areas holds significant importance for forest management and ecological restoration. Traditional methods for burned area extraction often face challenges such as low efficiency and suboptimal performance in extracting small burned areas when dealing with large-scale and topographically complex regions. This paper proposes a lightweight network model, FFANet (Lightweight Network Based on Feature Fusion Attention Mechanism), which incorporates a feature fusion attention mechanism, and applies it to the study of burned area extraction. [Methods] Initially, data preprocessing was conducted to create a forest fire change detection dataset. A lightweight model featuring a feature fusion attention mechanism was constructed and trained. The extraction of burned areas was achieved by extracting features from dual-temporal images, fusing feature information, and decoding to classify and output change result maps. The model utilized the U-Net network as its baseline, incorporating a Bitemporal Feature Fusion (BFF) strategy in the feature extraction phase to fully capture differential information between dual-temporal features. A Median-Enhanced Spatial and Channel Attention Block (MESC) was designed to capture detailed information in images by integrating channel and spatial attention mechanisms with various dilation rates. In the decoder section, a lightweight upsampling module was employed, combined with features from the previous level through skip connections, ensuring effective feature fusion while reducing computational costs. The final stage of the model utilized a classifier to output the change detection result map. [Results] This study selected the forest fire that occurred in Yuxi City, Yunnan Province, on April 13, 2023, as the experimental area. Sentinel-2 images taken before and after the fire were used for burned area extraction experiments. The results demonstrated that: (1) The model's lightweight architecture design enabled end-to-end learning, optimizing inference speed by reducing parameter and computational loads, allowing for rapid identification of changed regions. (2) When compared with Change Former, BIT, and STANet models based on various evaluation metrics, the FFANet model achieved an F1 score of 87.11%, representing improvements of 3.19%, 6.06%, and 1.4% over the other three models, respectively. Its Intersection over Union (IoU) value reached 82.64%, with increases of 4.06%, 0.97%, and 6.86% compared to the other models, respectively. These results validated the applicability of the FFANet model in extracting burned areas. [Conclusions] The proposed FFANet model effectively enhances extraction accuracy in complex terrains and for small burned areas, providing technical support for faster and more accurate burned area extraction. It also offers additional exploration avenues for future related research.
[Objectives] Martian rocks are important research content for Mars exploration missions, such as understanding the Martian geomorphology, analyzing the Martian geological evolution process, and landing site selection. Rocks are smaller objects on the Martian surface. Moreover, the different sizes and shapes of the rocks make it a great challenge to extract and analyze them. Boulder Halo is a typical landform in the middle-to-high latitudes of Mars. The extraction and analysis of the rocks around it allows for a better understanding of its geomorphological characteristics. [Methods] In this study, a method based on shadow and sliding window for rock extraction was utilized to extract Martian rocks (rocks with maximum diameters greater than 1.5 m) in HiRISE images. The method created an initial window based on the position and range of the Martian rock shadow, and determined the termination position of the window by sliding it. The range of the rock was determined based on the initial and termination positions of the sliding window. The rocks were represented as fitted ellipses. The length of the rock along the illumination direction was the maximum distance between the initial and termination windows. The length of the rock along the vertical illumination direction is the width of the sliding window in that direction. Twenty-three Boulder Halos between 60° N and 70° N on Mars were selected as the study area to analyze the rocks within and surrounding them. The analysis focused on rock density, rock diameter, and spatial autocorrelation. These analyses could characterize the spatial distribution of rocks around the Boulder Halo. [Results] Analysis of rock density revealed that Boulder Halos are mostly located near areas of high rock density. Moreover, the number of rocks with maximum diameters of 1.5 m to 2.5 m was the highest within 3 times the radius of the Boulder Halo. In most cases, the mean of the maximum diameters of the rocks was relatively large, within 1 to 1.5 times the radius of the Boulder Halo. Spatial autocorrelation analysis of rocks within 3 times the radius of Boulder Halo showed a negative spatial correlation between rock density and radial distance, a weak negative spatial correlation between maximum diameter of rock and radial distance, and a weak positive spatial correlation between maximum diameter of rock and rock density. Combining rock density analysis with bivariate spatial autocorrelation analysis between rock density and radial distance, it can be inferred that the rock density near the one-radius distance of the Boulder Halo is relatively high. [Conclusions] The analytical results can be used as reference data to assist future studies, such as in-depth analysis of the Boulder Halo formation process and revealing the related regions' geological evolution process.
[Objectives] Rural squares, as an essential component of public service infrastructure, face significant challenges in planning and construction, including inefficient layouts, oversized scales, and the illegal occupation of arable land. Current research on the site selection and layout of rural squares has focused primarily on general principles and strategies, leaving a gap in the development ofscientifically rigorous location-allocation models and effective optimization methods. Therefore, this study aims to address this gap by proposing a comprehensive location optimization model for rural public squares. [Methods] This study develops a multi-objective optimization model based on the classical Capacitated Facility Location Problem (CFLP). The model includes three objective functions: minimizing total travel cost, long-distance travel cost, and construction cost. It is subject to five constraints, including effective coverage rate, total square area, and per capita square area. The model is nonlinear and non-convex, characterized by complex spatial coupling. To enhance local search performance and convergence speed, this study improves the NSGA-II algorithm by optimizing the initialization settings, designing a new neighborhood mutation operator, and adopting a hybrid elite strategy. The resulting enhanced algorithm—named NSGA-II-N—integrates both global and local search capabilities and achieves faster convergence. [Results] A case study on rural square planning in Dazhong Village, Xiangyin County, was conducted to validate the model. The analysis examined the relationships among key parameters—effective coverage rate, total travel cost, long-distance travel cost, and construction cost—under different numbers of potential square locations. The findings reveal a nearly linear correlation between effective coverage rate and total travel cost (the determination coefficient: 0.739), indicating that higher travel costs lead to lower coverage. The relationship between effective coverage rate and construction cost follows a logarithmic trend, with the determination coefficient of 0.789. This suggests that initial investments in construction and the addition of more squares result in a rapid increase in effective coverage, followed by diminishing returns with further investment. Long-distance travel cost is negatively correlated with effective coverage rate. When the number of new squares is fewer than 10, long-distance travel cost rises sharply, with some trips exceeding 125 km. Comparative analysis demonstrates that the NSGA-II-N model has superior convergence relative to NSGA-II, NSGA-III, and SPEA2.The median construction cost of NSGA-II-N is only 76.98, with an IQR of 12.13, which is significantly lower than those of the NSGA-II, NSGA-III, and SPEA2, and it also demonstrates better solution diversity than MOEA/D. While the NSGA-II-N model performs similarly to traditional single-objective models in optimizing primary parameters (e.g., cost), it significantly outperforms them in optimizing additional parameters. [Conclusions] The NSGA-II-N model effectively balances construction costs, operational efficiency, and equity. It offers comprehensive optimization across multiple factors—such as construction cost, total travel cost, long-distance travel cost, and effective coverage rate—and offers a scientific basis for the site selection and layout of rural squares and other public service facilities.
