
Geospatial knowledge graphs effectively organize urban entities and semantic relations. When spatial analysis relies on an external spatial database or precomputed spatial relations, joint queries require conversion between the graph and geometry engine. Semantic traversal cannot prune spatial candidates, and maintaining precomputed topology incurs high costs. This paper proposes G3R, a graph reasoning framework that embeds the Geographical Coordinate Subdivision Grid with One-dimension Integer Coding on a 2n-Tree (GeoSOT) and uses hierarchical GeoSOT codes as a spatial representation layer for geospatial knowledge graphs. In G3R, Point, LineString, and Polygon features are jointly represented by a centroid code, vertex code sequence, and longest common prefix code, and stored alongside urban semantic entities and relations in a property graph. Using this representation, G3R supplies a set of spatial operators for distance and direction reasoning, forward and reverse overlay analysis, and K-nearest neighbor (KNN) retrieval. Experiments on urban data from Shenzhen, China show the framework maintains high spatial agreement with PostGIS, improves efficiency in selected retrieval scenarios, and reduces timed processing cost across two representative semantic and spatial analysis tasks. These results indicate G3R offers a feasible intermediate layer linking urban semantics to spatial analysis.
Surface melt is a key component of Antarctic ice-shelf surface hydrology and an important indicator of atmospheric forcing. The Shackleton Ice Shelf experiences particularly substantial summer surface melt, yet its seasonal evolution and spatial organization remain insufficiently resolved at fine scales. Here, we aim to characterize surface melt variability over the Shackleton Ice Shelf using a 40 m Sentinel-1 SAR record covering the 2015–2023 melt seasons. The record reveals a recurring seasonal progression from localized melt onset to widespread summer wetting and subsequent decay, together with pronounced variability among melt seasons. The 2019/2020 season exhibited the most extensive and persistent melt activity, with melt area exceeding [Formula: see text] in 14 five-day observation windows and a maximum melt area approaching [Formula: see text]. Surface temperature and melt proportion exhibited coherent seasonal phasing and a significant but moderate relationship, supporting the physical consistency of the Sentinel-1 record while showing that temperature alone does not fully describe melt evolution. Anomaly analysis further identified departures from the expected seasonal trajectory, including localized early-season activity in 2022. The high-resolution record provides an observation-based baseline for tracking future changes in melt timing and spatial extent and for improving regional understanding of evolving surface melt regimes across East Antarctic ice shelves.
Fine-grained land cover classification in desert ecosystems is essential for enhancing our understanding of ecological dynamics, enabling effective desertification monitoring. However, conventional classification schemes and single models often struggle to capture the intricate spectral and spatial heterogeneity of fine-grained classification. To address these limitations, we propose the SegFormer-Adaptive Random Forest (Seg-ARF) method, a hybrid framework that combines spectral index thresholding with a SegFormer-based encoder and an ARF classifier. A refined classification system is developed to distinguish sandy land into mobile, semi-fixed, and fixed types. This method integrates spectral and deep learning-based representations, thereby achieving enhanced pixel-level classification accuracy. Its effectiveness was demonstrated on Landsat imagery from multiple continents. Our findings indicate: (1) Comparisons between RF, XGBoost and the proposed Seg-ARF verify the advantages of our method. Relative to the RF baseline, Seg-ARF achieved a significant overall accuracy (OA) improvement of 7.09%. (2) Seg-ARF achieves an overall accuracy of 93.96% and a Kappa coefficient of 0.9247 across 100 test regions. (3) Seg-ARF exhibits remarkable adaptability across continents, effectively distinguishing fine-grained land cover types. The Seg-ARF method provides a scalable and transferable solution for large-scale desert mapping and long-term monitoring, with significant potential for applications in desertification mitigation.
Suspended particulate matter (SPM) monitoring in arid lakes is limited by gaps in satellite observations and field measurements. We reconstructed monthly SPM in Ebinur Lake, China, during the ice-free months (May–October) of 2014–2024 using imagery from the Moderate Resolution Imaging Spectroradiometer (MODIS), Landsat 8, and Sentinel-2, together with spatiotemporal fusion and machine-learning models calibrated with 220 in situ samples. The Spatial and Temporal Nonlocal Filter-Based Fusion Model (STNLFFM) outperformed the Enhanced Spatial and Temporal Adaptive Reflectance Fusion Model (ESTARFM) in reconstructing reflectance, achieving validation R2 values of 0.769 (Landsat 8) and 0.861 (Sentinel-2). Adding environmental predictors, especially wind-related variables, improved SPM retrieval. The selected Extra Trees model achieved R2 = 0.816 and a root mean square error (RMSE) of 113.39 mg/L in five-fold cross-validation while better preserving observed SPM variation than alternative models. Reconstructed monthly mean SPM ranged from 167 mg/L in July 2022 to 413 mg/L in September 2015, with seasonal and interannual variability but no consistent monotonic trend. These findings show that combining image fusion with spectral and environmental predictors supports long-term SPM monitoring in data-scarce arid lakes.
Intensive mining activities in the Yellow River Basin severely disrupt regional ecosystems, making fine-scale vegetation classification critical for ecological monitoring. However, mapping highly heterogeneous mining landscapes remains challenging due to spectral-structural confusion among trees, shrubs, and grasses. To address this challenge, we developed a multi-source data fusion pipeline and evaluated both single models and a Stacking Ensemble Learning (SEL) framework. The SEL framework integrates a novel deep learning model (CNN_RepMLP) with shallow machine learning algorithms, all driven by UAV LiDAR and hyperspectral data. An optimal 60-dimensional feature space was constructed. Results show: (1) Multi-source spatial-spectral integration is the primary driver of accuracy improvement, raising the standalone OA of CNN_RepMLP from 64.24% to 86.96%. (2) The SEL framework achieved an OA of 87.17% and a Kappa of 0.8305, not statistically significantly different from the best single model (p = 0.7545), indicating competitive and stable performance without overfitting. Quantitative calibration analysis confirms that Logistic Regression achieves the lowest Expected Calibration Error (ECE = 0.0221), demonstrating superior probability calibration over complex tree-based meta-learners.