Wetlands on the Qinghai–Xizang Plateau are crucial for ecosystem services and ecological security but are increasingly threatened by climate change and human activities. Previous studies have typically treated wetlands as homogeneous entities and relied on coarse-resolution data, overlooking the distinct responses of water and vegetation-soil components. To address this gap, we proposed a remote sensing-based assessment framework based on a dual water-vegetation structure to evaluate wetland dynamics from 2018 to 2023. By integrating landscape patterns, ecological states, and ecological functions, we constructed Wetland Health Index (WHI) using parallel unsupervised autoencoders. Geodetector and GeoShapley were further applied to identify the driving factors and quantify their spatial effects. Results showed the WHI fluctuated around a mean value of 0.42 at a moderate level during study period, with water components consistently exhibiting higher health scores than vegetated areas. Geodetector identified precipitation and livestock grazing as the dominant natural and anthropogenic drivers, respectively. GeoShapley further revealed geographic location inherently partly shaped the baseline ecological status, while external driving factors exert spatially heterogeneous impacts across geographic zones. These findings provide a spatially explicit perspective on recent alpine wetland health dynamics and support spatially differentiated wetland conservation and management on the plateau.
Spatial layout optimization of urban service facilities aims to maximize system-wide locational benefits through rational facility placement. However, due to pronounced spatial heterogeneity and nonlinear interactions among urban environmental features, locational benefits vary significantly across regions and are subject to varying degrees of uncertainty, making the optimization objective difficult to quantify and causing the spatial layout outcomes to fall short of real-world requirements. To address this issue, we propose a Bayesian-guided hierarchical reinforcement learning framework that learns locational benefits from historical data and incorporates benefit uncertainty into the spatial layout optimization process, enabling layout schemes to simultaneously maximize locational benefits while adapting to spatial heterogeneity. Specifically, we first develop a Bayesian neural network to characterize the nonlinear mapping between geospatial factors and locational benefits, which yields probabilistic benefit estimates with explicit uncertainty quantification for each candidate site. These estimates are then integrated into the optimization pipeline via a hierarchical multi-agent deep reinforcement learning model that generates region-specific location decisions under globally coordinated optimization. Experiments in Shenzhen, China demonstrate that our method outperforms mainstream benchmark methods, delivering over 10% higher total benefits and superior out-of-sample robustness. This work presents a novel and actionable uncertainty-aware spatial optimization framework for smart city development.
Arctic sea ice leads, which present as linear openings in sea ice, play a critical role in modulating ocean–atmosphere interactions and regional heat fluxes. High-resolution pan-Arctic observations remain limited due to the harsh polar environment. This study develops a deep-learning model, LeadNet, for detecting Arctic sea ice leads from Sentinel-1 synthetic aperture radar (SAR) imagery. LeadNet integrates HH-polarized backscatter and gray-level co-occurrence matrix (GLCM) texture features within a dual-branch architecture and retains sensitivity to narrow leads in the 100–300 m range. The trained model is applied to 14,093 SAR scenes acquired from October 2019 to April 2020 to generate an 80 m pan-Arctic LeadNet Arctic Lead Dataset, including monthly lead frequency maps and pixel-level LeadNet classification results for each scene. Based on the dataset, we analyze lead characteristics. The lead frequency maps show clear seasonal variability, with higher activity during autumn and early winter, a minimum in late winter, and an increase in early spring. In addition, analysis in the Beaufort Sea indicates that detected leads narrower than 300 m or shorter than 5 km account for 20–40% of the total lead area. The dataset provides a high-resolution observational benchmark for studies of Arctic sea ice dynamics, heat fluxes, and climate model evaluation.
Accurate estimation of forest aboveground carbon storage (AGC) is essential for understanding the global carbon cycle and carbon sequestration, yet conventional models often assume spatial stationarity, leading to systematic biases in heterogeneous landscapes. This study integrated Sentinel‑2 imagery and spaceborne LiDAR within a geographical random forest (GRF) framework coupled with quantile regression for uncertainty quantification and SHapley Additive exPlanations (SHAP) for interpretability, enabling confidence‑informed AGC mapping and a spatially explicit elucidation of variable importance. The fusion of LiDAR and Sentinel‑2 data consistently improved the accuracy over optical‑only data, and local models outperformed global models in both study areas. The fused GRF achieved R2 values of 0.710 in Jixi (medium effect) and 0.538 in Xintian (small effect). Uncertainty patterns differed markedly: bivariate local Moran’s I revealed a high‑AGC–low‑uncertainty cluster in Jixi versus a high‑AGC–high‑uncertainty cluster in Xintian. SHAP analysis revealed that textural features and vegetation indices dominated in Jixi, whereas spectral bands dominated in Xintian, with the dominant predictors shifting spatially. By integrating these methods into a unified workflow, the framework jointly addresses spatial nonstationarity, prediction uncertainty, and model interpretability, providing a methodological basis for AGC monitoring in complex and disturbed landscapes.
To the best of our knowledge, the joint use of Wi-Fi login records, OpenStreetMap (OSM), and remote sensing imagery for object-oriented campus land use and land cover (LULC) mapping under multimodal small-data conditions has not been previously reported. This paper proposes the first adaptive prototype last ensemble learning (FAPLEL) framework with a coarse-to-fine architecture for object-oriented campus LULC mapping using multimodal small-data. For land cover (LC) mapping, we propose an adaptive prototype zero-shot learning (APZL) method based on a lightweight foundation model. It offers label-free and training-free inference, adaptive adjustment of the number of subprototypes, and providing knowledge graph-like interpretability. For land use (LU) mapping, we propose a spatiotemporal clustering ensemble learning (SCEL) method. This method requires only a single label per cluster, achieving near-full automation and promising accuracy under small-data conditions. A campus case study shows that APZL achieved an mIoU accuracy of 0.796 for coarse-grained LC mapping. Subsequently, the fine LU mapping process, which uses K-means followed by random forest, yields the highest overall accuracy of 0.813 among various supervised and unsupervised classification methods. These results indicate that the proposed multimodal data-driven FAPLEL paradigm has the potential to be extended to urban LULC mapping.
The Himalayas host the planet's highest-elevation mountain forests, serving as natural sentinels of climate change. Despite evidence of accelerated shifting alpine treelines, changes in the entire Himalayan forest system remain understudied. In this study, we derive tree cover from long-term Landsat observations using machine learning algorithms and find a significant overall increase in tree cover by 2.5 percentage points from 1990 to 2020. At the scale of the entire Himalayas, the fastest increase occurred at higher elevations (1,000 to 3,500 m above sea level (a.s.l.)) at an annual rate exceeding 0.15% yr−1, whereas within the monsoonal Himalayas, even higher rates occurred between 1,000 and 2,500 m a.s.l. (>0.3% yr−1). We analyze forest change over the three decades in relation to climate and land cover datasets and validate the results with Google Earth Pro imagery. Forest gains in the monsoonal and westerly Himalayas were mainly associated with rising temperatures and changing precipitation patterns, respectively. Forest losses were dominated by the expansion of cropland and built-up areas, especially in the lower elevations. These findings highlight the value of long-term satellite records for detecting elevation-dependent forest changes and supporting sustainable forest management in the Himalayas.
Accurate identification of vegetation associations in coastal muddy tidal-flat wetlands is essential for understanding wetland ecosystem structure, function, and dynamics. However, spectral similarity and structural complexity limit conventional satellite-based mapping. Integrating UAV LiDAR with satellite imagery offers new opportunities for improving classification accuracy and enabling scalable wetland monitoring. In this study, we develop a UAV LiDAR-based point cloud classification method, CMFF-RandLA-Net, and construct a synergistic point cloud–satellite framework for multi-scale vegetation mapping. Multi-source features, including BRDF-corrected intensity, VDVI, and normalized scan angle, were incorporated to enhance feature representation, together with a feature fusion strategy to improve classification robustness. In the coastal wetlands of Yancheng, China, the proposed model achieved an overall accuracy (OA) of 95.53% in point cloud classification, with F1-scores above 0.80 for the dominant species Spartina alterniflora and Phragmites australis. The high-precision point cloud-derived ground truth was then used to train satellite classifiers for Sentinel-2 imagery. Among four evaluated algorithms (XGBoost, Random Forest, LightGBM, and Linear SVM), XGBoost performed best (OA = 92.85%, Kappa = 0.8381) and uniquely identified minority classes, enabling regional-scale vegetation mapping. The proposed framework bridges fine-scale UAV LiDAR classification and satellite-based regional mapping, providing a practical approach for coastal wetland monitoring.
Building generalisation transforms detailed large-scale building representations into suitable simplified forms for smaller-scale urban topographic maps. Existing deep learning-based approaches for raster building maps are largely built on convolutional neural networks (CNNs) or generative adversarial networks (GANs). However, their local inductive biases or limited global spatial awareness often cause geometric distortions in buildings. Although large-vision foundation models excel in holistic modeling, their application has focused mainly on building extraction. To address these issues, SAM-BG, a Building Generalisation framework that integrates the Segment Anything Model with multitask learning, was proposed. Specifically, low-rank adaptation was employed to efficiently fine-tune the vision transformer (ViT) encoder and transfer global representation capability to the cartographic domain. Subsequently, a geometry-aware multitask architecture with explicit boundary supervision was introduced to alleviate contour distortion. Moreover, a Boundary Feature Enhancement Module (BFEM) that uses dual-source-guided gating and multi-scale feature pyramids was designed to reinforce boundary responses. Multi-scale experiments on synthetic and real-world datasets showed that SAM-BG outperformed benchmark methods on boundary-sensitive metrics of buildings. This study presents a viable end-to-end solution for building generalisation and validates the potential of vision foundation models within domain-specific context of cartography.
Terrestrial ecosystems play a critical role in mitigating climate change by sequestering atmospheric carbon dioxide (CO₂), with aboveground carbon storage (AGC) serving as a key indicator of ecosystem carbon sink capacity. However, existing AGC estimation approaches remain constrained by the insufficient integration of natural and anthropogenic drivers, high update costs, and considerable uncertainties. In this study, we developed a prior-knowledge-constrained attention U-Net framework that incorporates land-use-specific carbon density information to improve the consistency and stability of AGC estimation. By integrating multiple environmental drivers, the framework effectively captures the nonlinear relationships between environmental factors and AGC, thereby enabling spatially continuous estimation of aboveground carbon storage across China's terrestrial ecosystems. The proposed model achieved an overall accuracy of 75.97% while significantly reducing estimation uncertainty. Based on this framework, a 1 km resolution AGC dataset for China was generated. The results indicate that forests are the dominant contributors to national AGC, with an annual average of 6.28 × 103 Tg C, accounting for 66.66% of the total AGC. Overall, this study provides an improved methodological framework for long-term AGC estimation and offers reliable data support for carbon sink assessment and carbon neutrality planning.
Accurate 3D forest reconstruction is essential for quantifying biophysical processes and ecosystem functions. However, the limited spatial coverage of LiDAR systems constrains the generation of structurally detailed forest scenes over large landscapes. This study introduces a scalable data-fusion framework to generate realistic 3D mountainous forest scenes, integrating the extensive coverage of unmanned aerial vehicle stereoscopic imagery (SI) with the high-fidelity vertical structure of unmanned laser scanning (ULS). ULS-derived tree templates were scaled and oriented using structural attributes from SI. To ensure structural realism, topographic variables (elevation, slope, aspect) were incorporated as strict constraints during template matching. The resulting point clouds drove a voxel-based 3D reconstruction validated across 500 m × 500 m coniferous and broadleaf stands. The simulated models demonstrated high structural congruence with benchmark ULS data. Furthermore, radiative transfer simulations based on these 3D scenes showed strong consistency with Sentinel-2 observations. In the near-infrared band, normalized absolute bidirectional reflectance factor differences dropped to 8.89% (RMSE: 0.0277) for coniferous and 8.28% (RMSE: 0.0392) for broadleaf scenes. Ultimately, fusing SI with topographically constrained LiDAR templates provides a highly transferable, geo-data-intensive approach for generating precise 3D forest representations, advancing regional ecological modeling and Earth science applications.
Artificial aquaculture ponds (AAPs) play a critical role in ensuring food security, supporting rural economic development, and shaping ecological landscapes. Yet, a nationally consistent, high-resolution, and individual-scale dataset of these ponds has been unavailable, hindering accurate assessments of their spatial distribution and temporal dynamics. In this study, we present CN-AAP10, the first 10-m resolution dataset of AAPs across China, delineating individual ponds from Sentinel-2 imagery. Using an improved ResUNet model applied to annual median composites of Sentinel-2 data for 2015, 2020, and 2025, we generated a dataset that captures pond-level features and their changes over a decade. Validation based on 3,049 randomly selected samples yielded an overall accuracy of 92.46%. Our results reveal a markedly uneven spatial distribution: eastern China accounts for 97.4% of total pond area, forming a continuous belt along coastal zones and major river plains. Temporally, the total pond area expanded from 17,280.59 km2 in 2015 to 18,968.79 km2 in 2020, followed by a modest decline to 18,110.83 km2 in 2025, indicating a shift from rapid expansion towards more regulated development. CN-AAP10 enables robust monitoring of aquaculture dynamics and offers a data foundation for sustainable aquaculture management, wetland conservation, and carbon emission assessments.
Remote sensing image super-resolution (RSISR) in real-world scenarios is challenged by spectral and geometric inconsistencies across sensors, caused by variations in sensor characteristics and acquisition conditions. These cross-sensor mismatches pose significant challenges for pixel-level supervision, and most existing methods overlook the valuable spatiotemporal metadata embedded in satellite imagery. To overcome these issues, a new metadata-guided framework named Spatiotemporal-Aligned Super-Resolution Framework (SASRF) is developed to incorporate spatiotemporal information into the reconstruction process. At its core, SASRF first leverages a Spatiotemporal Contrastive Learning Module (SCLM) to learn a meaningful alignment between image features and their corresponding spatiotemporal metadata. This learned context representation then guides a Dynamic Feature Alignment Module (DFAM), which performs metadata-aware feature modulation to mitigate spectral and geometric inconsistencies in cross-sensor imagery. The framework is optimized through a two-stage training strategy that decouples cross-modal representation learning from image reconstruction, thereby improving training stability and generalization. Comprehensive evaluations on a large-scale cross-sensor dataset with abundant spatiotemporal metadata verify that SASRF surpasses recent state-of-the-art techniques in reconstruction precision, spectral reliability, and geometric alignment. Moreover, its lightweight and modular design enables seamless integration with various backbone networks, making it a practical solution under real-world RSISR conditions. The code will be available.
The urbanization variability of several European cities is quantified in terms of fractal dimension [Formula: see text] and built-up density [Formula: see text] by using the WorldView (WV-2 European Cities and WV-ESA archive) multispectral satellite imagery collections at different spatial resolutions. The robustness and accuracy of the proposed approach are further assessed by analyzing Sentinel-2 L2A (S-2) data. The fractal dimension [Formula: see text] is estimated by implementing the Detrending Moving Average Algorithm (DMA). The built-up density [Formula: see text] is estimated by using the Masked Built-up Extraction Index (MBEI). The analysis yields values of [Formula: see text] and [Formula: see text] in the ranges [Formula: see text] and [Formula: see text], respectively, for dispersed and highly urbanized areas. The minimum volume ellipsoids (MVE) criterion is adopted to cluster the values of fractal dimension [Formula: see text] and built-up index [Formula: see text] and classify different urbanization areas.
Satellite solar-induced chlorophyll fluorescence (SIF) provides a direct proxy for vegetation photosynthesis, yet the short lifespan and degradation of satellite sensors limit long-term global SIF records. Here, we developed a hybrid modeling framework integrating light use efficiency (LUE) and machine learning to reconstruct a spatiotemporally continuous global long-term daily and monthly SIF dataset (LTSIF) at 0.05° resolution from 1981 to 2023. Key drivers were identified using a far-red SIF decomposition framework, LUE theory, and random forest (RF) modeling. Hybrid models combining LUE with RF, extreme gradient boosting, and neural networks were developed, among which the LUE-RF model achieved the best performance (R2 = 0.97, RMSE = 0.08 mW m−2 sr−1 nm−1, KGE = 0.97) and showed robust performance across vegetation types and climate zones. Independent validation against tower-based SIF, OCO-2 SIF, and widely reconstructed SIF datasets demonstrated that LTSIF reliably captured global photosynthetic dynamics. From 1981 to 2023, LTSIF revealed increasing photosynthetic activity over 89.01% of global vegetated areas, with a global mean trend of 0.016 mW m−2 sr−1 nm−1 per decade (p < 0.01). LTSIF provides a valuable long-term dataset for assessing terrestrial photosynthesis responses to climate change.
The growing use of geographic artificial intelligence (GeoAI) in volunteered geographic information is reshaping Digital Earth workflows through collaboration between human contributors and machine-learning models. OpenStreetMap is central to this ecosystem, yet evidence on how AI-assisted editors affect mapper behavior, data quality, and mapping performance remains limited. We evaluate fAIr, an open, locally trainable building-mapping environment developed by the Humanitarian OpenStreetMap Team. In a controlled experiment involving 26 participants, we compared fAIr with JOSM, a widely used editor for manual mapping. Across simple and complex study areas, we assessed mapping efficiency and effectiveness, classified error types, and examined participants’ adjustments to model parameters. Manual mapping in JOSM was faster and more accurate overall, mainly because experienced contributors performed well. fAIr reduced differences between novice and experienced contributors but produced AI-related errors, most commonly the merging of multiple buildings into a single footprint. Participants adjusted only some of fAIr’s parameters. Overall, fAIr offers a fast, accessible workflow for mapping simple structures, but its broader application depends on reducing recurring errors in building-footprint prediction. These findings informed the development of a new version of fAIr.
In the simultaneous prediction of soil properties, visible–near infrared (Vis–NIR) spectroscopy is widely used for rapid and non-destructive assessment of soil properties. However, traditional multi-task learning (MTL) approaches in such multivariate prediction tasks often show strong dependency on task correlation, limiting their robustness and generalization. To address this limitation, this study proposes the MSCA-CGC, a network that utilises Multi-Scale Convolution-Channel Attention within a Customised Gate Control framework. The model combines shared and task-specific experts to better capture local high-frequency spectral features. Each task uses an independent gating network to dynamically fuse shared and task-specific information, reducing the seesaw effect and negative transfer. Using the LUCAS 2009 dataset (19,032 samples, 11 soil properties), the model is benchmarked against PLSR, XGBoost, MMoE, and LSTM-CNN-Attention. Results show that MSCA-CGC outperforms the other models in prediction accuracy and stability, with average RMSE reduced by 3%–33%, [Formula: see text] improved by 3%–39%, and RPD increased by 3%–51%. In transfer tests on LUCAS 2015, MSCA-CGC demonstrated good generalisation for key properties, supporting its robustness and adaptability under spatiotemporal heterogeneity. This study confirms the effectiveness of MSCA-CGC for soil property prediction.
Maize occupies 12% of global grain production and plays a critical role in ensuring food security. However, most maize-mapping methods rely heavily on reference labels and repeated model training, limiting their utility in data-scarce heterogeneous agricultural regions. To overcome this limitation, a minimally sample-dependent dual-index constraint framework that directly exploits canopy-level spectral signals from Sentinel-2 imagery was proposed. Maize Chlorophyll and Water Content Index (MCWCI) was developed to capture spectral absorption associated with chlorophyll and water content, whereas Near-Infrared Difference Vegetation Index (NirDVI) was designed to enhance sensitivity to Nir-response differences between maize and other crops arising from canopy structure and integrated physiological characteristics. Cross-regional and cross-temporal assessments conducted in the Loess Plateau, Iowa, and Mato Grosso demonstrated the stability and transferability of the proposed framework. Results showed that the proposed framework achieved F1-scores ranging from 82.54% to 95.43% across all regions and improved F1-scores by 3.84%–41.80% compared with existing indices, including GWCCI, NDCI, DCNI, and TCARI. In addition, the framework shows better performance in delineating field boundary characterized by small and fragmented fields, providing a practical and transferable solution for large-scale maize mapping.
Monitoring fine-scale forest disturbances in karst regions is vital for ecological restoration but hindered by cloud cover, heterogeneity, and binary detection limits. We propose an integrated framework that couples spatio-temporal disturbance detection with multi-scale intensity analysis (SLAF). First, the Space-Time Extremes and Features (STEF) algorithm detected small-scale disturbances. This integrates the spatial-context normalization of the Landsat Normalized Difference Moisture Index (NDMI) time series (1986–2021) to suppress seasonal noise under data-sparse conditions, enabling the extraction of multi-level spatio-temporal features. Second, five ecologically interpretable disturbance-intensity levels were derived using Clustering Large Applications (CLARA) clustering by integrating detected disturbances with multi-scale landscape-pattern metrics—patch density and adjacency–which extend beyond binary classification. SLAF achieved 97.09% overall accuracy and a Kappa coefficient of 0.708, with a producer’s accuracy of 73.91%, a user’s accuracy of 70.83%, and an F1-score of 0.723 for the disturbed class. The identified peak year of forest-disturbance coincided well with key ecological policy shifts, indicating the reliability of SLAF. Moreover, SLAF reveals the gradient from isolated gaps to aggregated patches through multi-scale intensity analysis, providing an ecologically meaningful severity classification. This study offers a robust tool for translating satellite data into actionable management insights for fragile karst ecosystems.
Spatial optimization plays a pivotal role in urban analytics, underpinning critical decisions from facility siting to logistics routing. However, the practical application of these methods is often hindered by a fragmented software ecosystem: commercial GIS platforms offer usability but lack algorithmic flexibility, while specialized mathematical solvers provide computational power but require steep learning curves and extensive ‘glue code’ for data handling. To bridge this gap, we introduce HiSpot, an open-source Python framework designed as an integrated workbench for spatial optimization. Unlike disjointed tools, HiSpot unifies the entire workflow—data processing, model formulation, solution generation, and visualization—within a cohesive, object-oriented Application Programming Interface (API). The library supports a comprehensive suite of paradigms, including Facility Location Problems (FLP), Coverage Models (MCLP/LSCP), and Location-Routing Problems (LRP), seamlessly coupling them with standard geospatial data structures. Furthermore, HiSpot addresses the scarcity of standardized testing environments by incorporating the Spatial Optimization Benchmark Dataset, which includes multi-scale instances from real-world road networks. By lowering technical barriers and standardizing the analytical pipeline, HiSpot empowers researchers and practitioners to deploy robust, reproducible spatial optimization solutions for complex urban challenges.
Afforestation plays a crucial role in influencing global climate change through biophysical and biogeochemical processes. Age-related forest structure regulates the performance of forest biophysical feedback. This study investigated land surface temperature differences (ΔLST) between immature forests of varying age classes and mature forests using high-resolution remote sensing data, including MODIS and Landsat land surface temperature (LST) products, along with a 30-m forest age map. We further analyzed the complex relationships between ΔLST and its driving factors encompassing biophysical processes and vegetation condition. The study revealed that broadleaf forest (BF) presented a greater warming effect with (+1.63 °C ± 0.03 °C), followed by mixed forest (MF) (+1.43 °C ± 0.04 °C) and needleleaf forest (NF) (+1.41 °C ± 0.03 °C), than their mature counterparts. Forests with higher ΔLST across different age classes were predominantly concentrated along China's southwest-northeast forest belt. The albedo contributed up to 0.51 in the northern zone annually, and the influence of albedo and evapotranspiration on ΔLST gradually decreased as the forests aged. Spatially, the dominant drivers of ΔLST displayed significant regional heterogeneity, with their relative contributions showing non-linear relationships with forest age across climatic zones. Our findings provide empirical evidence of LST response to guide climate-smart afforestation strategies in China.
