Deep learning has advanced ship instance segmentation in synthetic aperture radar (SAR) imagery. However, segmentation performance remains limited by multi-scale ship variations, inaccurate small-target segmentation, and significant background interference in complex scenes. To address these challenges, we propose a novel dynamic attention and context fusion network (DACF-Net), which enhances mask prediction through geometry-aware multi-scale progressive feature refinement. DACF-Net integrates multi-scale context aggregation, cross-level semantic guidance, and geometry-aware proposal generation to improve ship representation and segmentation accuracy. Specifically, a multi-scale feature fusion module (MSFFM) is introduced to aggregate scale-adaptive contextual information through a multi-branch large-kernel convolution structure. A cross-level feature interaction module (CLFIM) modulates low-level features with high-level semantics to optimize cross-level semantic consistency. A dynamic context perception block (DCPBlock) jointly models mid-range spatial dependencies and global channel-wise context to suppress background interference. In addition, three variants are developed to balance segmentation accuracy and computational efficiency. Experiments on HRSID and PSeg-SSDD demonstrate the effectiveness of the proposed approach. DACF-Net‡ achieves segmentation AP values of 63.3% and 65.7% on HRSID and PSeg-SSDD, respectively, outperforming the second-best method by 4.2% and 1.4%. Further experimental results show that DACF-Net and its variants achieve competitive segmentation performance for ship targets across different scales.
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.
