Accurate extraction of city-scale road networks from remote sensing imagery remains challenging, as existing methods often suffer from topological fragmentation, directional ambiguity, and geometric distortion. This study proposes Intersection-Aware Orientation-Field-Guided Road Network Extraction (IAOFG-Road), a framework that integrates intersection perception, orientation-field constraints, and graph-based topological reasoning. A Swin Transformer backbone with a feature pyramid supports multi-task collaborative learning, in which semantic and structural cues are learned jointly through shared representations, avoiding the error propagation of sequential pipelines. During graph assembly, junction-anchored node linking, directional consistency checks, and segment completion jointly improve geometry and global connectivity. On public benchmarks, the method achieves the best TOPO-F1 on both datasets, the best TOPO-R on SpaceNet3 and the second-best TOPO-R on City-Scale (trailing only IS-RoadDet), together with competitive APLS and practical inference time.
Accurate soil moisture (SM) estimation is essential for precision agriculture; however, in southwestern China's karst mountains, terrain effects on SAR and persistent cloud contamination introduce uncertainty. We propose a parcel-constrained SAR–optical temporal reconstruction framework. Cultivated parcels are delineated from sub-meter imagery using semantic segmentation, and multi-source SAR, optical and environmental features are aggregated to reduce speckle and mixed-pixel effects. A SAR–Optical Transformer (SOT) reconstructs parcel-scale normalized difference vegetation index (NDVI) time series from SAR, which are integrated with multi-source features in a random forest for SM retrieval. The SOT achieved R2 = 0.854 and mean absolute error (MAE) = 0.065. Adding reconstructed NDVI to the SAR + environmental baseline (R2 = 0.721) increased SM R2 to 0.805 and reduced RMSE to 0.038 m3 m−3, with uncertainty decreasing by ~18%. Compared with pixel-scale retrieval (R2 = 0.762), the parcel constraint improved R2 to 0.805 and reduced uncertainty by ~29%.
Hurricane risk assessments under climate change are essential for resilient infrastructure planning, yet they remain constrained by limited and inconsistent historical damage records. This study develops a spatial framework for assessing storm surge-induced hurricane building damage under future climate scenarios. Ten thousand years of synthetic hurricane tracks and storm-surge simulations were generated and overlaid with building-level insurance data. A machine learning model trained on historical insurance losses was then applied to the simulated hurricane dataset to estimate building-level damage under RCP-4.5 and RCP-8.5 scenarios. The results indicate a spatial shift in projected hurricane damage from the western Florida coastline toward the eastern Atlantic coast. Although hurricane frequency is projected to decrease, storm intensity increases substantially, with Category 5 hurricanes rising by 58% under RCP-4.5 and 104% under RCP-8.5. Miami emerges as the highest-risk location. The proposed framework integrates synthetic hazard simulation and machine learning to support climate-resilient spatial risk assessment.
Under global warming, permafrost degradation in the Sanjiangyuan region has intensified freeze-thaw hazard risks. Deep-learning-based pixel-level extraction from high-resolution remote sensing imagery supports hazard mapping and ecological conservation. However, freeze-thaw hazards in complex alpine landscapes show irregular boundaries and severe background confusion, and single-network models cannot balance local details and global semantics. We propose a dual-backbone DeepLabV3+ integrating ResNet50 and Vision Transformer (ViT) in parallel. Shallow Convolutional Neural Network (CNN) features provide boundary-texture information, whereas deep Transformer features capture long-range context; the decoder fuses multi-scale semantics. Compared with DeepLabV3+ (MobileNetV2), PSPNet, TransUNet, SegNet, and HRNet, our model achieves a mean Intersection over Union (mIoU) of 97.85%, an Intersection over Union (IoU) of 96.44%, and an Overall Accuracy (OA) of 99.38%. Cross-region testing obtains an mIoU of 88.39%. Ablation verifies 4.52%–12.95% IoU gains over single-backbone schemes, demonstrating local-global feature fusion improves boundary continuity and object integrity for alpine freeze-thaw hazard extraction.
Accurate mapping of leaf chlorophyll content (Cab) is critical for monitoring plant growth and maximising crop yields in sustainable precision farming. This study developed a nonparametric, hybrid retrieval method to map winter wheat (Triticum aestivum) Cab from EnMAP spaceborne hyperspectral data, using in situ data collected in southern Hungary. A simulated training dataset was optimised using principal component analysis, active learning (AL), and spectral and sampling reduction, then used to train a Gaussian process regression (GPR) model. Validation yielded R² = 0.76 and RMSE = 8.22 μg/cm². The best results were achieved after adding 101 AL samples to 50 starting samples (NRMSE = 15.72%). Adding non-vegetated (NV) spectra slightly reduced validation accuracy (NRMSE = 16.93%, R² = 0.74) but produced more consistent models able to handle spectral variability across the heterogeneous EnMAP scene. GPR also provides uncertainty estimates alongside trait retrievals, indicating model reliability.
