Multisource remote-sensing fusion benefits port and near-shore ship surveillance: optical images provide rich textures in daylight, while SAR remains informative in day-and-night and all-weather conditions. Yet high-resolution SAR–optical ship detection is challenged by scarce near-simultaneous acquisitions, residual misalignment after co-registration, and condition-dependent modality degradation. We curate MSOSD-1.0, a high-resolution paired SAR–optical dataset for port/near-shore ship detection, containing 1,715 co-registered patch pairs with 17,915 oriented ship instances and diverse conditions (low illumination, clouds/fog, cluttered backgrounds). On this basis, we propose WMoE-Net, a correspondence-constrained and reliability-aware alignment–gating–fusion framework. A wavelet-directed cross-attention module (WCAM) enhances local correspondence under residual mismatch. A condition-informed quality gating network (CIQGN) performs sparse expert routing among optical-preferred, SAR-preferred, and collaborative-fusion branches, while a mutually-guided consistency-aware fusion expert (MCFM) refines the collaborative branch to enhance fusion quality. Experiments on MSOSD-1.0 and QXS-SAROPT-SHIP show consistent gains over strong single-modal and multimodal baselines, especially under low illumination and clouds/fog.
High-resolution bathymetric digital elevation models (DEMs) support marine geoscience, but global grids can smooth fine-scale terrain where observations are sparse. We present GGFA-SwinIR (Gradient-Guided Frequency-Aware SwinIR), a shared-backbone discrete multi-scale framework for ×2, ×4, and ×8 reference-grid reconstruction. The framework integrates gradient-magnitude guidance, bicubic-base residual prediction, and pixel, gradient, and complex-spectrum supervision. Under bicubic degradation of GEBCO_2025, GGFA-SwinIR achieved H-RMSE values of 5.13, 9.82, and 18.90 m and PSNR values of 70.28, 65.48, and 59.84 dB at ×2, ×4, and ×8, respectively. In a four-severity low-pass diagnostic, GGFA-SwinIR achieved the lowest H-RMSE and highest PSNR among six methods in all 12 severity-scale conditions and the lowest slope-angle RMSE in 11. Cross-product transfer comparisons with GMRT v4.3 showed the lowest H-RMSE across two evaluation regions and three scales. These results demonstrate the value of jointly modelling terrain gradients and spectral structure for multi-scale bathymetric reference-grid reconstruction and provide an empirically validated basis for exploratory grid enhancement.
Selective logging (SL) is a major driver of forest degradation, yet its small and short-lived canopy openings remain difficult to detect in persistently cloudy tropical regions. Optical approaches miss many events, while synthetic aperture radar (SAR) time series are challenging to model with architectures that either fuse multivariate features too early or smooth over transient signals. We present InterFormer, a Transformer-based architecture that jointly models feature interactions and temporal dynamics via a feature block with Multi-head Intercross Self Attention (MISA) and a temporal block with Multi-head Self Dissimilar Attention (MSDA). Texture features were further utilized as complementary input. Since precise SL labels are unavailable, the model was trained and internally tested using a weak-label benchmark, in which pixel labels were extracted within the official published Real-Time Deforestation Detection System (DETER) polygons. To complement this internal weak-label benchmark, we designed two independent practical experiments to evaluate the model performance in low-intensity Annual Production Units (UPAs) based on field-inventory data and at the municipality scale (> 30,000 km2) in the Brazilian Amazon. In the independent UPA validation, InterFormer localized 92.65% of tree-falling gaps in UPA areas and outperformed standard Transformer variants and Planet-based alerts from the Brazilian Forest Service by approximately 20% at the tree level. At the municipality level, our predicted alert polygons achieved the highest Detection Rate against DETER across multiple reference-overlap thresholds in both Paragominas and Feliz Natal, while also producing fewer unmatched alerts than RAdar for Detecting Deforestation (RADD) and Tropical Moist Forest dataset (TMF). However, substantial over-expansion of predicted alert regions remained. These results suggest that InterFormer provides a useful framework for representing weak and short-lived SL signals in Sentinel-1 time series, and that openly available SAR data have the potential to support large-scale SL monitoring under weak supervision. Nevertheless, the current municipality-scale workflow still relies on optical-derived masking as a supporting step, and the development of a fully SAR-only large-area monitoring framework remains an important direction for future work.
Hyperspectral and LiDAR images, as a class of multi-source remote sensing data, have their own advantageous features. HSI contains rich spectral information, compared to LiDAR, which is more capable of reflecting the height information of the target area. However, determining the intrinsic correlation characteristics between HSI channels and fully exploiting the complementary characteristics of HSI and LiDAR to improve classification accuracy remains a challenging issue. In this paper, we propose a graph-interactive collaborative learning network (GICL-Net) for joint classification of HSI and LiDAR, which consists of a multi-scale spectral-spatial graph feature extractor (MSGE) and a cross-modality adjacency matrix fusion (CAMF) module. Specifically, MSGE is divided into two branches. One branch starts from the spectral dimension, defining each channel of HSI as a node to construct a graph network and mine potential correlations at the channel level, thereby capturing the intrinsic spectral correlation features of the target region. The other branch focuses on the height information of the target area, constructing a graph network at the geometric spatial level to extract representative height features. The CAMF module leverages the fused adjacency matrix to establish a cross-modal graph interaction mechanism, enabling the spectral features and the spatial features to achieve deep synergy through iterative updates in graph convolution. This generates cross-modal fusion features with fine spectral discrimination capability and spatial dimensional structural integrity. This method has been validated for effectiveness across multiple HSI and LiDAR datasets. In particular, on the Houston2013 dataset, the overall accuracy (OA) reached 98.91%, with a Kappa coefficient of 0.9882. Furthermore, compared with the baseline network (i.e., the conventional two-branch graph network, the overall accuracy improved by 6.58%. Our code is available at the link: https://github.com/Xidian-AIGroup190726/GICLNet.
Large-scale vegetation greening has been widely reported across India; however, national assessments have largely relied on monotonic trend metrics without explicitly evaluating seasonal variability, spatial heterogeneity, or the structural persistence of vegetation dynamics. This study presents a physiography-resolved, multi-season assessment of vegetation change using a 24-year (2001–2024) MODIS 16-day Enhanced Vegetation Index (EVI; 250 m) dataset spanning 14 physiographic subdivisions. Seasonal Mann–Kendall analysis and Sen’s slope estimation reveal widespread but spatially heterogeneous greening, with all physiographic subdivisions exhibiting significant positive winter trends and the strongest greening rates (0.0032–0.0051 yr⁻1) over the Central Highlands, Northern Deccan, and Western Coastal Plains. Pixel-scale analysis indicates significant greening over 62.47 % of India during the pre-monsoon and 61.46 % during winter, declining to 34.57 % during the monsoon before partially recovering (∼43 %) during the post-monsoon, whereas browning remains limited (<2.2 %) and spatially restricted. Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) further reveals that vegetation dynamics across most physiographic subdivisions are dominated by sub-interannual (6–24 month) variability, indicating that long-term greening is embedded within persistent low-frequency oscillations rather than short-term fluctuations. In contrast, high-frequency variability is concentrated over climatically extreme and hydroclimatically stressed regions. Seasonal Intrinsic Mode Function 3 (IMF3) energy anomalies identify monsoon-concentrated, off-monsoon-amplified, and mixed variability regimes, while a persistence-based stability metric demonstrates that statistically significant greening does not necessarily imply structural resilience. The Western Plain exhibits the highest structural persistence, whereas plateau, mountain, and humid eastern regions remain dominated by temporal instability despite positive greening trends. By integrating seasonal trend analysis, frequency-resolved decomposition, physiographic stratification, and land-use differentiation, this study provides a scalable Earth observation framework for distinguishing persistent vegetation regime from variability-dominated change across heterogeneous landscapes.
