Soil moisture (SM) assessment underpins applications ranging from climatological studies to land management and emergency response. Microwave techniques already support operational SM products, but at a spatial resolution limited to 1 km. Optical multispectral indices have been proposed as a complementary source of information at finer resolutions (10–100 m), yet vegetation heterogeneity has generally been handled by masking or stratifying the land cover rather than by modeling its effect. Here, we compare three indices (the Normalized Difference Infrared Index, NDII; the Normalized Multi-band Drought Index, NMDI; and the Visible and Shortwave Infrared Drought Index, VSDI) over a heterogeneous agricultural area near Grosseto, Tuscany, using 31 Landsat-8 OLI images acquired over four summers. Evaluated against the Copernicus Soil Water Index (SWI), VSDI performed best both when averaged over the test area and when analyzed at the pixel level (1 km) and was therefore selected for further analysis. A linear mixed-effects model with pixel-dependent intercepts resolves a within-pixel VSDI–SWI repeated measures correlation of 0.714, against 0.411 for the pooled regression. Allowing the slope to vary as well shows that vegetation density governs both the baseline level of VSDI and its sensitivity to soil moisture, with the within-pixel correlation falling from 0.84 on bare soil to 0.30 under a dense canopy.
Coded aperture snapshot spectral compressive imaging (CASSI) enables the recovery of hyperspectral images (HSIs) from a single two-dimensional compressive measurement, serving as a core paradigm in modern computational imaging. However, HSI reconstruction from under-sampled measurements is a severely ill-posed inverse problem. Existing deep unfolding methods typically treat each iterative stage independently and model spatial–spectral features in a decoupled manner, which neglects intrinsic spatial–spectral coupling and lacks persistent cross-stage information propagation. To tackle these issues, we propose a deep dynamic 3D memory-augmented unfolding network (3DMAUNet) for HSI reconstruction. Based on the momentum-accelerated generalized alternating projection framework, 3DMAUNet retains the strict mathematical interpretability of iterative optimization while integrating data-driven learnable components. Specifically, we design two key modules: (1) a 3D memory-augmented learning module equipped with a persistent memory bank, which explicitly encodes cross-layer spatial–spectral correlations to enable efficient inter-stage information transfer and long-range dependency modeling; and (2) a learnable dynamic fast momentum updating module, which fuses cross-stage spatial–spectral attention features to estimate task-adaptive momentum coefficients, accelerating convergence and enhancing the robustness of feature updating. Comprehensive experiments on synthetic and real-world HSI datasets demonstrate that 3DMAUNet outperforms state-of-the-art methods in both quantitative and qualitative reconstruction performance.
The rapid proliferation of high-resolution remote sensing video (RSV) has made object segmentation an increasingly critical yet challenging task. Existing methods face two fundamental limitations in this domain: (1) the intrinsic complexity of RSVs arising from diminutive object sizes, dense distributions, and drastic scale variations; (2) prohibitive computational overhead and latency caused by the massive spatial resolution of remote sensing imagery. To address these challenges, we propose Trajectory-SAM, a novel object trajectory-guided segmentation framework tailored for rapid and accurate object segmentation in RSVs. Specifically, we first employ a Graph Neural Network to extract representative keyframes, improving data utilization efficiency while suppressing attention drift and interpretation bias induced by redundant frames. We then introduce Temporal-LoRA, a parameter-efficient fine-tuning strategy for the Segment Anything Model (SAM), substantially enhancing its spatiotemporal contextual understanding in remote sensing domains. Finally, we devise a sample-augmentation-based knowledge distillation strategy to accelerate inference without sacrificing segmentation quality. Extensive experiments on the SAT-MTB, iSAID, and VHR-10 datasets demonstrate that our proposed Trajectory-SAM achieves superior performance, establishing a compelling trade-off between accuracy and speed.
Hyperspectral image classification requires joint modeling of spectral signatures and spatial structure. In patch-based classification, each input patch is assigned the label of its center pixel, although neighboring pixels may belong to different land-cover classes. This setting motivates coordinating shallow spatial features, contextual refinement within the patch, and supervision of the resulting representation. We propose DSFormer++, an enhanced framework based on the original Dual Selective Fusion Transformer Network (DSFormer). Shallow Feature Compensation (SFC) combines shallow and patch-embedded features through gated residual fusion before the backbone. The Collaborative Global–Local (CGL) module applies an additional refinement operation with shared parameters before token selection at two backbone stages. Pixel Contrastive Loss (PCL) uses center-pixel labels to supervise projected patch embeddings during training. The proposed method achieves overall accuracies of 96.50%, 97.60%, 95.35%, and 94.83% on the Pavia University, Houston13, Indian Pines, and WHU-HH datasets, respectively. Component comparisons and approximately parameter-matched controls characterize configuration-dependent effects. Numerical checks assess consistency between the stated operations and their implementation, and an additional Botswana experiment extends the evaluation to a satellite wetland and woodland scene.
To address the challenges of reliance on manual expertise and insufficient automation and intelligence in settlement generalization during scale transformation, this paper proposes a multi-scale settlement generalization model (DSG-Trans) based on the fusion of dual-source data: raster maps and remote sensing images. Leveraging the cartographic correspondence between remote sensing image resolution and map scale, this model integrates source-scale maps with specific low-resolution remote sensing imagery to automatically generate target-scale raster maps that adhere to cartographic specifications. The model employs a bidirectional cross-attention mechanism and a Transformer architecture to facilitate the semantic integration of dual-source data, while a progressive upsampling strategy is utilized to reconstruct spatial details. Additionally, a composite loss function incorporating four constraints is designed to enhance the quality of settlement generalization. Experiments were conducted in New York City, USA, using a multi-source dataset constructed from 3D-GloBFP building footprint data and multi-source remote sensing imagery, focusing on two scale transformation scenarios: 1:5000 to 1:10,000 and 1:5000 to 1:25,000. The results indicate that the proposed method effectively learns and simulates expert generalization knowledge, significantly improving the automation level and scale adaptability of map generalization while preserving geometric morphology and spatial distribution patterns.
