Global navigation satellite system reflectometry (GNSS-R) is widely used for soil moisture retrieval. To mitigate heterogeneity between delay–doppler map (DDM) and auxiliary features and to impose physical constraints in fusion modeling, we propose a physics-guided cross-feature fusion network (PCF-Net) for the Yellow River Delta, comprising a multi-feature input, cross-feature fusion, and a physics-constrained retrieval module. Specifically, the multi-feature input module employs convolutional neural networks (CNNs) to extract spatial–local features from DDM and deep semantic information from auxiliary data, enabling structural alignment and unified embedding of heterogeneous modalities. The cross-feature fusion module adopts mamba-based cross-feature fusion (MCF) block and transformer-based cross-feature fusion (TCF) block to achieve bidirectional interaction and deep coupling between one-dimensional auxiliary features and DDM image features. The physics-constrained retrieval module constructs a physics-consistency loss between measured and simulated GNSS-R surface reflectivity to regularize the training process. We further construct a multi-source GNSS-R dataset from cyclone global navigation satellite system (CYGNSS), Tianmu-1, and fengyun-3E (FY-3E) and validate against soil moisture active passive (SMAP). The joint use of multi-source GNSS-R data markedly improves retrieval accuracy, achieving R of 0.9805 and RMSE of 0.0176 cm³/cm³, demonstrating effectiveness at the regional scale.

Multi-source data fusion offers great potential for land cover classification. However, the substantial differences in data structures and content representations across various remote sensing sources present significant challenges in heterogeneous feature extraction and information fusion, ultimately constraining the effectiveness of fusion-based classification. To address the aforementioned limitations, this article proposes a multi-source data fusion classification method based on a cross-modal cascaded encoder-decoder network (CCEnd-Net). The proposed algorithm comprises two primary components: a blur feature extraction module and a cascaded encoder-decoder feature fusion module. Specifically, the blur feature extraction module utilizes multi-dimensional convolution to extract deep features from multi-source data and incorporates a blur pooling module to enhance aliasing resistance. This approach mitigates the original spatial discrepancies among heterogeneous data while preventing feature distortion. Meanwhile, the cascaded encoder-decoder feature fusion module reconstructs multi-source data features by integrating a multi-scale channel-spatial interaction attention (MCIA)-enhanced convolutional neural networks (CNNs) with a transformer-based cross-modal fusion (TCMF) block. Additionally, a multi-level fusion strategy is employed to comprehensively exploit the complementary information from different remote sensing data sources, thereby improving classification performance. To validate the effectiveness of the proposed method, comprehensive experiments were conducted on four benchmark datasets covering light detection and ranging (LiDAR), hyperspectral imaging (HSI), synthetic aperture radar (SAR), and very high resolution (VHR) data. The proposed CCEnd-Net was systematically evaluated against state-of-the-art models, including transformer-based architectures, CNNs, and traditional classifiers. The experimental results demonstrate that CCEnd-Net achieves superior performance across all datasets.

Synthetic Aperture Radar (SAR) imagery is indispensable for rapid, weather-independent marine oil spill monitoring. Yet the acute shortage of annotated SAR spill imagery severely limits deep learning detectors. While Generative Adversarial Networks (GANs) have been used to generate synthetic data, their inherent limitations—training instability and mode collapse—often result in blurred and semantically inconsistent outputs. To overcome these challenges, we introduce a Classification-Guided Diffusion Model (CG-DM), which integrates the expressive power of diffusion processes with task-specific guidance. CG-DM incorporates three key innovations: (i) Morphology-Aware Classification Guidance: SAR oil spill images are categorized into different morphological categories (blocky, elongated and patchy). This category information conditions every step of the reverse-diffusion process, enabling fine-grained control over the global geometry of generated spills while preserving intra-category diversity. (ii) Label-Synchronized Generation: The model simultaneously generates the SAR image and its corresponding pixel-level annotation mask, which eliminates the need for the time-consuming and error-prone manual labeling process. (iii) Spatial-Aware Attention Mechanism: A lightweight attention mechanism performs localized window self-attention with relative positional offsets, which significantly enhances the sharpness of spill edges and the fidelity of speckle-textured details. Evaluated on the M4D benchmark (ITI/EMSA’s semantic segmentation dataset with Sentinel-1 SAR imagery from 2015–2017), CG-DM achieves a Fréchet Inception Distance (FID) of 203.14 and a Kernel Inception Distance (KID) of 0.124, surpassing state-of-the-art GAN baselines by substantial margins. Besides, ablation studies confirm the critical contributions of each component: spatial-aware attention mechanism significantly enhancing generation quality, while classification guidance effectively preserving morphological feature diversity. Crucially, under extreme data scarcity, training augmentation with CG-DM synthetic samples improves oil-spill detection IoU by up to 81%, demonstrating strong practical utility. This study establishes the first annotation-free paradigm for SAR oil-spill data generation, paving the way for high-accuracy maritime disaster monitoring systems.

Crop field edge is the key characteristic of agricultural crop management. Edge detection with dual-polarization SAR time series have been widely studied, with the data advantages of sensitivity to crop growth. Existing methods rarely consider time-varying dynamic polarimetric characteristics, making it difficult to detect complete crop field edges. Based on this, this proposes a joint edge strength, which combines two kinds of polarimetric distances with a novel spatial-temporal homogeneity measure. This measure applies spatial- and temporal-varying contexts to pre-identify edge and homogenous area, and adaptatively allocate various distances in one pixel. There are 14 Sentinel-1 SAR time series images are utilized to evaluate the proposed method. By the comparison of the visual differences, our method has lower missing rate and lower false alarm rate than conventional methods.

With the development of the polarimetric statistical measures, change detection algorithms have been widely applied in the flood disaster detection. The representative methods utilize the Wishart distance to generate the difference map between the pre- and the post-disaster data. Furthermore, the difference map is applied into the threshold segmentation to generate the binary result. However, the existing methods rarely consider the impact of rainfall on the polarimetric distance, which causes abnormal polarimetric differences and incorrect detection results. Based on the fact of different sensitivities of the co- and the cross-polarization intensity for the rainfall, this paper uses the cross ratio of two polarization intensities to eliminate the impacts of the rainfall events on the polarimetric distances. Then the Markov Random Field (MRF) model is utilized to detect the flood disaster with the initialization of OTSU threshold segmentation. Two Sentinel-1 SAR data in the Poyang Lake Basin is used to assess the effectiveness of our method. The results have demonstrated the superiority of our method over the conventional methods, especially in the rainy areas.

Precision agriculture management relies on the delineation of crop field edges. Multi-polarization SAR technology has the ability to penetrate clouds and capture morphological structures or moistures, suited for extracting crop field edges. Due to the time-dependent characteristics and phenological evolutions of crops, the methods with single-date data are difficult to detect complete edges. Moreover, the existing methods fail to extract the dynamic time-varying patterns, limiting the improvement of edge detection accuracy. Based on this, this paper proposes a novel crop field edge detection method based on the time-varying polarimetric characteristics. First, a spatial-temporal homogeneity measure is proposed to pre-identify the edge and homogenous area, for guiding the adaptive calculation of edge strength. Based on the time-series polarimetric stationarity and the trace moment estimation theory, the proposed measure enlarges the separating degree of various crop parcels. Second, a joint edge strength is proposed to enlarge strength contrast between edge and homogenous area. With the spatial-temporal homogeneity measure, it combines the similarity with the root mean square and the similarity with time-series average covariance matrix. Based on the advantages of two kinds of similarities, it highlights the field edges and reduces the impact of speckle noises. Evaluated by 8 quad-polarization and 14 dual-polarization SAR images, the proposed edge detection method achieves better visual presentations and detection accuracies than traditional methods. With the statistics of the signal-noise ratio (SNR), the joint edge strength also has higher strength contrast than conventional strengths. The relevant codes can be found in https://github.com/DawnHanGeo/TSPolEdge.git.

Due to the high operational complexity of manual sample labeling for hyperspectral images (HSIs), few-shot learning (FSL) has been introduced to cope with the lack of training samples in the field of HSI classification and achieved good results by virtue of its excellent performance. However, factors such as category labeling noise, category distinguishability in the metric space, and completeness of effective feature mining are still the primary considerations that significantly affect the stability and robustness of FSL. Therefore, this study proposes a framework of FSL with label smoothing and metric space optimization (LMFSL) for HSI classification. The framework first incorporates a category label smoothing strategy into FSL, which mitigates the effect of noise by constructing a novel category label smoothing module (CLSM), to reduce the confidence of the classifier. Meanwhile, the study also designs a metric space optimization module (MSOM), which prompts similar samples in the metric space to largely aggregate together by maximizing the intraclass similarity and minimizing the interclass similarity in a more flexible way, so as to improve the decision boundary of the model and enhance the recognition performance of the model. Furthermore, to achieve effective feature extraction in the context of a few labeled samples, a lightweight feature extractor, LFE-MAs, incorporating multiple attention mechanisms is designed to extract the HSI features with high efficiency and low computational cost. Experimental results on four public HSI datasets show that LMFSL outperforms other state-of-the-art methods in terms of classification accuracy with limited labeled samples and has lower computational complexity.

Deep learning methods have been proved outperforming the traditional methods in the field of hyperspectral image classification (HSIC). However, in pursuit of higher accuracy, HSIC networks have become deeper and more complex, resulting in excessive parameters and computational cost. To deploy neural networks on small platforms such as mobile or embedded devices, many studies have focused on lightweight HSIC networks. Currently, these researches are dominated by patch-based networks, which suffer from the low accuracy caused by lightweight scale and slow inference speed derived from structural deficiencies of such networks. It is worth noting that full convolutional networks are able to achieve fast inference, but they tend to consume massive memory. To this end, this paper proposes a novel lightweight HSIC method, which consists of a successive spatial rectified network (SSRNet) and a non-central positional sampling (NCPS) strategy. SSRNet is composed of a local channel attention based spectral full convolutional network and several separable atrous spatial pyramid modules. These shallow sub-networks are concatenated together to progressively optimize their outputs by successive spatial rectified learning. For decreasing memory access cost, SSRNet makes little patches as input to perform patch-wise pixels-to-pixels learning. After training, SSRNet is able to adapt to any size of hyperspectral images and complete fast inference of the full image directly. In particular, the NCPS sampling strategy enables all labeled pixels to equally traverse all spatial positions of each training patch through the positional shift sampling, which effectively alleviates the sparse problem of hyperspectral semantic labels. Experiments upon three public benchmark datasets indicate that SSRNet is comparable to the state-of-the-art methods in classification accuracy with less than 0.15M parameters and only occupy less than 10MB memory for single forward computation. Moreover, SSRNet behaves significantly superior to the traditional patch-based networks in term of the inference speed. The source codes can be available from the website of https://github.com/Pancakerr/HSIC-platform.

Polarimetric Synthetic Aperture Radar (PolSAR) can record the backscattering information of ground objects by capturing the phase difference of echo signal in different polarization combinations. Due to the characteristics of coherent imaging, there is a lot of coherent speckle noise in PolSAR images. However, most of the current PolSAR image despeckling algorithms are difficult to balance the ability in aspects of speckle suppression, polarization information retention and edge structure preservation. This paper proposes an adaptive PolSAR trilateral filter (PTF) based on the mechanism of scattering consistency, which consists of following three steps. (1) The marker map of ground object scattering category is firstly obtained based on the dominant scattering mechanism determined by Freeman decomposition of the polarization covariance matrix. (2) Next, the neighborhood pixels with the same scattering mechanism as the center pixel are screened out within an adaptive filtering window. (3) A PolSAR trilateral filter is established through integration of the spatial proximity, polarization similarity and local structural characteristics of the center pixel and the screened pixels to effectively suppress the speckle noises. Experimental results upon the data of Radarsat-2 and AIRSAR demonstrate that PTF can not only suppress coherent speckle noise effectively without sacrificing the edge structure information, but also retain the unique polarization information of PolSAR image. In particular, compared with the traditional six widely used PolSAR despeckling algorithms, PTF have been significantly improved in terms of Equivalent Number of Looks (ENL) and Degree of difference (Dod) indexes by 2.49 times and 6 orders of magnitude, respectively.

This letter presents a semi-automatic approach to delineating road networks from very high resolution satellite images. The proposed method consists of three main steps. First, the geodesic method is used to extract the initial road segments that link the road seed points prescribed in advance by users. Next, a road probability map is produced based on these coarse road segments and a further direct thresholding operation separates the image into two classes of surfaces: the road and nonroad classes. Using the road class image, a kernel density estimation map is generated, upon which the geodesic method is used once again to link the foregoing road seed points. Experiments demonstrate that this proposed method can extract smooth correct road centerlines.