Hao Cui

The geometric precision of building segmentation is frequently bottlenecked not by the feature extraction capabilities of models, but by the crude post-processing step of binarizing their uncertain confidence map outputs. This forced decision-making on ambiguous probabilities is the primary cause of boundary degradation and structural distortion. This paper introduces a paradigm shift, moving from direct result binarization to a learned optimization of the underlying confidence distribution itself. We propose the Confidence Distribution Optimization Network (CDO-Net), a novel post-processing framework designed to learn a transport-inspired mapping. It transforms the uncertain source distribution from any segmentation model into a sharp, geometrically-regular target distribution that mirrors an ideal building mask. To achieve this, CDO-Net integrates a dual-branch CNN-Transformer architecture, which captures both fine-grained local uncertainties and global structural context. This transformation is driven by a multi-faceted objective function that synergistically enforces pixel-level fidelity, distributional realism via adversarial learning, and structural coherence through geometric priors. As a universal plug-and-play module, CDO-Net provides a powerful solution for refining the outputs of diverse segmentation models. Extensive experiments on the Inria, Massachusetts, and WHU datasets verify the general effectiveness of the proposed module. When integrated with DeepLabv3, CDO-Net achieves a relative improvement of up to 5.67% in Boundary F1-score compared with the second-best model on the Inria dataset, while consistently enhancing overall metrics such as IoU, thereby substantiating the superiority of the proposed distribution optimization paradigm.

Deep learning has made significant progress in the semantic segmentation of remote sensing images. However, its great performance heavily depends on high-quality annotations, which greatly restrict further advancement. Recently, the Segment Anything Model (SAM), enabled by large-scale pre-training and flexible prompt mechanisms, has shown exceptional zero-shot segmentation ability for any object. Hence, we devise RSPoint-SAM, a SAM-based point-supervised semantic segmentation model that can achieve excellent results using point annotations. RSPoint-SAM has two main branches: convolutional neural network (CNN) and SAM-driven fusion. The CNN branch uses EfficientNet as its feature extraction backbone and employs a cross-level feature fusion strategy to combine multi-level features, effectively capturing local contextual dependencies under sparse supervision. Next, high-confidence pseudo-labels are produced by combining the first two outputs of the CNN branch with a credible assessment mechanism. The SAM-driven fusion branch then integrates EfficientViT and EfficientNet to utilize the ability of both global context modeling of the vision transformer and the structured boundary perception of SAM, thereby capturing object-level semantics under limited pseudo-label supervision. Moreover, we develop a knowledge dynamic interaction mechanism to feed global semantic and boundary information from the SAM-driven fusion branch back into the CNN branch, reducing overfitting and further reinforcing pseudo-label quality. Experiments on two benchmark datasets show that RSPoint-SAM outperforms state-of-the-art weakly supervised methods. The source code will be available at: https://github.com/Yujia73/RSPoint-SAM.

Water bodies are typical natural features that exhibit significant variations in scale and appearance in high-resolution remote sensing imagery due to the influence of factors such as topography, temperature, sediment content, and aquatic vegetation. Existing deep learning-based water body segmentation networks typically adopt a local sliding-window approach for training and inference. However, this strategy limits the receptive field and hampers the model’s ability to utilize contextual information for identifying large-scale, complex-textured water bodies, often resulting in misclassification. To address these limitations, we propose a scale-progressive prompt water body extraction framework (SPPrompt-Water), inspired by the hierarchical “global-to-local” cognitive mechanisms of human vision. The framework employs progressive prompting to facilitate cross-scale contextual knowledge transfer, thereby enabling high-precision water body extraction. To support this framework, we design an enforced prompt-based parallel simulation training strategy, which leverages cross-scale sample simulation and label decomposition to obtain object-level supervision within the same scene. Through parallel contrastive training, the model learns to perform segmentation with enforced prompt guidance. Extensive experiments conducted on the GID, Complex Water Bodies Segmentation Dataset (CWBSD) and GLH-Water datasets demonstrate that the proposed method effectively balances spatial continuity in large-scale water bodies with boundary precision in small-scale ones. Our method achieves mean Intersection over Union (mIoU) scores of 83.22%, 92.27%, and 86.91% on the respective datasets, outperforming existing approaches by 8.1%, 6.65%, and 6.79%. These results validate the effectiveness of the proposed framework and offer a novel solution for high-precision dynamic monitoring of water bodies using remote sensing imagery. The source code and datasets will be made publicly available at https://github.com/XinMe/SPPrompt-Water

Accurate crop identification is crucial for agricultural trade, market risk management, and food security. Current research on automatic interannual sample extraction for crop mapping often emphasizes multi-source feature fusion but overlooks the importance of feature distance differences in crop seeding processes. This study focuses on crop mapping in the Hetao Plain from 2020 to 2023, using high-resolution Sentinel-1 and Sentinel-2 remote sensing data. We introduce a method called BGSI-DFF-WMRF, which combines the Bidirectional Global Selection Index (BGSI) and Distance Fusion Feature (DFF) under the interannual Weaving net Month Probability Random Forest (WMRF). BGSI captures the coupling between phenological and spatial factors necessary for crop growth in different regions under feature fusion. Additionally, sample migration under feature fusion enhances the accuracy and representativeness of sample points across different years. WMRF integrates a monthly classifier with multi-source feature distance interpolation. The BGSI-DFF-WMRF method achieved over 82% classification accuracy for crops like wheat, corn, and sunflower in the HID region, with an accuracy of 91.81% in the western area. Field samples from 2023 were successfully applied to previous years (2020-2022) through feature fusion expression (DFF) transfer. The method outperformed existing products and local statistical data, particularly for corn, demonstrating high accuracy and robustness in crop mapping. Coupling spatio-temporal factors enhances large-scale crop identification and holds great significance for the advancement and widespread adoption of large-scale crop identification techniques.

Although significant advances have been made in the semantic segmentation of high-resolution remote sensing images, obtaining accurate pixel-wise annotations remains resource-intensive. We propose SparseFormer, a credible dual-CNN expert-guided Transformer model designed for semantic segmentation using point-level annotations to reduce this annotation burden. SparseFormer comprises three branches, where two CNN branches employ different attention mechanisms to encourage diverse outputs. To enhance the local consistency of pseudo-labels, we introduce a pixel-adaptive refinement module that dynamically refines CNN output probabilities by incorporating image information during training. A credible assessment is then performed to combine the CNN outputs, producing highquality pseudo-labels that supervise the CNN-Transformer hybrid branch. This hybrid branch integrates global representations with local features, achieving precise segmentation. To further strengthen the CNN branches, we introduce a knowledge distillation strategy that steadily feeds back information from the hybrid branch to CNN branches, mitigating overfitting risks caused by sparse supervision. SparseFormer employs credible assessment to reduce pseudo-label uncertainty, followed by continuous interaction and dynamic information enhancement among the three branches in an end-to-end training process. Extensive experiments on two benchmark datasets demonstrate that SparseFormer significantly outperforms state-of-the-art methods. Our code is available at https://github.com/Yujia73/SparseFormer.

Temporal-based methods effectively improve the utilization rate of remote sensing images but large ratios of missing information still need to be improved in the reconstruction models. In this article, based on the imaging theory with the help of a radiation correction model, a decoupling-reconstruction network (DecRecNet) for image reconstruction is proposed. The network uses a ground content radiation (GCR) correction module and imaging environment radiation (IER) correction module and their corresponding loss functions to decouple the image radiation information into GCR related to the ground objects and IER related to the imaging conditions and carry out targeted processing to achieve the purpose of missing information reconstruction. The GCR consistency loss function is used to preserve the ground content information, and the IER consistency loss function and IER smoothness loss are used to coordinate the imaging environment. The radiation-guiding module further performs targeted radiation adjustment on the foreground and background images to transfer the background imaging environment to the foreground and consistent radiation information of the same ground object. Compared with the classical U-Net, DeepLabV3+, RFR-Net, and spatial-temporal-spectral deep convolutional neural network methods, our model showed remarkable advantages in cloud occlusion and stripes of Landsat-8 (30 m), GaoFen-1 (2 m), and Landsat-7 (30 m) at various missing ratios, data sources, resolutions, and scenes, and achieving the goal of missing information reconstruction in large missing ratios.