Deep learning approaches have been widely used in building automatic extraction tasks and have made great progress in recent years. However, the missing detection and wrong detection causing by spectrum confusion is still a great challenge. The existing fully convolutional networks (FCNs) cannot effectively distinguish whether the feature differences are from one building or the building and its adjacent non-building objects. In order to overcome the limitations, a building multi-feature fusion refined network (BMFR-Net) was presented in this paper to extract buildings accurately and completely. BMFR-Net is based on an encoding and decoding structure, mainly consisting of two parts: the continuous atrous convolution pyramid (CACP) module and the multiscale output fusion constraint (MOFC) structure. The CACP module is positioned at the end of the contracting path and it effectively minimizes the loss of effective information in multiscale feature extraction and fusion by using parallel continuous small-scale atrous convolution. To improve the ability to aggregate semantic information from the context, the MOFC structure performs predictive output at each stage of the expanding path and integrates the results into the network. Furthermore, the multilevel joint weighted loss function effectively updates parameters well away from the output layer, enhancing the learning capacity of the network for low-level abstract features. The experimental results demonstrate that the proposed BMFR-Net outperforms the other five state-of-the-art approaches in both visual interpretation and quantitative evaluation.

Clouds and snow in remote sensing imageries cover underlying surface information, reducing image availability. Moreover, they interact with each other, decreasing the cloud and snow detection accuracy. In this study, we propose a convolutional neural network for cloud and snow detection, named the cloud and snow detection network (CSD-Net). It incorporates the multi-scale feature fusion module (MFF) and the controllably deep supervision and feature fusion structure (CDSFF). MFF can capture and aggregate features at various scales, ensuring that the extracted high-level semantic features of clouds and snow are more distinctive. CDSFF can provide a deeply supervised mechanism with hinge loss and combine information from adjacent layers to gain more representative features. It ensures the gradient flow is more oriented and error-less, while retaining more effective information. Additionally, a high-resolution cloud and snow dataset based on WorldView2 (CSWV) was created and released. This dataset meets the training requirements of deep learning methods for clouds and snow in high-resolution remote sensing images. Based on the datasets with varied resolutions, CSD-Net is compared to eight state-of-the-art deep learning methods. The experiment results indicate that CSD-Net has an excellent detection accuracy and efficiency. Specifically, the mean intersection over the union (MIoU) of CSD-Net is the highest in the corresponding experiment. Furthermore, the number of parameters in our proposed network is just 7.61 million, which is the lowest of the tested methods. It only has 88.06 GFLOPs of floating point operations, which is less than the U-Net, DeepLabV3+, PSPNet, SegNet-Modified, MSCFF, and GeoInfoNet. Meanwhile, CSWV has a higher annotation quality since the same method can obtain a greater accuracy on it.

Deep learning-based disparity estimation methods have demonstrated significant potential in optical satellite stereo image applications. However, learning-based methods remain susceptible to domain shifts caused by spatiotemporal variations and stereo-sensor heterogeneity. To address these challenges, we propose a Hierarchical Domain Adaptation Disparity Estimation framework (HDADE) for optical satellite stereo images. HDADE was structured with a four-stage technique pipeline to improve the training data quality and diversity, explicitly align the spectral and stereo distribution, implicitly enhance the robustness of feature extraction and matching, directly facilitate feature alignment with the target domain. This hierarchical framework systematically mitigates disparity estimation accuracy degradation in cross-domain scenarios. Cross-spatiotemporal and cross-payload generalization experiments were conducted based on the WHU_Stereo and US3D datasets. The experimental results show that HDADE significantly outperformed other advanced methods and possessed plug-and-play versatility. Notably, greater domain shift scene transfer experiments indicated that, with limited annotation data, HDADE has the potential for large-scale automatic applications.

Accurate height estimation from satellite imagery is essential for digital surface model (DSM) generation in remote sensing applications. Although recent deep learning-based multi view stereo (MVS) methods have advanced height prediction, they often struggle with edge degradation, particularly around building outlines and terrain discontinuities. To address this issue, we propose SP-MVS, a structure-preserving MVS network that explicitly enhances edge-aware height estimation. The network integrates a triple-branch structural encoder equipped with a U-Net backbone and the Convolutional Block Attention Module (CBAM) to capture multi-scale spatial and texture features. A Mamba-based cost volume regularization module is introduced to model long-range cross-view dependencies with low computational overhead. Furthermore, we design a multistage Edge–texture consistency loss to guide the network in preserving sharp structural boundaries throughout the estimation process. SP-MVS directly produces high-quality height maps and serves as a critical component in precise DSM generation. Extensive experiments on the WHU-TLC and US3D datasets demonstrate that SP-MVS achieves superior accuracy and sharper boundary delineation compared to state-of-the-art methods.

This paper proposes an unsupervised knowledge distillation framework for satellite multi-view stereo (MVS) reconstruction under label-free settings. A teacher–student paradigm is adopted, where a teacher MVS network is first trained using self-supervised multi-view geometric constraints, including photometric consistency, feature similarity, smoothness, and structural similarity. The teacher produces multi-view height predictions, and geometry-aware uncertainty is inferred from cross-view consistency during multi-view fusion. Based on the fused height estimates and uncertainty cues, reliable pseudo-labels are constructed in the form of expectation-based height maps, pixel-wise uncertainty maps, and visibility masks. These pseudo-labels supervise a lightweight student network through probability-volume distillation, combined with uncertainty-weighted height regression and cross-view geometric consistency constraints. Extensive experiments on representative satellite benchmarks demonstrate that the proposed framework achieves competitive accuracy among unsupervised methods while significantly reducing computational cost and eliminating the need for LiDAR ground truth or pre-trained models, providing a scalable solution for large-scale satellite 3D reconstruction.