Aerospace photogrammetry serves as a critical methodology for large-scale 3D reconstruction, where disparity estimation constitutes the fundamental technology. With various types of complex scenes covered in high-resolution satellite images, traditional stereo matching methods are facing a great challenge. Although deep learning-based disparity estimation networks have gained significant traction in recent years, research within satellite image stereo matching predominantly prioritizes model accuracy while neglecting computational efficiency and model complexity, which runs counter to the general trend of onboard intelligent processing in modern remote sensing satellites. To bridge this gap, we propose SIGLNet, a semantic information-guided lightweight end-to-end stereo matching network specifically designed for high-resolution satellite stereo imagery. The SIGLNet first employs a lightweight backbone to extract multi-scale deep features, which are fused into the downsampled low-resolution feature maps. Utilizing these representations, we then construct a low-resolution cost volume and deploy efficient cost aggregation via inverted residual blocks and multi-branch adjustable bottleneck modules, producing an initial low-resolution disparity map. Finally, a semantic information-guided disparity upsampling module reconstructs high-quality full-resolution disparity maps. Extensive experiments conducted on the US3D and the WHUStereo benchmark demonstrate that the proposed SIGLNet achieves the highest disparity estimation accuracy while maintaining competitive model complexity compared to widely used methods.

Dynamic imaging of optical remote sensing satellites refers to the active acquisition of images while the satellite is maneuvering at a high angular velocity, which significantly enhances the efficiency and application value of remote sensing imagery. However, due to the influence of rapid maneuvering, the random error of star sensors significantly increases, resulting in a decrease in the geometric positioning accuracy of dynamic imaging remote sensing images. This paper proposes a dynamic fusion method for multisource attitude measurement data based on the noise adaptive estimation and bidirectional filter to achieve high-precision attitude determination and geometric positioning in dynamic imaging. Based on the measurement error model of star sensors, the weights of the star sensor and gyroscope are adaptively adjusted in multisource data fusion to reduce the impact of star sensor measurement errors on the gyroscope. Moreover, a bidirectional fusion filter that includes a low-velocity maneuvering stage is proposed to realize the optimal estimation of the satellite attitude parameter. The simulation data and onboard data of the Luojia3-01 (LJ3–01) satellite were tested to verify the effectiveness of the proposed method. The geometric positioning accuracy of the staring images of LJ3–01 improved from 10.048 m to 7.538 m. The registration accuracy of the sequential images improved from 3.568 pixels to 1.179 pixels. The proposed method can significantly improve the attitude determination accuracy and the geometric positioning accuracy of LJ3–01 satellite staring images. Moreover, for simulation data with various angular velocities, the attitude determination accuracies of the proposed method are better than 0.93”. The experimental results show that the proposed method can achieve high-precision attitude determination in dynamic imaging, reaching the accuracy in the traditional passive imaging.

Representation learning is one of the core problems in machine learning research. The transition of input representations for machine learning algorithms from handcraft features, which dominated in the past, to the potential representations learned through deep neural networks nowadays has led to tremendous improvements in algorithm performance. However, the current representations are usually highly entangled, i.e., all information components of the input data are encoded into the same feature space, thus affecting each other and making it difficult to distinguish. Disentangled representation learning aims to learn a low-dimensional interpretable abstract representation that can identify and isolate different potential variables hidden in the high-dimensional observations. Disentangled representation learning can capture information about a single change factor and control it by the corresponding potential subspace, providing a robust representation for complex changes in the data. In this paper, we first introduce and analyze the current status of research on disentangled representation and its causal mechanisms and summarize three crucial properties of disentangled representation. Then, disentangled representation learning algorithms are classified into four categories and outlined in terms of both mathematical description and applicability. Subsequently, the loss functions and objective evaluation metrics commonly used in existing work on disentangled representation are classified. Finally, the paper summarizes representative applications of disentangled representation learning in the field of remote sensing and discusses its future development.