A critical analysis of satellite stereo pairs for digital surface model generation and a matching quality prediction model

Abstract The geometric analysis and data acquisition of satellite photogrammetric images are often regarded as a direct extension of traditional aerial photogrammetry, with the only difference being the sensor model (linear array vs. central perspective). The intersection angle (or base-height ratio) between two images is seen as the most important metadata of stereo pairs, which directly relates to the base-high ratio and texture distortion in the parallax direction, thus both affecting the horizontal and vertical accuracy. State-of-the-art DIM algorithms were reported to work best for narrow baseline stereos (small intersection angle), e.g. Semi-Global Matching empirically takes 15–25° as “good” intersection angles. However, our experiments found that the intersection angle is not the only determining factor, as the same DIM algorithm applied to stereo pairs of the same area with similar and good intersection angle may produce point clouds with dramatically different accuracy (demonstrated in the graphical abstract). This raises a very practical and often asked question: what factors constitute a good satellite stereo pair for DIM algorithms? In this paper, we provide a comprehensive analysis on this matter by performing stereo matching using the very typical and widely-used Semi-Global Matching (SGM) with a Census cost over 1000 satellite stereo pairs of the same region with different meta-parameters including their intersection, off-nadir, sun elevation & azimuth angles, completeness and time differences, thus to offer a thorough answer to this question. Our conclusion has specifically outlined an important yet often ignored factor – the Sun-angle difference to be one decisive in determining good stereo pair. Based on the analytical results, we propose a simple idea by training a support vector machine model for predicting potential stereo matching quality (i.e. potential level of accuracy and completeness given a stereo pair). Experiments have shown that the model is well-suited and generalized for multi-stereo 3D reconstruction, evidenced by a comparative analysis against three other strategies: (1) pair selection based on an example patch where partial ground-truth data is available for computing a priori ranking (2) based on intersection angles and (3) based on a recent algorithm using intersection angle, off-nadir angle and time intervals. This work will potentially provide a valuable reference to researchers working on multi-view satellite image reconstruction, as well as for practitioners minimizing costs for high-quality large-scale mapping. The trained model is made available to the academic community upon request.