Optimized Image Combination Selection and Confidence-Guided DSM Fusion for 3-D Reconstruction From Multiview Satellite Imagery

The 3-D reconstruction based on multiview satellite imagery effectively leverages the redundant information to mitigate the effects of occlusions, noise, and other factors. Deep learning-based multiview stereo (MVS) methods are primarily applied to tri-stereo satellite data. Traditional MVS methods often fail to fully consider the comprehensiveness of image coverage and the information gain during image pair selection. Additionally, digital surface model (DSM) fusion typically relies on simple median filtering (MF) or its variants, which lack the flexibility required for handling heterogeneous information. To address these challenges, this article presents a novel MVS method based on optimized image combination selection and confidence-guided DSM fusion. First, to ensure comprehensive coverage and low redundancy, multiview images are partitioned based on their meta-parameters, and the interval boundaries are adjusted to ensure that the images are evenly distributed across the intervals. Next, one image is randomly selected from each interval for combination. A hierarchical heuristic strategy is then designed to select the optimized image combination that satisfies conditions such as intersection angles and sun angles. Subsequently, a stereo matching method based on iterative optimization of hierarchical graph structure consistency (GSC) cost is employed to generate multiple DSMs. Meanwhile, confidence features and estimated confidence are enhanced using contextual information. Finally, a confidence-guided DSM fusion method is designed, which effectively suppresses the interference of outliers by balancing local statistical properties with single-point confidence. Experimental results demonstrate that the proposed method outperforms mainstream methods, achieving an average mean absolute error (MAE) of 1.31 m, a root-mean-square error (RMSE) of 2.51 m, a median height error (MHE) of 0.43 m, and a COMP of 70.27%.