A Learning-Based Dual-Scale Enhanced Confidence for DSM Fusion in 3-D Reconstruction of Multiview Satellite Images

Abstract

Compared to two-view reconstruction, multiview imagery leverages redundant information to mitigate the effects of occlusion and noise. Deep-learning-based multiview stereo (MVS) methods are primarily applicable to tristereo data captured simultaneously and rely heavily on training samples. Traditional MVS methods typically rely on simple filtering and weighting techniques for digital surface model (DSM) fusion based on image pair selection and pairwise stereo matching, which are usually affected by poor image pairs and fail to fully exploit the complementary advantages of DSMs. To address these limitations, this article proposes a novel DSM fusion method incorporating learning-based dual-scale enhanced confidence for three-dimensional reconstruction from multiview satellite imagery. First, a generalized stereo matching method is adopted, which considers radiometric differences and small feature variations. Next, auxiliary information generated during pairwise stereo reconstruction is utilized to construct a high-dimensional confidence vector that includes classical confidence measures and a newly designed topological structure relationship consistency measure. Then, a guided regularized random forest regressor is employed to identify influential confidence measures and establish their correlation with reconstruction accuracy, leading to the estimation of enhanced confidence. Additionally, to preserve fine details and boundary information, the dual-scale enhanced confidence is introduced to facilitate cross-scale DSM fusion. Finally, each view is sequentially treated as the master view to obtain DSMs, which are then fused to produce the final DSM. Experimental results demonstrate that the proposed method achieves superior performance across various datasets, including tristereo Beijing-3 data acquired nearly simultaneously, multiview WorldView-3 data captured at different times, and self-made tristereo data collected from different sensors at different times. The proposed method achieves an average MAE of 1.14 m, RMSE of 2.16 m, median height error of 0.47 m, and COMP of 75.13%, outperforming several mainstream methods.