Highly Applicable Iterative Network for 3-D Reconstruction Based on Multiview Satellite Images

Abstract

Multiview satellite images play a crucial role in reconstructing Earth's surface information. Its application in deep-learning-based multiview stereo (MVS) methods has recently exhibited impressive performance. However, there is still ample room for improvement, particularly in the reconstruction of the digital surface model (DSM) within scenes characterized by dense features, high repetition, and significant elevation variations. To address the challenges, we present a novel network, Appli-SatNet. The network initially introduces a multiscale transformer module to emphasize feature correlations within and across multiple images. Subsequently, a novel gated recurrent units iterative optimizer is introduced to derive iterative indices for the cost volume within implicit states. This iterative process enhances the correlation among the cost volumes and enables multiple iterative refinements for prediction elevation maps. Ultimately, this leads to the generation of a high-resolution DSM that matches the resolution of the nadir image. The method is validated on a publicly available TLC SatMVS dataset, which includes plain areas with repetitive textures and inconspicuous features, mountainous regions with significant continuous elevation changes, and urban areas with dramatic elevation variations and high repetition. Comparative to mainstream methods, our approach demonstrated remarkable improvements in reconstruction quality. Moreover, when applied to low-resolution images captured by the Ziyuan-3 01 satellite over various complex terrains, it consistently exhibits robust reconstruction accuracy and superior visual quality, underscoring the applicability of our model.

Keywords

• Deep learning
• multiview satellite images
• three-dimensional (3-D) reconstruction