Research on 3D Reconstruction Methods of Remote Sensing Images Combined with Deep Learning and GIS

Abstract

Three-dimensional reconstruction of remote sensing images represents a key research direction in integrating geographic information systems (GIS) with remote sensing data. This study proposes a comprehensive technical approach that combines deep learning techniques with GIS to enhance the reconstruction of remote sensing imagery, addressing common challenges such as limited accuracy, low efficiency, and difficulties in semantic interpretation. Specifically, an improved U-Net network is employed to perform semantic segmentation on remote sensing images, enabling the extraction of critical land feature information while preserving spatial and structural details. Following feature extraction, a three-dimensional registration method is integrated with dense point clouds to achieve high-precision terrain reconstruction, ensuring accurate spatial alignment and continuity across the reconstructed surface. In addition, GIS-based procedures are applied to perform spatial positioning, attribute integration, and three-dimensional visualization, allowing the reconstructed terrain and land features to be effectively interpreted and analyzed within a geographic context. Compared with traditional reconstruction methods, this integrated approach demonstrates higher positioning accuracy, improved model fidelity, and superior semantic reconstruction capabilities. By combining deep learning-based feature extraction with GIS-enabled spatial analysis, the method offers a more effective and robust solution for three-dimensional remote sensing reconstruction, providing enhanced applicability for geographic analysis, environmental monitoring, and urban planning applications.