GPE-DNeRF:Neural radiance field method for surface geological bodies reconstruction

Three-dimensional (3D) geological models are crucial for a comprehensive understanding of regional geological formations. Deep learning-based 3D reconstruction technologies offer highly automated approaches for recognizing complex data patterns and generating realistic reconstruction results. The application of these methods in the reconstruction of surface geological bodies is particularly significant in the context of advancing the construction of digital mines nationwide. Neural Radiance Fields (NeRF) have been employed to generate 3D scenes by training models on images captured from different viewpoints. However, parallax errors across viewpoints may lead to misalignment or overlapping of details in the generated images, especially in regions with complex geometric structures. These errors can hinder the model's ability to accurately reconstruct surface details, resulting in substantial distortions in the final output. To address this issue and reduce artifacts and noise in the reconstructed 3D surface geological model, this study explores the use of NeRF for geologic body reconstruction. We propose an enhanced method, GPE-DNeRF, which integrates depth information with Gaussian positional encoding to achieve high-quality reconstruction of geological surfaces. The performance of the proposed method is evaluated, and comparative analyses are conducted with the SfM-MVS and NeRF methods. The GPE-DNeRF method demonstrates a strong capability to eliminate artifacts and retain detailed terrain features, thereby enhancing reconstruction quality and ensuring a closer alignment with actual surface geological conditions.