Abstract:To address the challenges of dataset scarcity and insufficient reconstruction accuracy in synthetic aperture radar (SAR) image three-dimensional reconstruction, a novel approach based on the Neural Radiance Fields (NeRF) model is proposed. First, the bidirectional analytic ray tracing (BART) method is used to generate a simulated SAR dataset, and the ColMap tool is applied to obtain camera poses and sparse reconstruction data. For data processing, techniques such as image enhancement, speckle noise suppression, and sidelobe suppression are employed to ensure high-quality input data. Based on this high-quality data, the NeRF model is trained to achieve three-dimensional reconstruction of buildings, with notable performance in recovering the details of disk-shaped structures, especially in cylindrical and stair-like structures.In the experiment, the dimensions of the optical simulation model were recorded as a ratio of height: length: width = 7.3: 4.55: 1, while the final reconstructed model"s dimensions were measured at a ratio of height: length: width = 7.874: 5.058: 1, with an error margin kept within 6%. The experimental results demonstrate that the proposed method effectively restores the main contours and structural features of the building, particularly achieving high precision in the reconstruction of the columnar sections. Although the precision of the foundation"s reconstruction requires further enhancement, the overall approach offers innovative insights and technical support for the application of SAR building image in 3D reconstruction.
