Ortho-NeRF: generating a true digital orthophoto map using the neural radiance field from unmanned aerial vehicle images

True Digital Orthophoto Maps (TDOMs) have high geometric accuracy and rich image characteristics, making them essential geographic data for national economic and social development. Complex terrain and artificial structures, automatic distortion elimination and occluded area recovery in TDOM generation pose significant challenges. Hence, the need for further improvements in both mapping accuracy and automation is highlighted. In this paper, we present an approach for generating a TDOM based on a Neural Radiance Field (NeRF) without utilizing prior three-dimensional geometry information called an Ortho Neural Radiance Field (Ortho-NeRF). The Ortho-NeRF divides a large-scale scene into small tiles, implicitly reconstructing each tile by selecting pixels on posed images, and individually generate TDOMs of all tiles using a true-ortho-volume rendering before mosaicking. Additionally, the Ortho-NeRF uses a strategy to skip empty spaces and adaptively set the spatial resolution of a voxel grid, improving the generated TDOM quality with fewer computational resources. Many experiments showed that our approach outperforms ContextCapture, Metashape, Pix4DMapper, and Map2DFusion, especially in challenging areas. Owing to its global consistency and continuous nature, Ortho-NeRF was able to effectively reconstruct the geometry information and details, generating TDOMs without distortion or misalignment. Eight ground control points were randomly selected to evaluate the geometric accuracy of the TDOMs, with an average median error of 0.267 m. The length between two points on a plane was also measured for quantitative evaluation, with a mean absolute error of 0.08 m and a mean relative error of 0.14%. Compared with the NeRF efficiency, that of the Ortho-NeRF increased 104 times in training and about 1000 times in rendering.