Creating Digital Surface Models from satellite imagery using Neural Radiance Fields (NeRF)

Abstract

Neural Radiance Fields (NeRFs) are increasingly used to generate Digital Surface Models (DSMs) from satellite imagery. However, research is far from mature and questions considering the accuracy in different environments, sensitivity and applicability have not been addressed. This research aims to answer these questions by firstly conducting a thorough literature research, resulting in the selection of a suitable NeRF model called SAT-NGP. The SAT-NGP model is tested on accuracy and runtimes in both urban and rural environments using the Data Fusion Contest 2019 (DFC2019) dataset, both prepared and unprepared. Furthermore, sensitivity analysis on the number of iterations and the ray batch size is conducted. The sensitivity analysis is followed by an applicability research considering the implementation of Superview-1 data of the Netherlands. This research has found that SAT-NGP, using the DFC2019 dataset, creates more accurate DSMs for urban areas compared to rural areas. Additionally, challenging situations are identified where NeRF produces DSM accuracy errors. These challenging situations include edges of buildings, shadows, flat surfaces and trees. The sensitivity analysis findings indicate that a lower number of iterations can be used with minimal to no effect on the accuracy, reducing runtime significantly by 50%. Altering the ray batch size yielded no improvements concerning accuracy or runtime. Furthermore, the applicability analysis results show that it is complicated to apply SAT-NGP to data different from DFC2019. This is mainly due to the numerous input files that are needed, which all have to be perfectly adjusted and aligned to each other. Through the different analyses and literature research, the findings of this study contribute to the DSM generating NeRF research field. The results show that NeRFs are a promising development in the field of DSM construction from satellite imagery, but many improvements can still be made concerning accuracy methods, sensitivity research and applicability in general.