Comparison of Long Deep Learning–Based 3D Reconstruction Models for Visualizing the Fukushima Daiichi Nuclear Power Station

The Fukushima Daiichi nuclear power station presents an unprecedented challenge for reactor decommissioning, requiring visualization of environments with extreme radiation, poor lighting, and underwater conditions. This study investigates deep learning–based three-dimensional (3D) reconstruction methods capable of processing long video sequences typical of reactor surveys. Three representative approaches were evaluated: LocalRF [Neural Radiance Fields (NeRF)–based model], LongSplat (Gaussian splatting–based model), and VGGT-Long (foundation model). Publicly released Tokyo Electric Power Company videos from multiple reactor inspections were used to benchmark performance under realistic conditions. Evaluation metrics included peak signal-to-noise ratio, structural similarity, and computational time. Results indicate that LongSplat achieves higher pixel-level fidelity and faster processing than LocalRF, while LocalRF demonstrates superior structural consistency and robustness across all datasets. VGGT-Long offers the shortest computation times but produces significantly lower perceptual quality, highlighting limitations in current 3D foundation models for this domain. These findings underscore trade-offs among accuracy, stability, and efficiency and suggest that hybrid approaches utilizing foundation models could enable real-time, high-quality reconstruction in hazardous environments such as nuclear reactors after meltdown. Advancing these capabilities is critical for accelerating nuclear decommissioning and improving situational awareness in extreme conditions.