Fast and Generalizable NeRF Architecture Selection for Satellite Scene Reconstruction

2026-03-18

AI Summary

This work addresses the limitations of Neural Radiance Fields (NeRF) in satellite scene reconstruction, where per-scene independent training and prohibitively expensive neural architecture search hinder scalability. The authors propose PreSCAN, a novel framework that, for the first time, demonstrates multi-view consistency as the dominant factor governing reconstruction quality—surpassing the influence of model architecture itself. Leveraging lightweight geometric and photometric descriptors, PreSCAN predicts NeRF reconstruction performance across scenes without any training, selecting optimal architectures within 30 seconds with prediction errors below 1 dB. Integrated with SHAP-based interpretability, offline power modeling, and edge deployment optimizations, the method reduces power consumption by 26% and latency by 43% on a Jetson Orin platform, achieves zero-shot generalization on the DFC2019 dataset, and accelerates architecture search by up to three orders of magnitude.

Technology Category

• Computer Vision: Scene Analysis & Understanding
• Search and Optimization: Learning to Search
• Machine Learning: Deep Neural Architectures and Foundation Models

Application Category

• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines
• Web Mining and Content Analysis: Large pretrained models with web data
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

Neural Radiance Fields (NeRF) have emerged as a powerful approach for photorealistic 3D reconstruction from multi-view images. However, deploying NeRF for satellite imagery remains challenging. Each scene requires individual training, and optimizing architectures via Neural Architecture Search (NAS) demands hours to days of GPU time. While existing approaches focus on architectural improvements, our SHAP analysis reveals that multi-view consistency, rather than model architecture, determines reconstruction quality. Based on this insight, we develop PreSCAN, a predictive framework that estimates NeRF quality prior to training using lightweight geometric and photometric descriptors. PreSCAN selects suitable architectures in < 30 seconds with < 1 dB prediction error, achieving 1000$\times$ speedup over NAS. We further demonstrate PreSCAN's deployment utility on edge platforms (Jetson Orin), where combining its predictions with offline cost profiling reduces inference power by 26% and latency by 43% with minimal quality loss. Experiments on DFC2019 datasets confirm that PreSCAN generalizes across diverse satellite scenes without retraining.

Problem

• Neural Radiance Fields
• satellite scene reconstruction
• Neural Architecture Search
• multi-view consistency
• edge deployment

Innovation

• Neural Radiance Fields
• Neural Architecture Search
• multi-view consistency
• satellite scene reconstruction
• edge deployment