An Adaptive NeRF Framework for City-Scale Emergency Awareness

Rapid situational awareness during disasters such as floods, earthquakes, and wildfires demands 3D scene representations that can update seamlessly as new aerial data arrive. Neural Radiance Fields (NeRFs) are a technique that constructs such 3D scene representations. Conventional NeRF-based pipelines remain static and require full retraining when additional observations become available, leading to delays that hinder real-time decision-making. In this paper, we introduce a city-scale NeRF framework for high-speed adaptation and real-time rendering under rapidly evolving data conditions. Our approach combines three key components: meta-continual learning (MCL) for fast adaptation without catastrophic forgetting, modularized neural architectures to support large-scale spatial decomposition, and optimization based on Instant Neural Graphics Primitives (Instant-NGP) for efficient training and compact representation. Together, these components enable real-time refinement of city-scale NeRFs as new observations become available.