Learning Neural Radiance Fields of Forest Structure for Scalable and Fine Monitoring

TL;DR

The paper tackles the challenge of scalable, high‑resolution forest monitoring by leveraging neural radiance fields (NERF) to fuse multi‑view RGB imagery with LiDAR (ALS and TLS). It formalizes a joint loss that combines color consistency with LiDAR depth priors, enabling accurate 3D reconstruction of forest structure from diverse sensing modalities and viewpoints. Through terrestrial and aerial experiments, the work demonstrates that NeRF can recover fine vertical structure, improve sub-canopy details with LiDAR priors, and enhance forest factor metrics such as tree counts and DBH, often approaching in‑situ TLS performance while maintaining scalability. The results suggest that neural fields offer a practical pathway to scalable, cost-effective forest monitoring with near in-situ accuracy across large landscapes.

Abstract

This work leverages neural radiance fields and remote sensing for forestry applications. Here, we show neural radiance fields offer a wide range of possibilities to improve upon existing remote sensing methods in forest monitoring. We present experiments that demonstrate their potential to: (1) express fine features of forest 3D structure, (2) fuse available remote sensing modalities and (3), improve upon 3D structure derived forest metrics. Altogether, these properties make neural fields an attractive computational tool with great potential to further advance the scalability and accuracy of forest monitoring programs.

Figures (6)

• Figure 1: Even though SFM reconstruction is capable of extracting the 3D structure of tree, its recontruction suffers from sparsity. Such sparsity limits the spatial variability of structure that can be captured thorugh such models.
• Figure 2: Neural field models are capable of extracting fine 3D structure from terrestrial multi-view images in forestry. Reconstructions demonstrate their potential to represent fine scale variability in heterogeneous forest ecosystems.
• Figure 3: Forest structure from TLS and ALS: ALS provides sparse spatial information and is not capable of resolving sub-canopy detail. TLS on the other hand, provides fine spatial variability and resolution along full 3D vertical stands.
• Figure 4: TLS to ALS co-registration: Forest features are well aligned qualitatively between both ALS and TLS sensing.
• Figure 5: AI-based extraction of 3D structures from aerial multi-view 2D images + 3D point cloud data inputs. Imposing point cloud priors into 3D structure extraction improves distance ambiguities in structure and resolves artifact issues likely at far ranges.