NeRF-Accelerated Ecological Monitoring in Mixed-Evergreen Redwood Forest

Abstract

Forest mapping provides critical observational data needed to understand the dynamics of forest environments. Notably, tree diameter at breast height (DBH) is a metric used to estimate forest biomass and carbon dioxide sequestration. Manual methods of forest mapping are labor intensive and time consuming, a bottleneck for large-scale mapping efforts. Automated mapping relies on acquiring dense forest reconstructions, typically in the form of point clouds. Terrestrial laser scanning (TLS) and mobile laser scanning (MLS) generate point clouds using expensive LiDAR sensing, and have been used successfully to estimate tree diameter. Neural radiance fields (NeRFs) are an emergent technology enabling photorealistic, vision-based reconstruction by training a neural network on a sparse set of input views. In this paper, we present a comparison of MLS and NeRF forest reconstructions for the purpose of trunk diameter estimation in a mixed-evergreen Redwood forest. In addition, we propose an improved DBH-estimation method using convex-hull modeling. Using this approach, we achieved 1.68 cm RMSE, which consistently outperformed standard cylinder modeling approaches. Our code contributions and forest datasets are freely available at https://github.com/harelab-ucsc/RedwoodNeRF.

Figures (5)

• Figure 1: TreeTool process applied to a forest NeRF reconstruction. (A) ground segmentation, (B) trunk segmentation, and (C) trunk modeling. Our tree modeling approach considers trees as stacks of convex-hull slices which outperformed other approaches by 3-4$\times$ in terms of DBH estimation accuracy.
• Figure 2: NeRF scene representation flow. Sparse images with corresponding poses are sampled using ray-tracing to generate 5D input vector comprised of location ($x, y, z$) and viewing direction ($\theta, \phi$). A cascaded MLP learns the weights to map this 5D vector to output color (r, g, b) and volume density $\sigma$. Volume rendering composites the learned rays to novel views.
• Figure 3: Quadruped robot creating a dense LiDAR-inertial reconstruction in a forest environment (left). LIOSAM visualization of estimated trajectory (torqouise), loop-closure events (yellow), and tightly aligned LiDAR scans (magenta).
• Figure 4: Four comparisons of RANSAC and convex-hull modeling approaches. Deltas between manual DBH and each modeling approach are provided on the top line. RANSAC cylinder modeling consistently under-fits well-represented trunk projections. Convex hull DBH estimation outperformed RANSAC by 3-4$\times$.
• Figure 5: Forest reconstructions produced by SLAM (bottom row) and NeRF (top row) methods of both datasets. Adjacent plots are data collection trajectories for each reconstruction. The figure also compares a zoomed-in section of a tree trunk. The NeRF reconstruction is 4$\times$ denser and is of higher surface quality.