Sapling-NeRF: Geo-Localised Sapling Reconstruction in Forests for Ecological Monitoring

Abstract

Saplings are key indicators of forest regeneration and overall forest health. However, their fine-scale architectural traits are difficult to capture with existing 3D sensing methods, which make quantitative evaluation difficult. Terrestrial Laser Scanners (TLS), Mobile Laser Scanners (MLS), or traditional photogrammetry approaches poorly reconstruct thin branches, dense foliage, and lack the scale consistency needed for long-term monitoring. Implicit 3D reconstruction methods such as Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS) are promising alternatives, but cannot recover the true scale of a scene and lack any means to be accurately geo-localised. In this paper, we present a pipeline which fuses NeRF, LiDAR SLAM, and GNSS to enable repeatable, geo-localised ecological monitoring of saplings. Our system proposes a three-level representation: (i) coarse Earth-frame localisation using GNSS, (ii) LiDAR-based SLAM for centimetre-accurate localisation and reconstruction, and (iii) NeRF-derived object-centric dense reconstruction of individual saplings. This approach enables repeatable quantitative evaluation and long-term monitoring of sapling traits. Our experiments in forest plots in Wytham Woods (Oxford, UK) and Evo (Finland) show that stem height, branching patterns, and leaf-to-wood ratios can be captured with increased accuracy as compared to TLS. We demonstrate that accurate stem skeletons and leaf distributions can be measured for saplings with heights between 0.5m and 2m in situ, giving ecologists access to richer structural and quantitative data for analysing forest dynamics.

Figures (14)

• Figure 1: We propose a joint NeRF and LiDAR SLAM system for sapling reconstruction in situ in the forest using data from a handheld sensing device. 1) The core LiDAR SLAM system supports hectare-scale multi-session map merging as well as GNSS alignment. 2) We extract a sub-trajectory from a mapping session which encircles an individual sapling of interest and run Structure-from-Motion (SfM, COLMAP) while using the SLAM-derived trajectory to determine consistent metric scale and global localisation. The images and poses are then used to train NeRF models for each sapling. 3) Finally, we demonstrate that a NeRF-derived point cloud s sufficiently accurate to detect the tree skeleton and to measure leaf-wood separation and the leaf distribution. In Section \ref{['sec:ev_pc']} we demonstrate longitudinal monitoring of saplings - from summer to winter.
• Figure 2: The proposed system maps a forest environment at three levels of representation. Levels 3 is coarse GNSS-based localisation. Level 2 is a centimetre-accurate representation from a own multi-session LiDAR SLAM system while Level 1 is generated using images captured around a sapling, using SfM to estimate image poses and localising and scaling those poses according to the SLAM system, and finally training a scale-consistent, geo-localised NeRF model for each sapling.
• Figure 3: By co-registering multiple mapping sessions, we can create a unified map made up of trajectories across different time periods. In this way, we can monitor saplings from summer to winter.
• Figure 4: Rescaling and aligning multiple SfM reconstructions: From each SLAM-derived session of a plot (upper), we extract a subtrajectory where a sapling is scanned in detail, and then estimate a COLMAP SfM trajectory which has a scale ambiguity. This is corrected using the SLAM-derived trajectory (lower) to produce the (co-aligned) trajectories shown in red in the example.
• Figure 5: Leaf segmentation: (a) A detailed view of the NeRF-derived point cloud of sapling 01 showing individual leaves. (b) Leaf segmentation with overskeletonization for segmentation purposes, which can be avoided with appropriate parameter turning.