Using LiDAR on a Ground-based Agile Robot to Map Tree Structural Properties

The widespread use of diameter at breast height (DBH) and tree height attributes as a non-destructive indirect estimation of tree parameters (e.g., above-ground biomass, volume, age, and carbon stock) demands efficient and accurate surveying methods. However, traditional surveys, which are primarily manual, are often time-consuming, inaccurate, inconsistent, and might suffer from observer-bias. This study applies an agile quadruped robot, Spot from Boston Dynamics, and a mounted LiDAR system for mapping and measuring tree height, diameter at breast height (DBH), and tree volume. This project uses the Spot Enhanced Autonomy Payload (EAP) navigation module as the source of LiDAR data. The use of this module has two main advantages. First, Spot EAP's VLP-16 sensor is a low-beam LiDAR that, as demonstrated in previous research, is capable of estimating tree structural parameters while consuming less time and data than robust systems such as Terrestrial Laser Scanning (TLS). Second, using an existing payload as the primary source of data without disabling its default function results in more efficient payload capacity utilization and, as a result, lower energy consumption, in addition to making room for additional payloads. The experiment was conducted for 41 trees (23 Erythrina variegata and 18 Ficus altissima) in a park on the campus of King Abdullah University of Science and Technology (KAUST) in Saudi Arabia. TLS data were used to compute the height and volume reference data, while manual measurements were used to obtain DBH reference data. The robot-derived point cloud generation methodology was based on a multiway registration approach in which a total of 76 scans were acquired from 4 different locations using multiple poses of the robot to overcome the short field of view of the LiDAR sensor. As a result of processing the scans, a point cloud for each of the trees was obtained. The height estimations, which consist of a difference within Z coordinates, obtained a mean absolute error (MAE) and a mean percentage error (MPE) of 6.71 cm and 1.31% respectively. The DBH estimation based on circle-fitting algorithms obtained an MAE and an MPE of 2.55 and 12.99% respectively. The volume estimation obtained a coefficient of determination of 0.93. When compared to the most recent approaches available in the literature, the results for height and volume were satisfactory, yielding higher accuracy than other studies in some cases. The results for DBH estimation were also comparable to those in the literature. The main sources of error were tree occlusion and inclined trees, both of which are solvable by including more scanning locations and increasing the robustness of software estimation. Consequently, the acquisition system is not a barrier to future improvements. This work successfully introduced one of the first methods for using agile robots in high throughput field phenotyping. The use of agile robots addresses some of the major challenges for deploying ground-based robotics in high throughput field phenotyping, allowing for a higher assessment frequency without causing soil compaction and damage, as well as bringing unprecedented adaptation to difficult terrains.