Using Drone Based LiDAR to Reconstruct Forests

Abstract

Forests are an essential ecosystem for the sequestration of CO2, the dominant greenhouse gas driving climate change, in the environment. The ability to accurately determine the amount of carbon stock and sequestration within this system through biomass estimation is crucial to informing carbon budgets, carbon offset projects, and commercial forestry. However, national, and regional biomass models rely heavily on laborious stand-level, typically field-derived, metrics such as Diameter at Breast Height (DBH) of individual trees, that are then scaled up via models with satellite imagery. To facilitate easier biomass estimation, this study employed the use of a small-footprint Light Detection and Ranging (LiDAR) sensor, a Small Unmanned Aircraft System (SUAS), and advanced LiDAR point cloud processing to extract and estimate the DBH of individual stems at a well-studied ForestGEO site (12.5ha.) in Virginia. Unmanned aerial vehicle borne Laser Scanning (ULS), as performed here, can significantly improve standxv level biomass estimates which can then be used to develop empirical models that predict regional biomass using satellite imagery. Our specific objectives were to (i) assess the ability to automatically detect and extract individual tree stems using Density Based Spatial Clustering of Applications with Noise (DBSCAN), and more prominently (ii) test the accuracy of four DBH estimation methods adopted from Terrestrial-Laser-Scanning (TLS) and Airborne-Laser-Scanning (ALS) at stand-level scales. The DBH estimation methods assessed were (i) Convex Hull approach (CH), (ii) Pratt (Pt) and (iii) Levenberg-Marquardt (LM) circle fitting, and (iv) Random Sample Consensus (RANSAC). We demonstrated that through DBSCAN, individual stems larger than 18cm DBH could be detected across the full study area with an accuracy of 65%. Estimation bias was the lowest in small stems ranging from 10-50cm (67% of the known stems); where all DBH estimation methods displayed a relationship of increasing negative bias (underestimation) for progressively larger stems. For stems approximately 10-20cm DBH, LM and RANSAC had a positive bias of 1.6 and 3.8cm, which turned negative and increased to - 10.7 and -9.3cm for stems 40-50cm DBH. Pt failed to reconstruct small stems 10-20cm DBH with an initial bias of 14.2cm which then decreased to -0.2cm at 40-50cm DBH. CH similarly failed to reconstruct small stems but had the smallest overall range in bias across the 10-50cm DBH interval of 10.8-4cm. Due to errors in co-location between the ForestGEO data and the ULS point cloud, initial R2 values were low for the full study area with the highest being .17 for LM, followed by .16, .06, and .04 for CH, RANSAC, and Pt, respectively. Limiting the analysis to only high-confidence matches between in-situ and ULS clusters drastically improved R2 to .69, .71, .23, and .13 for LM, CH, RANSAC, and Pt correspondingly. This underscores the importance of reliably aligning the two datasets before analyses. With these findings, this study hopes to pave the way for ULS DBH estimation for individual stems and provide a significant contribution towards the improvement of nondestructive biomass estimation. Through this study and its successors, rapid stand-level metrics will be attainable from UAS LiDAR and could supplement regional satellite and ALS-based biomass estimates.