Evaluating the transferability of airborne laser scanning derived stem size prediction models for Pinus taeda L. stem size estimation to two different locations and acquisition specifications

• Published in: International journal of remote sensing Vol. 45; no. 16; pp. 5267 - 5294
• Main Authors: Sumnall, Matthew J., Carter, David R., Albaugh, Timothy J., Platt, Erik, Host, Trevor, Cook, Rachel L., Campoe, Otávio C., Rubilar, Rafael A.
• Format: Journal Article
• Language: English
• Published: London Taylor & Francis 17.08.2024; Taylor & Francis Ltd
• Subjects: Accuracy; Airborne lasers; ALS; Area; Biometrics; Biometry; Data acquisition; Datasets; Environmental conditions; Equipment specifications; Estimation; Evergreen trees; Forest management; Forests; ITC; Laser applications; Lasers; LIDAR; loblolly pine; Pine; Pine trees; Pinus elliottii; Pinus taeda; Prediction models; Predictions; Size estimation; Specifications; Stems; Training; UAV; Uncertainty
• ISSN: 0143-1161; 1366-5901
• DOI: 10.1080/01431161.2024.2370499

Airborne laser scanning (ALS) datasets are used widely for estimating forest biometrics. The transferability of predictive models among ALS acquisitions is a topic of research due to differences in timing, flight parameters, equipment specifications, environmental conditions, and processing methods. The transferability of predictive models therefore is subject to uncertainty. This paper presents an evaluation of the transferability of models for the estimation of stem volume and diameter at breast height (DBH) based on individual tree crown size and competitive neighbourhood metrics derived for managed loblolly pine (Pinus taeda) and slash pine (Pinus elliottii) forest in the Southern USA. Two predictive models types were tested: multiple linear regression (MLR) and Rand Forest (RF). We also evaluated the inclusion of additional training data to model development. Models were able to be transferred to other locations with similar structural and management conditions as the original training dataset with little decrease in accuracy, specifically unthinned stands, despite different ALS acquisitions (Plot stem volume: R 2 0.7-0.8; NRMSE 10-12%; mean DBH: R 2 0.4-0.7; NRMSE 10-17%; plot basal area: R 2 0.7-0.8; NRMSE 12%). Increases in structural differences between the training and test data, driven by age or thinning status, introduced unacceptable levels of uncertainty (Stem volume: R 2 0.4-0.7; NRMSE 12-16%; mean DBH: R 2 0.4-0.5; NRMSE 18-20%; plot basal area: R 2 0.5-0.6; NRMSE 22-40%). Generally, RF models most accuracy estimated DBH, and MLR for stem volume. Improvements to estimate accuracy can be achieved through the addition of relatively small datasets, representing features which were not present in the original data. ALS's ability to provide accurate and near-complete inventories of forests hold a great deal of potential for forest management. The existence of a transferable model that can be used across different acquisitions represents a saving in terms of cost and time, we would argue that future research is therefore warranted. Novel LiDAR-based allometric models were tested using new LiDAR acquisitions; LiDAR predictions were based on tree crown size and immediate neighbourhood metrics; Model inputs were derived from drone and crewed aircraft LiDAR acquisitions; Model predictions were most accurate when applied to similar stand structures as the original model; Normalized root means square error for stem volume and diameter in pine stands with no thinning was <15%; Stem size estimates in thinned stands were consistently overestimated.