Optimising prediction of forest leaf area index from discrete airborne lidar

Abstract

Discrete airborne laser scanning (ALS) data has emerged as a useful tool for mapping forest leaf area index (LAI). Both empirical and physically-based approaches linking pulse penetration to LAI through gap probability theory have been widely used. We contrasted these approaches using field measurements of LAI (n=135) acquired in stands of pure Pinus radiata D. Don in New Zealand. For the empirical approach, we addressed several methodological questions: (1) Identification of important covariates from an extensive list of lidar metrics with empirical or theoretical links to LAI. (2) Evaluation of the impact of lidar plot radius on metric importance and model accuracy by trialling fixed radii and radii based on mean top height (MTH). (3) For ratio metrics, which require selection of a height threshold (HT), identification of the optimum fixed HT and to evaluate a novel variable HT set as a percentage of canopy height. Custom lidar software evaluated all combinations of metric, radius, and HT. For model development, we tested elastic net linear regression for regularisation and variable selection, as well as random forests to explore potential nonlinear relationships and to provide insight into variable importance using conditional importance scores accounting for intercorrelation. For the physically-based model, a proxy for vertical canopy gap fraction was sought from ALS metrics measuring pulse penetration for use in a nonlinear model based on the Beer-Lambert law. Empirical models were strongly impacted by calibration and larger plot radii and higher HTs generally reduced RMSE and highlighted a common set of ratio metrics characterising pulse penetration. Elastic net models performed best with the lowest RMSE=0.57 LAI at radius=100% of MTH and HT=20% of canopy height. Models with low RMSE often had radii in the range of canopy height - supporting theoretical links to instrument view distance. The best fixed-radius model (RMSE=0.64) had radius=20m and HT=20% of canopy height. Random forests results were similar, with little evidence of nonlinear relationships (lowest RMSE=0.64). Physically-based models produced results close to the best calibrated empirical models (RMSE=0.72) using a single metric. This approach offered the potential to estimate forest type coefficients that could allow ALS-LAI mapping without the need for calibration and with greater potential for transferability between ALS campaigns. These results support the use of physically-based models for discrete ALS-LAI mapping.