Modeling Miscanthus Biomass from UAS-LiDAR Data Using Machine Learning

Abstract:

Biomass crops such as Miscanthus (Miscanthus x giganteus), a highly productive C4 grass, are being used for feed, fiber, and fuel in the United States and Europe. Estimating biomass in Miscanthus fields is therefore, of great interest for field management. Light detection and ranging (LiDAR) is an active remote sensing technique, and has been proved to be capable of estimating forest biomass accurately. However, its capability to estimate Miscanthus biomass has been rarely investigated. Here, we used multi-temporal unmanned aerial system (UAS)-LiDAR data to quantify the three-dimensional physical characteristics of Miscanthus and evaluate the potential of using LiDAR-derived physical attributes to estimate miscanthus yield. Results showed that the yield of Miscanthus is best predicted when combining the variability of Miscanthus height and density at the early growth stage, capturing the variability in stand density across the field, and average height values in the late growth stage when canopy closure had occurred masking variability in plant density. Compared to stepwise regression, the random forest model had lower predictive power. The results of our study provide guidance in selecting LiDAR features for Miscanthus yield estimation. The experimental design and the methods used in this study could be applied to other grass yield estimations using LiDAR data in general.