Evaluation of 3D point cloud-based models for the prediction of grassland biomass

Farmers, as well as agronomists, are intrigued by efficient quantification of grassland biomass at field-scale. Canopy surface height (CSH) based non-destructive grassland biomass estimation over a larger area has important advantages compared to destructive methods. 3D point clouds generated from remote sensing (RS) data offer a systematic methodology to derive CSH and estimate grassland biomass. This study evaluated 3D point clouds derived from a terrestrial laser scanner (TLS) and an unmanned aerial vehicle (UAV)-borne structure from motion (SFM) approach for grassland biomass estimation over three grasslands with different composition and management practice in northern Hesse, Germany. TLS data, UAV-borne images and reference biomass data were collected two days before each harvest from the selected grasslands in 2017. Three levels of linear empirical models (i.e. sampling date-specific, grassland-specific and global) were developed to estimate grasslands fresh and dry biomass using CSH derived from TLS and SFM 3D point clouds. The aforementioned three level models resulted in an average nRMSE of 17.2%, 25.3%, and 28.7% respectively for the grassland dry biomass estimation based on CSH from TLS data. Similarly, models based on CSH from SFM data estimated dry biomass with somewhat higher average nRMSE of 19.5%, 27.1%, and, 36.2% respectively. In general, models with 3D point clouds from UAV-borne SFM was slightly outperformed by models with TLS point cloud data. This study also identified that the utilisation of UAV-borne SFM developed digital terrain model as a reference layer to derive CSH could enhance the performance of the models with SFM data. Furthermore, the performance of the biomass estimation models was affected by both species richness and sward heterogeneity of the grasslands. Overall, these results disclosed the potential of 3D point cloud derived from RS for estimating field-scale grassland biomass.