Multispectral Vegetation Indexes and 3D Image Analysis for Aboveground Biomass Estimation in Meadows: A Comparison between Models

The implementation of remote sensing technologies for agricultural applications has undergone continuous growth in the last couple of decades. Aboveground biomass (AGB) estimation for grasslands and pastures using remote sensing technologies and machine learning is gaining significant interest throughout the scientific community. In this study, we investigated and compared the use of vegetation indices (NDVI, GNDVI, SR, NDRE) calculated from multispectral images and volumetric roughness values derived from a Digital Elevation Model (DEM) obtained with the Structure from Motion (SfM) photogrammetric technique estimating AGB. We carried out this study in two adjacent permanent meadows located in Schio (VI, northeastern Italy) whose total surface was 3.29 ha. We trained two models using the Random Forest (RF) machine learning algorithm to estimate AGB over the study area, the vegetation indices-based model performed better (R2 = 0.84) but the results obtained by the model based on DEM roughness were very close (R2 = 0.82). The results obtained show an excellent potential of new analysis techniques based on Structure from Motion models, which are cheap, widely known, and do not need multispectral cameras. However, many characteristics of the vegetation, such as the botanical composition of the meadow, reflecting in its density, strongly influence the biomass content, and the use of parameters based on volume and external texture may limit the accuracy of this model.