Maize biomass estimation using structure from motion data and volumetric approaches

Dissertation submitted in partial fulfilment of the requirements for the Degree of Master of Science in Geospatial TechnologiesPlant phenotypic traits such as biomass work as predictors of important biological outcomes like fitness, disease, and mortality. Plant biomass is an essential parameter for crop management, growth monitoring, and yield estimation. Traditionally, the manual approach known as destructive sampling is the most accurate technique to estimate biomass. However, the large scales of modern agricultural research or production schemes are turning this approach into an impractical technique. Remote Sensing (RS) is a technology that has been applied to facilitate this task. Light Detection and Ranging (LiDAR) is a popular RS tool for evaluating and understanding plant canopy structure due to its accuracy and ability to build 3D point clouds, but its high cost currently limits its application. Structure from Motion (SfM) is a low-cost alternative to LiDAR that involves acquiring images using a digital camera from multiple positions to generate a 3D point cloud similar to that which is produced with LiDAR. In this study, SfM point clouds derived from Unmanned Aerial Systems (UAS) were used to build volumetric models evaluating and comparing two different methodologies for estimating plant biomass in maize (Zea mays L.); voxel-counting and convex-hull approach. The voxel-counting approach works by encapsulating the point cloud into volumetric pixels to generate a voxel grid. The volume is estimated by counting the number of voxels with at least one point inside. The convex-hull approach splits the point cloud into two parts and progressively calculate the extreme points to generate a polygon that encompasses the full point cloud. The result of this study showed that volumetric models based on SfM point cloud data are suitable for estimating maize biomass combining with volumetric models. Voxel-counting was more accurate predicting maize biomass (2=0.973) than the convex-hull approach. SfM point cloud data coupled with the voxel-counting approach offers a low-cost alternative for providing accurate biomass estimations