Document Type

Dissertation - Open Access

Award Date


Degree Name

Master of Science (MS)

Department / School

Agronomy, Horticulture, and Plant Science

First Advisor

Melanie Caffe


Above ground biomass, machine learning, oats, structure features, UAV, VI


Conventional methods of phenotyping for above-ground biomass under field conditions are time consuming, costly, and labor intensive. Unmanned Aerial Vehicles (UAV) offer a rapid and non-destructive approach for collecting high temporal and spatial resolution imagery of multiple field plots. Vegetation indices (VIs) can be derived from those imagery and have been used broadly for high-throughput biomass estimation, however, they suffer from spectrum saturation issues at high canopy density. In addition, VIs only reflect the canopy spectral information and fail to incorporate the 3-dimensional (3D) canopy structure and architecture information. It also ignores the spatial relationship of pixels to each other within an image. These limitations can be filled by incorporating canopy structure and texture information. No studies have utilized all three types of features for predicting oat biomass. Two oat field trials were conducted to investigate the potential of using spectral, structure, and texture features from UAV collected multispectral imagery to predict oat biomass. The first trial was set up to develop the prediction model (using only six oat genotypes planted at two seeding rates and harvested at three growth stages). The second trial consisted of breeding materials (35 different oat genotypes) and was used to validate our findings for use in breeding. High-resolution aerial images were acquired prior to each harvest using the Sentera multispectral sensor attached on a DJI Phantom 4 pro UAV. Plot-level canopy spectral, structure, and texture features were extracted from multispectral imagery and fed into three machine learning models: (1) Partial Least Squares Regression (PLSR), (2) Support Vector Regression (SVR), and (3) Random Forest Regression (RFR). Three types of extracted features were used individually and in different combinations as input variables, and a comparison was made among models. For the first trial, results showed that: (1) in addition to canopy spectral features, canopy structure and texture features are important indicators for oat biomass prediction; (2) fusing spectral, structure, and texture features improved biomass prediction accuracy over using single features; and (3) machine learning algorithms showed good predictive ability with slightly better prediction accuracy for RFR models. However, while prediction models with high accuracy (R2 = 0.9) were successfully developed using the first trial, poor validation accuracy was obtained when these models were validated using the breeding trials. Overall, this study demonstrated the benefits of UAV multispectral imagery-based multiple features fusion using machine learning algorithms in the context of biomass estimation in oats. However, significant limitations of model transferability were observed when considering actual breeding nurseries. Inclusion of data from multiple years and using a bare-ground elevation model for calculating canopy height metrics should be considered in future studies for estimating biomass in oats.

Number of Pages



South Dakota State University

Available for download on Thursday, August 15, 2024



Rights Statement

In Copyright