Phenotyping Canola (Brassica napus L.) Agronomic Traits and Estimating Seed Yield Using Unoccupied Aerial Vehicle (UAV)-based Multispectral Imagery

Abstract

Canola (Brassica napus L.), as an important oilseed crop, is widely grown in Canada. In plant breeding, there has been great improvement in genetic techniques, but traditional field phenotyping methods are a limitation in breeding genotype selection for new cultivar development. These conventional phenotyping methods are labour-intensive, time-consuming, and can be subjective and destructive. With the development of phenotyping technologies such as unoccupied aerial vehicles (UAVs) and multispectral sensors, desirable phenotypic traits and seed yield can be estimated digitally. In this thesis, the main objective was to develop an efficient and non-destructive method to estimate canola flowering number, flowering layer depth, canopy height, and seed yield using UAV-based multispectral imagery collected during the entire crop season. Canola field experiments were conducted using 56 diverse Brassica genotypes under diverse environments from 2016 to 2018 in central Saskatchewan. A UAV mounted with a multispectral sensor was used for imagery collection. In the flowering number estimation study, the normalized difference yellowness index (NDYI)-based pixels were significantly correlated with actual canola flower numbers with coefficient of determination (R2) ranging from 0.54 to 0.95 (p < 0.05). Moreover, seed yield could be estimated using cumulative NDYI-based pixels extracted from multi-temporal imagery collected during the flowering stage with R2 up to 0.42 (p < 0.05). In the flowering layer depth and canopy height estimation study, canopy height and flowering layer depth could be quantified using a crop surface model generated from UAV-based imagery with R2 up to 0.90 (p < 0.05) and 0.42 (p < 0.05), respectively. Furthermore, the cumulative UAV-derived canopy height at the flowering stage and the cumulative UAV-derived flowering layer depth were significantly correlated with seed yield (R2 up to 0.46 and 0.34, respectively; p < 0.05). In the last seed yield estimation study, a machine learning method (i.e., random forest regression model) was used to investigate 30 digitalized input variables including 28 cumulative vegetation indices and 2 cumulative canopy structural phenotypes (i.e., cumulative UAV-derived canopy height at the flowering stage and cumulative flowering layer depth) for yield estimation. The ranking of variable importance by the random forest model indicated that the cumulative blue normalized difference vegetation index at the flowering stage, the cumulative UAV-derived canopy height at the flowering stage, and the cumulative normalized difference vegetation index at the vegetative stage were the most important indicators to estimate seed yield in canola. Generally, all results demonstrated that flowering number and canopy structural feature (i.e., canopy height and flowering layer depth) could be efficiently assessed by UAV-based imagery. The digital cumulative phenotypes at the vegetative and flowering stages selected by a random forest regression model can assist crop researchers and farmers in early yield estimation and crop management decisions.