By Darell van der Voort (the Netherlands)
Quantitative analysis of plant phenotypes in field trials has become a major bottleneck due to the large amount of breeding trials that need to be scored and analyzed. This phenotyping is traditionally done by visual scoring and manual measurements, which is labor intensive, non-systematic and susceptible to human error. Plant breeding would therefore benefit from fast, high-throughput and non-destructive sensing techniques to score and evaluate field-based breeding trials. In controlled circumstances like greenhouses significant progress has been made using image based techniques in high-throughput phenotyping facilities. The next step would be to adopt camera based technology in field experiments, however, a multitude of factors are influencing these observations: light conditions, wind, moisture on leaves, etc. Both ground based vehicles and aerial platforms have been considered for field-based phenomics. Aerial platforms like Unmanned Aerial Vehicles (UAVs) have much potential to be a suitable tool for plant breeders. They can be equipped with multiple sensors and fly at low-altitude in a relative small timespan without affecting field conditions.
In this research we evaluated the usability and accuracy of high-resolution hyperspectral images acquired from an Unmanned Aerial Vehicle (UAV) to quantify the growth parameters biomass and plant height for a large-scale maize breeding trail. Furthermore, this research tested which spectral range and growing stages gave the best relation with biomass and plant height in order to find out which sensors were most suited and what flying periods were essential for phenotyping these traits. The field under research consisted of a phenotyping experiment with 3838 plots with a size of 10.2 m2 each and planted with different maize (Zea Mays) cultivars. Conventional measurements for biomass and height of each cultivar were done at harvest. At four different growth stages, flights with a multi-rotor UAV platform equipped with a combined RGB and hyperspectral camera we carried out, resulting in three geo-rectified products: RGB-orthomosaic, digital surface model and a hyperspectral dataset with 100 bands ranging from 450-950 nm. In the subsequent method development, both individual spatial images and also the fusion (including time) was evaluated using vegetation indices and multivariate statistical approaches.
For the spatial estimation of height at the end of the growing season, a good accuracy could be achieved when correlating UAV derived heights to LiDAR measurements (R2 = 0.78). Accuracy over the whole field differed due to geometrical errors and influence of ground control points (R2 = 0.61). Overall there was a systematic underestimation of -0.30 meters. For the estimation of final biomass, the prediction accuracy was R2 = 0.79 when the spatial layer of the canopy plant model was combined with spectral information from specific parts of the growing season in a multivariate approach. The multispectral indices showed a better performance in estimating biomass than the hyperspectral indices. The beginning and end of the growing season were most important for biomass prediction. The main limitations of this research were the geometrical distortions caused by the hyperspectral push broom scanner and non-linear deformations of the digital surface models due to the lack of sufficient amount of ground control points. The results of this study have shown that crop surface models derived from the UAV based RGB sensor are capable of capturing height variability with high spatial detail. Furthermore, our results suggest that plant height measured at harvest, combined with vegetation indices at the start and end of the growing season may be used to predict maize biomass. Future research should explore the use of more ground control points for improved rectifications of 3D models and the use of multi- or hyperspectral frame cameras in order to minimize geometrical distortions.
Keywords: Unmanned Aerial Vehicle; Remote Sensing; Phenotyping; Breeding Trials; Maize; Structure from Motion; Hyperspectral; Plant Height; Biomass; Multivariate Analysis.