High-throughput Field Morphological Phenotyping Using UAV-based Proximal Sensing

Plants are affected by complex genome(G)x environment(E)× management(M)interactions which determine phenotypic heterogeneity in breeding populations.In contrast to high throughput genotyping large number of plants cost-efficiently,phenotyping under field conditions remain inefficient,expensive,laborious and time-consuming.Large scale of modern breeding programs comprising thousand plots in multi-location trials make it a big challenge for breeders to efficiently obtain phenotypic information inexpensively.Therefore,developing high throughput phenotyping methods that enable large-scale screening of genotypes in field plays an important role in increasing breeding efficiency and is helpful for researchers and crop breeders to monitor and evaluate phenotypic traits and its response to genetic variation and environmental stresses.Here we developed an unmanned aerial vehicle(UAV)based high throughput phenotyping platform,which is capable of collecting aerial imagery through mounting sensors and extracting phenotypic information through image analysis.And we evaluated effects of flight configurations(i.e.flight height)of the platform on extracting phenotypic information(i.e.ground coverage).The main contents are as follows:1.Estimation of plant heights using UAV survey.We compared a new method to existing algorithms to estimate plant height for a sorghum breeding trial.Images were captured by a RGB camera on an UAV before emergence and near maturity,digital surface models(DSMs)were generated.Two existing methods(i.e.'point cloud' and 'reference ground')and a new method(i.e.self-calibration)were used to estimate ground level and plant height at plot level.UAV-derived plant heights from each method were compared to manual measurements.The self-calibration method obtained the best performance with R2 = 0.63,RMSE = 0.07 m and repeatability = 0.74,which was a similar repeatability to manual measurement(i.e.0.78).The point cloud and reference ground methods achieved lower repeatabilities(i.e.0.34 and 0.38,respectively).For the self-calibration method,we tested different sampling strategies of calibration plots(plant heights of these plots need to be manually measured)to balance estimation accuracy and workload of manual measurement,finding about 30?40 of 1440 plots could obtain precision similar to manual measurement of the entire trial.2.Evaluating effects of flight height of UAV on accuracy of extracting ground coverage from aerial imagery through simulation study.High-resolution reference images of canopies were manually taken in a wheat trial(28 treatments).The image resolutions of original/reference images were gradually degraded into coarse resolutions(26 levels in total)to mimic the image acquisition at different flight heights of UAV.Vegetation segmentation was conducted for these images(reference and simulated images)through a machine learning method and thus ground coverages were computed.Ground coverages of simulated images were compared to corresponding reference ground coverage.Results indicated that,as pixel size increased,ground coverage tended to be underestimated and overestimated when ground coverages were smaller and greater than about 0.5,respectively.The greatest errors were observed when ground coverages were around 0.3 and 0.7.And the minimum pixel sizes to distinguish between two treatments depended on the relative difference between ground coverages of two treatments,and were exponentially increased when difference were greater than the specific values(i.e.about 0.1,0.08 and 0.06 for P<0.01,0.05 and 0.1,respectively).The self-calibration method offers a pragmatic,robust and universal approach to high throughput phenotyping plant height using UA V survey.And the concept of data fusion of manual measurement and UAV survey can provide new insight into data analysis in high throughput phenotyping.By means of theoretical simulation,we evaluated effects of flight height(i.e.pixel size)of UAV-based platform on accuracy of computing ground coverage from aerial imagery.This study provides a guideline to choose appropriate image resolutions and flight plans to estimate ground coverage and other traits related to vegetation segmentation in plant breeding using UAV based platforms.UAV-based high throughput phenotyping can be used in plant breeding to characterise phenotypic information of large number of lines with given accuracy to improve breeding efficiency and accelerate delivery of new cultivars.