UAV-based imaging platform for monitoring maize growth throughout development

Abstract

Plant height (PH) data collected at high temporal resolutions can give insight into how genotype and environmental variation influence plant growth. However, in order to increase the temporal resolution of PH data collection, more robust, rapid, and low-cost methods are needed to evaluate field plots than those currently available. Due to their low cost and high functionality, unmanned aerial vehicles (UAVs) provide an efficient means for collecting height at various stages throughout development. We have developed a procedure for utilizing structure from motion algorithms to collect PH from RGB drone imagery and have used this platform to characterize a yield trial consisting of 24 maize hybrids planted in replicate under two dates and three planting densities. PH data was collected using both weekly UAV flights and manual measurements. The comparisons of UAV-based and manually acquired PH measurements revealed sources of error in measuring PH and were used to develop a robust pipeline for generating UAV-based PH estimates. This pipeline was utilized to document differences in the rate of growth between genotypes and planting dates. Our results also demonstrate that growth rates generated by PH measurements collected at multiple timepoints early in development can be useful in improving predictions of PH at the end of the season. This method provides a low cost, high throughput method for evaluating plant growth in response to environmental stimuli on a plot basis that can be implemented at the scale of a breeding program.

Keywords: UAV phenotyping; maize; plant height.

FIGURE 1

Procedure for feature extraction from UAV images. (a) Image generation and processing pipeline. (b) UAS flight mission structure for gathering images. (c) Dense cloud (top), orthomosaic (middle) and DEM (bottom) output from Agisoft Photoscan Professional Edition. (d) Plot boundary extraction from grid shapefile for plot segmentation. UAV, unmanned aerial vehicle

FIGURE 2

Method for extracting mean PH_UAV values for a given plot. The plot is segmented and broken down into 20 bins. The 3rd and 97th percentiles for each bin are extracted and used to calculate the average plot height by subtracting the minimum 3rd percentile of all bins from the 97th percentile of each bin

FIGURE 3

Pearson correlation of mean plot PH_UAV and mean plot PH_R(2) across all dates

FIGURE 4

Adjusted r‐square values, root mean square error, and normalized root mean square error by mean for the linear correlation of various PH measurements. (a) UAV‐derived (PH_UAV) plot mean height values compared to the respective hand‐measured plot mean height (PH_R(2)) for plots where two plants per row were hand measured. (b) UAV‐derived (PH_UAV) plot mean height values compared to the respective hand‐measured plot mean height (PH_R(All)) for plots where all plants for one of the middle two rows were hand measured. (c) Means derived from two iterations of random sampling of hand‐measured height values for two plants across plots were all plants for one of the middle two rows were hand measured. (d) 95th percentile of height values from entire plot compared the respective hand‐measured plot mean height values (PH_R(2)). PH, plant height; UAV, unmanned aerial vehicle

FIGURE 5

Height through time for various genotypes and treatments. (a and b) Height through time as measured by the UAV (a) and hand measurements (b) for two replicate plots of a single genotype (LH82 × DK3IIH6) planted under high density across two planting dates. Dots indicate timepoints of measurement. (c) Height through time as measured by the UAV for two replicate plots of a single genotype (LH82 × PHK76) in the early planting date treatment and the three planting densities. UAV, unmanned aerial vehicle

FIGURE 6

Predicting terminal height using rate of change in height at certain intervals. (a) Height through time for an individual plot as recorded by the UAV. (b) First derivative of the height through time spline curve in blue and the selected timepoints were the slope contributed most to the prediction of terminal height in yellow. (c) Correlation of predicted terminal height on the test dataset utilizing a linear regression model derived from the training dataset based on slope values of selected timepoints and the hand‐measured height values at maturity. UAV, unmanned aerial vehicle