High throughput phenotyping in agriculture seeks to automatic the measurement of plant phenotypes by automating aspects of the measurement process. For example, in field grown sorghum, measurements like plant-height and canopy cover, and geometric measurements of leaf length, width and angles, and visual phenotypes that may be measurable automatically. Characterizing these phenotypes over time in large scale field trials opens up opportunities for improving plant-breeding and better understanding of the genotype-phenotype relationships. Attempts to automate these analysis with computer vision pipelines based on classical approaches (edge detection, segmentation and explicit coding for more complicated analysis) such as PlantCV [3] have made many analytics possible, but such approaches often struggle in field conditions.
In this study they approach this problem from another viewpoint. They hypothesize that if a convolutional neural network is trained to differentiate between a large number of different cultivars based on imagery, then it must be learning visual features that capture relevant visual phenotypes. This extended abstract captures our initial experiments in this direction. The researchers propose one approach calculate these features, and show preliminary results that those features can be used to predict the phenotypes and the presence or absence of several common genetic markers. Figure 1 gives a high-level overview of the process.
