High Throughput Plant Phenotyping in the Field: Machine Learning Enabled by Data from Crowdsourcing

We present a high throughput computing workflow for the automated extraction of phenotypic traits from digital images of growing maize plants in a large scale field experiment. This framework enables automated data collection and curation towards answering questions on tassel morphology and the progression of anthesis. We discuss a data processing framework that consists of a crowdsourcing-enabled, machine-learning algorithm for image registration that isolates tassels from the rest of a cluttered image. Crowdsourcing provides a scalable, distributed and statistically consistent approach of acquiring labeled data to train machine learning algorithms for plant phenotyping. We report on trends related to training, intensity and length of crowdsourcing tasks with the ensuing quality of labels as well as the performance of the image registration algorithm trained using this labeled data. This work provides potential best practices as applied to crowdsourcing for plant phenotyping. Finally, we discuss the downstream analytics pipeline that includes (1) a graph-based image processing framework to extract features from the tassel images (2) a work distribution system that deploys the above framework to extract features from individual images using a massively parallel approach. This is a collaborative work by an interdisciplinary group that includes psychologists, statisticians, engineers, computational biologists, and plant scientists.