Using Deep Learning to Develop SNAP, The Soybean Nodule Acquisition Program

The Soybean Nodule Acquisition Program, or SNAP, is a process designed to identify and count the nodules on a 2D soybean (Glycine max) image. Soybean nodules are formed by a symbiotic relationship with the host plant, and nitrogen-fixing rhizobia bacteria (Bradyrhizobium japonicum). In this relationship, soybeans provide simple sugars to the bacteria which in return fixes atmospheric nitrogen into an ammonia form for the soybeans to use. Additional knowledge and information on the amount and quality of nodule formation can be implemented in breeding programs and genetic improvements to develop more efficient soybean cultivars. The current state of the art is manual identification and counting each unique nodule by visual assessment; and therefore counting nodules on soybean roots takes large quantities of time and human resources, and human fatigue introduces inadvertent rater variability.

We propose the SNAP, which uses computer vision and deep learning to identify nodules from a robust data set of 2D soybean roots from the field from various growth stages in order to count the number of nodules per root. Using 1400+ root images collected from a diverse set of soybean accessions, data were separated into training, validation and test sets. We manually annotate all the nodules in the training data to build a Faster R-CNN based algorithm for nodule detection. The robustness and accuracy of the SNAP program was determined on the test set.

The overall goal of the SNAP program is to determine root regions with most dense nodules for its exploitation in soybean genetic enhancements programs. Using maps of nodule locations based on skeletonized roots, further studies will be conducted to compare nodulation density and root location across a diverse panel of genotypes to understand how the location of nodules impacts the roots and productivity of soybean plants.