Image processing and machine learning identify high-yield branching phenotypes in soybean

A challenge for precisin agriculture is developing automated computer methods to accurately estimate fruit and seed yield in the standing crop. Soybean (Glycine max (L.) Merr.) pods are hard to distinguish from stems, which causes inaccurate predictions of yield from images of mature shoots. We developed image analysis tools to estimate morphological traits in the vertical canopy profile that are associated with high seed yield in soybeans. Using common image processing methods involving thresholding and particle analysis, higher circularity of the shoot convex hull vertical profile was found to correlate with high seed yield (number and grams per plant) in both an indeterminate cultivar (P49T80R) and in a determinate cultivar (Glenn). These soybean cultivars achieved high yields using different growth and production strategies. Glenn had a smaller shoot but exhibited a high pod density phenotype throughout its canopy (PT1, where PT stands for phenotype), while P49T80R achieved high yield through a combination of increased height and greater branching width, which compensated for lower pod density in its branches (PT2). We trained a deep machine learning model to automate shoot phenotyping using nearly 400 images of soybean shoots. The resulting model distinguished between PT1 and PT2 shoot images with 80% overall accuracy. The highest prediction accuracy in the model, 95%, was attained for shoots exhibiting the PT2 phenotype. Our work illustrates real-world application of image analysis technologies to identify high-yield trait analysis in field-grown soybeans and emphasizes the importance of including pod density positioning in machine learning training models.