A novel image-based approach for soybean seed phenotyping using machine learning techniques

Soybean is important for protein and oil worldwide, requiring investments in research and production technology, mainly breeding programs, to meet the increasing demand. Seed morphology, including area, perimeter, and circularity, significantly impacts soybean yield. However, the analysis of these factors has been hindered by the lack of efficient computational approaches that offer accuracy and flexibility to user preferences and needs. This study presents a soybean phenotyping system framework that includes: seed segmentation in soybean images, morphological evaluation, and image-based prediction of hundred-seed weight (HSW). We used genotypes from a partial diallel cross design and collected red-green-blue images of seeds from each plot. We developed an in-house image processing pipeline for seed segmentation, which enabled a full morphological evaluation of the seeds. For predicting the HSW, we compared machine learning algorithms using the obtained morphological characteristics as input and features from state-of-the-art convolutional neural network (CNN) architectures. Our image segmentation methodology correctly identified over 98% of the seeds in the images, even when they were in proximity. Our morphological phenotyping approach's adaptability to other plant species was verified, fully demonstrating the pipeline's generalizability. The morphological measurements were effective in predicting the HSW with an accuracy of 0.71 predictive ability and mean squared error (MSE) of 3.15 (mean HSW of 19.74). The same results were observed for CNN features, highlighting the efficiency of the morphological measurements. ResNet-50 had the most effective feature extraction, with a mean accuracy of 0.60 (MSE of 4.15). Our study can aid automated phenotyping tool development in research and industry.