X-ray Imaging Assessment of Internal Seed Morphology as a Nondestructive Viability Prediction for Triploid Watermelon Seeds

Abstract

Highlights X-ray imaging techniques were used to assess the internal morphology of triploid watermelon seeds. Structural integrity of triploid watermelon seed was quantified through image-processing and analyzed according to multiple viability classes. Integrity and CNN-based viability prediction models were developed and evaluated for multiple viability criteria. In the integrity analysis and modeling results, there were differences in the correlation between internal seed morphology and viability depending on the condition of the seed lot. Abstract. Watermelon (Citrullus lanatus) is a tropical fruit consumed worldwide in various forms. Triploid watermelons—or seedless watermelons—have remained popular for decades because of the absence of hard seeds and their flavor. However, triploid watermelon seeds have lower viability than diploid watermelon seeds because of their thick seed coats, underdeveloped embryos, and larger internal cavity spaces. This poor viability characteristic of triploid watermelon seed leads to low crop productivity. Therefore, a nondestructive inspection technology is deemed necessary for sorting triploid watermelon seeds. In this study, we assessed the internal morphology of triploid watermelon seeds by applying the X-ray imaging technique to predict seed viability. More specifically, we analyzed the association between the structural integrity and viability of the seeds by X-ray image processing. Furthermore, prediction models based on integrity and convolutional neural networks (CNN) were developed and evaluated for multiple viability criteria and seed lots. As a result, first-grade class seeds were shown to significantly differ from the rest of the classes in terms of integrity. Similarly, the performance of classifying the first-grade class from other classes was the highest among classification criteria in prediction models. Although the CNN model showed better performances than the integrity-based model, seed integrity was considered to be the most important feature even in the CNN model. The CNN model in this study showed accuracies of 73.64%–90.63% depending on the seed lot, suggesting that the correlation between seed internal structure and viability may differ depending on the conditions of the seed lot. Keywords: Deep learning, Seed, Seed integrity, Triploid watermelon, Viability, X-ray.