Modeling the Relationship between Root Color, Root Shape, and Weight of Radish using Machine Learning

Abstract

The quality of radish depends on the production environment and postharvest management during a supply chain. Quality monitoring is therefore important for post-harvest management in the supply chain. This study aims to estimate the quality of radish non-destructively based on color and shape information using random forests as a tool of predictive data-driven modelling. The explanatory variables, namely color and shape information, were obtained by capturing images of radish under a controlled shooting environment. Color information was converted from RGB to HSL or HSV for minimizing potential effects of light conditions on an object surface. Model performance was assessed using Pearson’s correlation coefficient (COR), Nash-Sutcliffe efficiency (NSE), and root mean squared error (RMSE). Experimental results indicated high model performance, supporting the applicability of nondestructive weight estimation using color and shape information of radish. Among the models using different color components, the HSV model exhibited the best performance in all performance measures in this study (i.e., COR, NSE, and RMSE were 0.889, 0.776, and 1.55, respectively). Since the radish variety in this study was red, the R value was the most important variable among the color information. Partial dependence plots further visualized the relationships between each pair of the color component and the radish weight. Further study is needed for the application of this method for other photographic environments specifically where sunlight is present. Assessment of internal conditions of fresh radish such as vitamin C and anthocyanin could be useful for consumer-oriented quality assessment.