Prediction of Firmness of Sweetpotatoes using VNIR Hyperspectral Imaging and Machine Learning

Sweetpotatoes are used in a variety of ways and are regarded as a commercially important agricultural product due to their exceptional nutrition. Consumer perception is greatly influenced by the sweetpotato's physicochemical characteristics, such as firmness. The traditional methods for measuring firmness are arduous, time-consuming, and destructive. Recently, hyperspectral imaging coupled with machine learning has been considered an innovative approach for quick and non-destructive analysis of agricultural and biological products. This study used both linear [partial least squares regression (PLSR), multiple linear regression (MLR)] and non-linear (support vector regression (SVR) machine learning-based regression methods to predict the firmness of sweetpotatoes of different varieties using spectral data extracted from the images collected using a portable visible near-infrared hyperspectral imaging system (400-1000 nm). Important feature wavelengths were identified using the recursive feature elimination (RFE) technique to show the pixel-wise distribution of firmness of the sweetpotatoes and to aid the development of a low-cost multispectral system for industrial application. The results of the predictive analysis suggested that visible near-infrared hyperspectral imaging technology and machine learning could provide a quick and non-invasive prediction of sweetpotato firmness.