Sweetpotato moisture content and textural property estimation using hyperspectral imaging and machine learning

Abstract

Quality parameters, such as moisture content (MC) and textural properties, are critical indicators reflecting the quality of fruits and vegetables and significantly influence their shelf life. Monitoring these parameters of agricultural products during post-harvest processing, drying, and storage is crucial for ensuring product quality, safety, and cost-efficiency. This study proposes the integration of machine learning (ML) algorithms with hyperspectral imaging (HSI) to effectively estimate the moisture content and texture characteristics like firmness and consistency of sweetpotatoes. In light of this, orange-fleshed and purple-fleshed sweetpotato samples were imaged using a hyperspectral camera with a spectral range of 400–1000 nm. The extracted spectral data underwent preprocessing to select key wavelengths, construct various models, and compare their accuracy and efficacy. The multiplicative scatter correction-competitive adaptive reweighted sampling-radial basis function (MSC-CARS-RBF) model (RMSE = 0.066%, R² = 0.97) demonstrated superior performance for the moisture content prediction, while the standard normal variate-competitive adaptive reweighted sampling-extreme learning machine (SNV-CARS-ELM) model showed the best predicting results for the texture characteristics. The results indicated that selecting key wavelengths can enhance the predictive ability for sweetpotato quality assessment. Furthermore, this study demonstrates that combining HSI with ML algorithms have the potential to improve the quality assessment of sweetpotatoes by enhancing the accuracy, consistency, and speed of evaluating moisture content and firmness, ensuring uniformity in grading, and enabling near-real-time, non-destructive assessment during handling and processing, thereby ensuring a higher quality product for consumers.