Using hyperspectral imaging technology and machine learning algorithms for assessing internal quality parameters of apple fruits

The important internal quality parameters of apples are firmness and soluble solids content (SSC); features that affect consumers' preferences and the marketing of apples. Hyperspectral imaging systems are novel and nondestructive methods that have various applications in the food industries. The regression methods offer many advantages due to the learning capability, non-destructive measurements, reduced assumptions, process compliance, and tolerance of missing data in agricultural commodities. In the present study, hyperspectral images of Pink Lady apples at different harvest stages (three periods) were analyzed to predict some internal characteristics (firmness and SSC). The hyperspectral camera was used to acquire reflectance data in 300 spectral bands in the range of 386 and 1028 nm in a total of 100 samples for each harvest period. In addition, prediction performance of artificial neural network (ANN), k-nearest neighbors (KNN), decision tree (DT), partial least squares regression (PLSR) and multiple linear regression (MLR) were evaluated. Determination coefficients (R2) were respectively determined as 0.910 and 0.684 for ANN, 0.881 and 0.679 for DT, 0.781 and 0.684 for KNN, 0.666 and 0.762 for MLR and 0.819 and 0.661 for PLSR in the prediction of firmness and SSC. Besides, better firmness prediction was achieved by using spectral bands of 511, 505, 704, and 689 nm. The present study demonstrated the potential use of hyperspectral imaging with ANN and DT methods was more effective for firmness, while DT and MLR were more effective for SSC and the methods were proved to be quite feasible for industrial applications.

Automated summary

The present study analyzed hyperspectral images of Pink Lady apples at different harvest stages to predict firmness and soluble solids content. The results demonstrated that hyperspectral imaging combined with artificial neural network and decision tree methods were more effective for firmness prediction, while decision tree and multiple linear regression methods were more effective for soluble solids content prediction.