Estimation of different ripening stages of Fuji apples using image processing and spectroscopy based on the majority voting method

Abstract

Non-destructive determination of the different stages of fruit ripening has important advantages over traditional methods, such as selective robotic harvesting or adapting fertilization operations depending on the ripening stage. In this regard, the purpose of the present study was to investigate the non-destructive estimation of the ripening stages of Fuji apples combining different classifiers with the majority voting (MV) method. This process is based on five constituent classifiers, including hybrids of artificial neural network (ANN) classifiers adjusted with the genetic algorithm, the particle swarm optimization algorithm and the firefly algorithm, and classifiers based on support vector machines, and the k-nearest neighbor algorithm. The input of the MV classifiers consists of four alternatives: (1) color data extracted from the second channel of L*a*b* color space, and the hue angle in L*a*b*; and multispectral data including wavelengths ranging: (2) from 465 to 485 nm; (3) from 675 to 700 nm; and (4) from 870 to 890 nm. The first two ranges are in the visible spectrum, while the second is within the near-infrared. To evaluate the reliability of the MV method, the classification procedure was repeated 1000 times with different seeds. In order to assess the obtained performance, the proposed method has been compared with an alternative technique based on an ANN classifier, in this case using all the spectral data in the range from 450 to 1000 nm, and with the hyperparameters adjusted by a grid search. The results indicate that the correct classification rate of the MV method using color data, and using spectral data from 465 to 485 nm, 675 to 700 nm, and 870 to 890 nm were 95.12%, 99.37%, 97.56% and 97.80% respectively, while the correct classification rate of the ANN method including all the spectral data from 450 to 1000 nm reached an average classification of 92.12%. Thus, the optimal selection is the MV method using spectral information from 465 to 485 nm, which is able to achieve a very accurate result, feasible to be used in practical applications.