Visible-near infrared spectrum-based classification of apple chilling injury on cloud computing platform

This paper evaluates the feasibility of applying cloud computing technology for spectrum-based classification of apple chilling injury. The reflectance spectra of Fuji apples with four different levels of chilling injury (none, slight, medium, and severe) were collected. During data processing, the spectra at 400-1000 nm were selected, and first- and second-order-derivative spectral data sets were obtained through integral transformations. Five optimal wavebands were chosen as inputs for the classification models. A cloud computing framework based on Spark and the MLlib machine learning library was used to realize multivariate classification models based on an artificial neural network (ANN) and support vector machine (SVM). The ANN and SVM classification models were used for multivariate classification and analysis of the spectral data sets (raw, first derivative, second derivative) and corresponding optimal wavebands. Of the total data samples, 70% were used for training, while the remaining 30% were used for prediction. The experimental results showed that, by using the cloud computing platform, we could establish an efficient spectrum classification model of apple chilling injury; the ANN model had slightly higher accuracy than the SVM model (not including the second-derivative spectra), but the SVM model was more efficient. Moreover, the classification accuracy using full-waveband spectral data sets was higher than that of data sets using five optimal wavebands. Furthermore, the Spark framework and MLlib were used to implement binary classification models (decision tree and random forest), and these were compared with the multivariate classification model; the binary classification method had better performance in near-infrared spectrum-based classification of apple chilling injury. Finally, we extended the existing spectrum data set to verify the efficiency of the cloud computing platform and desktop PC for handling larger data sets. The results showed that the efficiency of the cloud computing platform was significantly improved by increasing the spectral data set capacity or number of working nodes. Owing to processor and memory limitations, the classification algorithm and model of abundant spectral data sets cannot complete all of the tasks on a desktop PC.