Mango Internal Defect Detection using NIR Spectroscopy Data

Near-Infrared (NIR) Spectroscopy is a crucial non-destructive technique for the internal defect detection of mangoes in effectively identifying spongy tissue. In this study, we are proposing a method for classifying mango spectroscopy data into two classes such as internal defective and healthy mangoes. The proposed method uses savitzky-golay (savgol) filter as a pre-processing step for smoothening the wavelength or spectroscopy data because it effectively smooth’s noisy data by computing derivatives that enhance critical features which are crucial for further processing. Then we have adapted partial least square discriminant analysis (PLS-DA) with cross decomposition for the effective reducing dimensionality by transforming the original data into a lower-dimensional space, as spectroscopy data typically involves high-dimensional features due to the large number of wavelengths measured and they suffer from multi-collinearity problem. These issues are addressed by increasing the correlation among the spectral data (predictors) and the defect classification (response variable). This approach helps in not only preventing overfitting, but also ensuring that the model generalizes well to new unseen data. After determining new axis or principal components that sum ups the variations in the data, learning algorithms such as random forest (RF), support vector machine (SVM), logistic regression (LR), are applied for the purpose of classification. For experimental analysis the dataset of 76 mango samples of the spectroscopy data is considered [3] and achieved an average accuracy of 93.78% using SVM, 93.72% with Logistic Regression and 93.42% with Random Forest for the lower range wavelength. This study also presents the results for the higher range wavelength. Moreover, other dimensionality reduction methods such as LDA (Linear Discriminant analysis) and PCA (Principal Component analysis) are compared against proposed method. An extended analysis is also done with advanced feature selection techniques such as recursive feature elimination (RFE) and sequential feature selection (SFS) methods for selecting optimal features and classified with average accuracy of 91.16% and 84.61% for the lower range wavelength and 79.310% and 81.39% for higher range wavelength. The empirical results shows that, the proposed technique has a superior accuracy over the existing contemporary state-of-the-art model.