Spectra Dimensional Reduction Coupled with Machine Learning for the Detection of Pepper Yellow Leaf Curl Virus

Spectra of Fourier Transform Infrared (FTIR) Spectroscopy have wide wavenumber ranges. The resultant large data set contains noise as well as unwanted information. Dimensional reduction techniques are thus needed to process and highlight important spectral information. This study focuses on optimizing the dimension reduction technique to process FTIR spectra to detect Pepper Yellow Leaf Curl Virus (PYLCV) disease infection in chilli plants. In this paper, Principal Component Analysis (PCA) and Discrete Wavelet Transform (DWT) were used for the dimension reduction. To find out the best technique, the resulting data reduction is processed with three types of learning machines, namely Support Vector Machine (SVM), Linear Discriminant Analysis (LDA), and Random Forest (RF). Here, we split the training data into only 20% and 80% for the testing. From the results of the analysis, it was found that for PCA, the best classification results were dimension reduction with the first five principal components (PCs), which produced the highest score of 79.17% in the LDA and RF techniques. For DWT, level 4 decomposition gives 95.83% accuracy for LDA and RF. The DWT level result was then fed into a feedforward artificial neural network (ANN), which showed that the network could process the input and give the same best accuracy result as LDA and RF. This demonstrates that the best dimensional reduction technique for the classification of PYLCV-infected chilli plants is DWT level 4.