Optimization of Raman Spectra Peak Fitting for Oil Palm Classification

Abstract

The selection of fresh fruit bunches (FFB) plays an important role to ensure the quality and quantity of palm oil. Current harvesting method relies on the experience of the harvester and by counting the number of fruitlets that have loosely fallen on the ground. A rapid and non-invasive method such as the Raman spectroscopy, that can determine the ripeness of the fruits can serve as an assisting tool during pre-harvesting. The challenge lies in the complex signal processing methods during the pre-processing of the raw Raman spectrum which highly relies on human expertise. Thus, this study aims on developing an automated and optimized signal processing algorithm that could eliminate the need of an experienced personnel in processing the raw spectra which mostly relies on experience and intuition. The process involves peak selection, noise smoothing using Savitsky-Golay filters, interpolation and peak fitting. This article focuses on optimizing the Raman spectra peak fitting during pre-processing before significant features are extracted from the peaks to be fed as input into the machine learning model. Three profiles for peak fitting that have been applied are Gaussian, Lorentzian and Voigt. The findings show that after optimization, the best fit percentage for Gaussian profile is 33%, Lorentzian 42% and Voigt 24% which is in line with the manual peak fitting method where Lorentzian profile dominates the best fit. This result shows that the Lorentzian profile or its derivative could be used as the dominant profile for automating peak fitting during raw Raman signal pre-processing.