Cocoa beans moisture content prediction using Machine Learning Model based on the color image features

Abstract

The moisture content of cocoa beans is an essential factor in their quality. Modeling it during drying is still problematic due to the wide variation in drying conditions and the wide variation in cocoa bean varieties. This article aims to investigate the possibility of modeling the moisture content of cocoa beans as a function of RGB images features of unshelled cocoa beans. The approach is to extract features, analyze them and then use the most relevant ones to study Machine Learning models. Features are extracted by calculating mean, standard deviation, energy, entropy, kurtosis and skewness of the components of the rgb (RGB normalized), HSV, L*a*b*, YCbCr color spaces without the brightness components. These features are extracted from 4 types of samples, namely 10, 30, 50 and 70 bean samples per image. Features analysis using the F-test and RReliefF methods shows that the features based on the energy and entropy of the components rg, yb, Cr, Cb, a*, b* and h* are fairly relevant for predicting the water content of cocoa beans. However, they are highly correlated. The selected predictors allow the analysis of linear models, such as Ridge Regression (RR), PLS Regression (PLSR) and non-linear models, such as polynomial, Support Vector Regression (SVR) with rbf kernel, and Decision Trees Regression (DTR). Except RR and PLSR, the other models were preceded by a principal component analysis (PCA) to handle the collinearity problem. The non-linear models give good predictions for the training dataset, with coefficients of determination R2 ranging from 0.94 to 0.96 and RMSE from 3.85 to 4.81. However, there is a significant difference between these results and the predictions of the new datasets. RR and PLSR are stable models, but their predictions are less than non-linear ones. It is therefore possible to predict the moisture content of cocoa beans from the features of RGB images.