Assessing the Performance of a Machine Learning System to Predict Geometrical Properties of Ahmad Aghaei Pistachio Kernels

The use of machine learning techniques such as artificial neural networks (ANN) improves the performance and speed of prediction processes as well as their reliability in the design of agricultural processing machines. Machine learning as a subset of artificial intelligence makes it possible to develop a unique way to create a predictive model system in the form of a known dataset by developing machine learning models (MLM).

In this study, first the geometric properties of pistachio kernels including the major diameter (L), intermediate diameter (T), minor diameter (W), geometric mean diameter (Dg), and surface area (S) were calculated at four moisture levels of 4.33, 22.64, 29.11, and 41.35% (w.b). Then, the data obtained in this step were used as the input values (L, W & T) and the output value (S) into the machine learning system. Multi-layer perceptron (MLP) and radial basis functions (RBF) were used as two machine learning models to predict the surface area of pistachio kernel during rehydration.

The data analysis revealed that the neural network model of RBF with 42 neurons in the hidden layer (N1st=42) had the lowest mean relative error (MRE=0.01414), and the highest coefficient of determination (R2=0.954) and chosen as the best model for predicting the surface area of pistachio kernel.

Following the findings of this study, it can be concluded that the MLM as one of new forecasting techniques can be used to estimate the engineering properties of agricultural products.