Comparison of Neural Network Techniques and Multi-Linear Regression to Predict Properties and Classify Pepper Seeds

Abstract

Pepper seed quality is determined using the mechanical and physical properties through artificial neural networks (ANNs) to enable accurate and timely agricultural planning. The objective of this study is to develop a model that provides simple, precise, rapid, and cost-effective predictions based on thousand-grain weight, porosity, and various classifications for pepper seeds. To achieve this, three different models-artificial neural networks (ANN), radial basis function (RBF), and multiple linear regression analysis (MLR) were employed to estimate thousand-grain weight and porosity. The best-selected model was then used to classify 12 different pepper seed varieties. This applied model's performance was evaluated using the determination coefficient (R2), the root mean square error (RMSE), the mean relative percentage absolute error (MRPE), and the mean square error (MSE). A comparison of the ANN model results indicated that the input parameters-width, length, thickness, and bulk density-provided the optimal prediction model concerning R2, RMSE, MRPE, and MSE. For the testing dataset, the ANN model achieved values ranging from 0.88 to 0.92 for R2, 0.276 to 0.016 for RMSE, 1.647 to 0.232 for MRPE, and 0.138-0.008 for MSE using 5-8-1 and 8-10-1 network structures, respectively. These selected models can be used as a neurocomputing-based approach to predict the mechanical and physical properties of pepper seeds, assisting in variety classification and genotype prediction.