Advanced Predictive Analytics in Agriculture: Case Study on Wheat Kernel Weight

This research, was aimed at modeling the thousand-grain weight of 13 different wheat varieties using five different input parameters. We used multiple linear regression (MLR), artificial neural networks (ANN), principal component analysis (PCA), and two different hybrid models consisting of PCA + MLR and PCA + ANN for this purpose. The MLR models were tested with various input configurations, demonstrating moderate explanatory power, with R² values ranging from 0.37 to 0.44. Increasing the number of independent variables increased prediction accuracy but also increased the risk of overlearning. ANN models showed significantly higher performance in prediction accuracy. The best performance was achieved in the ANN20 architecture with an R2 value of 0.866. In this architecture, a combination of the gradient descent training function, the hyperbolic tangent sigmoid transfer function, the linear transfer function, and 18 neurons were used. The PCA+MLR hybrid model was not effective in predicting thousand-grain weight. The fact that R² values obtained with different input configurations vary between 0.24 and 0.31 shows that the prediction accuracy of the model is low. In contrast, the PCA+ANN hybrid model significantly improved the prediction accuracy, and the best model achieved an R2 value of 0.981, an RMSE of 0.0829, and an MAE of 0.0359. The PCA+ANN model, which preserved the necessary variance by reducing the complexity of the input data, enabled the ANN to focus on the most critical components for accurate prediction. This study demonstrates that whereas ANN and PCA+ANN models give significantly increased accuracy in predicting wheat varieties' thousand-kernel weights, MLR models only offer moderate prediction capabilities.