Estimation of bread wheat yield by multiple linear regression (MLR) and artificial neural network (ANN) methods and their fair comparison

Abstract

Abstract This study investigates the accuracy of Multiple Linear Regression (MLR) and Artificial Neural Networks (ANN), specifically a hybrid Genetic Algorithm-ANN (GA-ANN), for predicting wheat yield in plant breeding. This research, conducted using 782 wheat genotypes in Rafsanjan, Iran, compares MLR and ANN methodologies. MLR, using seven traits selected via stepwise regression, achieved an R² of 0.90, the root of mean square of error (RMSE) of 14, Average absolute percentage error (MAPE) of 13.3, and Average deviation of prediction from the actual value (MAE) of 10. Key traits identified were biological weight (weight of total plant per line, WPP) and harvest index (HI). Conversely, the GA-ANN model, employing six selected traits, demonstrated superior performance with R² values of 0.94, 0.96, and 0.94 for training, testing, and combined datasets respectively. Validation metrics for the ANN model were MSE of 144.3, RMSE of 12, and MAE of 5.7. GA-ANN selected height, peduncle length, days to flowering, spike length, biological weight, and harvest index as significant predictors. The results underscore that ANN models, particularly when combined with genetic algorithms for feature selection and optimization, can improve prediction accuracy by modeling complex, non-linear relationships in agricultural data, therefore providing more precise yield prediction tools for plant breeders. This study emphasizes the need for advanced predictive techniques in achieving more accurate assessments of crop yield for sustainable agriculture.

Plant phenotyping relevance

小麦遺伝子型レベルの収量という植物形質を対象に、MLRとGA-ANNの予測精度を比較・検証しており、計算による形質推定手法が研究の中心です。

Code and data availability

The supplied blocks describe a wheat yield prediction study (MLR vs GA-ANN on 782 genotypes) but contain no data availability statement, no public phenotype dataset, no author code/model repository, and no URLs. Analysis was done in MINITAB-18, SPSS.16, Path2, Excel, and MATLAB with no deposit or sharing language.

No evidence-backed public reproduction asset is currently recorded.