Comparative evaluation of hybrid and individual models for predicting soybean yellow mosaic virus incidence.

Forecasting the severity of crop diseases is crucial for agricultural productivity and can be achieved through statistical and machine learning techniques. Predictive models that consider weather conditions during critical growth stages of crops have shown promising accuracy. However, selecting the most suitable forecasting model remains a challenge. This research investigates the impact of various weather factors on Soybean Yellow Mosaic Virus (SYMV) incidence. Specifically, six multivariate models Stepwise Multiple Linear Regression (SMLR), Artificial Neural Networks (ANN), Least Absolute Shrinkage and Selection Operator (LASSO), Ridge Regression (RR), Elastic Net (ELNET), and SMLR_ANN both direct and with Principal Component Analysis (PCA)-were developed using 20 years of data (2001 to 2020) to predict the severity of soybean disease in Pantnagar, Uttarakhand. The dataset was divided into two parts, with 80% used for calibration and the remaining 20% for validation. Model accuracy was evaluated using several statistical criteria, including R², RMSE, nRMSE, MAE, PE, and EF. The results indicated that the PCA-SMLR-ANN (nRMSE = 0.76%) model was the most effective predictor of soybean disease severity, closely followed by the PCA-ANN (nRMSE = 3.67%) model. Hybrid models such as PCA-SMLR-ANN and PCA-ANN outperformed individual models like SMLR (nRMSE = 47.72%) and ANN (nRMSE = 6.82%). The performance ranking of the models is as follows: PCA-SMLR-ANN ≈ PCA-ANN ≈ SMLR-ANN ≈ ANN > PCA-ELNET > PCA-Ridge > ELNET ≈ RR > PCA-LASSO > LASSO > PCA-SMLR ≈ SMLR. These findings highlight the superior efficiency of hybrid models in predicting soybean disease severity based on weather indices in the study region.