A Hybrid NBINGARCH–Machine Learning Framework for Forecasting Rice Leaf Folder (Cnaphalocrocis medinalis) Populations in Andhra Pradesh, India

Abstract

Rice (Oryza sativa) is a staple food crop that supports global food security, but its production is threatened by the rice leaf folder (Cnaphalocrocis medinalis), a major pest associated with yield losses. Accurate forecasting of leaf folder (LF) populations is important for timely pest management. Traditional statistical approaches, including multiple regression and time-series analysis, may not adequately represent nonlinear dependencies and environmental interactions. This study integrated statistical and machine-learning (ML) methods to improve forecasting accuracy for weekly LF count data from major rice-growing locations in Andhra Pradesh, namely Nellore, Ragolu, Maruteru, Bapatla and Nandyal. A hybrid forecasting framework was evaluated by combining the Negative Binomial Integer-Valued Generalized Autoregressive Conditional Heteroskedastic (NBINGARCH) model with Extreme Learning Machine (ELM), Artificial Neural Networks (ANN) and Support Vector Regression (SVR), yielding NBINGARCH-ELM, NBINGARCH-ANN and NBINGARCH-SVR models. Model performance was assessed using mean squared error (MSE), root mean squared error (RMSE) and the Box-Pierce residual autocorrelation test for training and testing datasets. The results indicated that NBINGARCH-ELM performed best in several datasets, particularly in Ragolu (Rabi), where it produced the lowest error values and no significant residual autocorrelation. However, higher errors in datasets such as Maruteru (Kharif) and Nandyal (Kharif) indicate the need for further refinement. Overall, the findings support the cautious use of hybrid statistical-ML models for location-specific pest early warning.