An innovative two-stage, zero-inflated, hybrid count time series model for predicting rice yellow stem borer using weather parameters in hotspot locations of India

The Yellow Stem Borer (YSB) ( Scirpophaga incertulas Walker) is a major pest in rice agroecosystems, and timely prediction of its occurrence is crucial for effective management. This study considered data from 2013 to 2023 from five YSB hotspot regions in India, namely Warangal, Rajendranagar, Pattambi, Raipur, and Chinsurah to develop a reliable forewarning model for predicting YSB populations using weather parameters. Daily YSB catches were recorded using light traps with 200W incandescent bulbs, and various weather variables were also considered. Stepwise regression identified key weather parameters influencing YSB population density, including minimum and maximum temperatures, evening and morning relative humidity, sunshine hours, and rainfall. The study utilized weekly cumulative YSB populations and average climatological data as inputs to several count time series models, including the Integer-valued Generalized Autoregressive Conditional Heteroscedastic (INGARCH) model, Zero-Inflated Poisson Autoregressive (ZIPAR) model, Zero-Inflated Negative Binomial Autoregressive (ZINBAR) model, and the Artificial Neural Network (ANN), machine learning model. Additionally, innovative two-stage hybrid models viz., INGARCH-ANN, ZIPAR-ANN, and ZINBAR-ANN were developed and evaluated. Classical count time series models, such as INGARCH, underperformed when a high proportion of zeros were observed due to the absence of YSB in certain Standard Meteorological Weeks (SMWs). Zero-inflated models were found to be better suited for such cases. Classical models showed significant residual patterns, indicating the need for model correction. To address this, hybrid models were constructed to normalize the residuals and enhance forecasting accuracy. Among all the models, the two-stage ZINBAR-ANN model outperformed the others, showing the lowest Mean Squared Error (MSE) and Root Mean Squared Error (RMSE) across most locations in both training and testing datasets for both rainy and post-rainy seasons. This innovative two-stage zero-inflated hybrid modeling approach offers an effective early warning system for YSB population prediction, supporting sustainable and localized pest management strategies, and helping to mitigate rice yield losses.