DEVELOPMENT OF ZERO-INFLATED COUNT TIME SERIES MODELS FOR FORECASTING YELLOW STEM BORER (Scirpophaga incertulas) POPULATIONS OF RICE
Loading...
Date
2023-01-28
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
PROFFESSOR JAYASHANKAR TELANGANA STATE AGRICULTURAL UNIVERSITY
Abstract
Aim of this study was to develop the reliable statistical model to forecast the
yellow stem borer pest data. Count time series modelling is a popular statistical
approach in which auto-correlated discrete count observations are considered as
dependent variables, and the observations are assumed to be derived from Poisson or
negative Binomial distributions. Crop pest modelling is one of the major areas of
count time series modelling wherein daily or weekly counts of insects (pests) are
considered as dependent variables of interest. As the pest population does not occur
regularly on a daily basis there used to be many zeros in weekly count data under such
conditions the classical count time series models may not yield better results
alternatively zero excess count models can be used to model the data with excess
zeros.
In this study, the rice Yellow Stem Borer (YSB) populations recorded using a
light trap with an incandescent bulb along with weather parameters of major centres
generated under All India Coordinated Rice Improvement Project (AICRIP) from
2013-2021 were considered.
The residuals of the fitted models namely, Integer-valued Generalized
Autoregressive Conditional Heteroscedastic (INGARCH), Zero Inflated Poisson
Autoregressive (ZIPAR), Zero Negative Binomial Autoregressive (ZINBAR) models
were found significant for most of the centres. Therefore, to correct the classical
model a two stage count time series methodology is proposed in this study. In the
first stage, the YSB populations are modelled using the count time series models and
diagnostically tested using the multivariate Box-Pierce test. If this test is not
significant, the procedure ends. In the second stage, when the residuals are
significant, the BDS test is used to test for the nonlinearity of the residuals. If the
residuals are found to be significant, the ANN model is fitted to the residuals to
obtain the predicted values.
Finally, the ANN fitted residuals are combined with the count time series
fitted values to obtain the final YSB population forecast. The proposed two stage
methodology performed better thanclassical count time series models in both training
and testing data sets, further, two stage zero inflated models outperformed all models
which as it provides lowest root mean square error values. The proposed
methodology for an efficient early warning system to predict the YSB population
could greatly contribute to the sustainable site-specific pest management strategies to
avoid significant riceyield losses.