Development of count time - series models for predicting pest dynamics using weather variables

Abstract

Agriculture being highly cost intensive and full of uncertainties have great impact on the livelihood of farmers, if timely measures are not taken to minimize the risk from incidence of pest attacks, they may fall in the trap of vicious cycle. Incidence of pest and diseases in crops have made agriculture very risky venture and about 15-25 per cent of crops yields is lost each year due to this. To mitigate these problems, reliable and timely forecast provides an important and extremely useful input in formulation of policies. An attempt has been made using count time series of pests using advanced models viz. generalized linear (GLM), integer autoregressive (INAR) and integer artificial neural network (IANN) models. This study has been conducted to develop count time series models for modelling and forecasting pest dynamic prediction. The developed models viz GLM, INAR and IANN were applied in aphids and Jassids of cotton pests at different centres of India. Pest count time series data (2008-2013) for Bt. cotton along with standard meteorological weekly data on weather parameters viz., maximum temperature (MAXT), minimum temperature (MINT), rainfall (RF), maximum relative humidity (MAX_RH) and minimum relative humidity (MIN_RH) were used to build the model. Developed models are compared using MAE, MSE and RMSE and based on these reported values it was observed that integer based artificial neural network (IANNX) models performed better compared to other models for Aphid of Akola centre. For Jassids of Akola, Banswara, Faridkot, Guntur and Perambalur centres integer based neural network model with exogenous variables performed better compared to other models. For Aphid of Vadodra centre INAR model outperformed over all the models. Based on the results obtained in this study it was concluded that whenever data set is nonlinear and complex in nature integer based neural network model performed better. On the other hand if data set is linear in nature, INAR model performed better compared to other models. Better performance of integer based neural network can be due to its ability to model the complex relationship existing in the data set.