Development of Hybrid Time Series Model using Machine Learning Techniques for Forecasting Crop Yield with Covariates

Abstract

Meeting food demand of the growing population is a major challenge. In order to ensure food security of this mega country in upcoming couple of decades, a precise Government policy is required. Although, some of the institutes of ICAR have formulated their vision 2030 or 2050, but, the projection of the food demand and supply mentioned in the vision are not based on sound statistical foundation. Enhancing the accuracy and reliability of the forecast of the food production for upcoming couple of decades is a major challenge. Production is multiplication of cropped area with yield. In this study, we are focused on yield as we have considered time series weather variables for forecasting yield. Although, few work on short term time series forecast of the crop yield are available in literature but hardly any work is available on long term forecast of crop yield using hybrid time series model using weather variables as covariates. In an attempt to get long term forecast of the crop production, we have used Box Jenkins linear time series approach in first instance and obtained the residuals and improved the forecast values of yield by hybrid linear time series approach using machine learning techniques like artificial neural network and support vector machine. We have also proposed a technique to compute the long term forecast using the improved short term forecast through the hybrid approaches. The improved short term forecast values of yield have been considered as the baseline data and through the proposed approach we get the long term forecast of yield up to desired forecast horizon. Apart from this, autoregressive integrated moving average using exogenous variable (ARIMAX) model has also been used for yield forecast using weather variables. The forecasted yield through ARIMAX model has been further improved through hybrid approaches and has been considered as baseline data for further long term yield forecast up to the desired year.