Application of statistical and machine learning models in combination with stepwise regression for predicting rapeseed-mustard yield in Northern districts of West Bengal

Abstract

Rapeseed-mustard crop is an important oilseed crop in India. District-wise yield prediction is essential for various location specific decision making. The performance of two machine learning models namely Support Vector Regression (SVR) and Artificial Neural Network (ANN) were compared with basic linear regression model for district-wise yield prediction of rapeseed-mustard crop. The study area for the present investigation were Cooch Behar, Malda, Jalpaiguri and Uttar Dinajpur districts of West Bengal. Yearly unweighted and weighted weather indices were calculated from weekly weather parameters. The indices that significantly affecting yield were selected using stepwise regression for fitting the models. The ANN model was fitted using backpropagation algorithm. The optimum number of neurons in hidden layer for ANN were ranging between two to four. The Tangent hyperbolic function was found to be suitable hidden layer activation function. The nonlinear Radial Basis Function kernel was the best kernel for Support Vector Regression. While evaluating the performance of fitted models in both calibration and validation stages, the ANN model was the best fitted model for Cooch Behar and Malda and SVR was the best fitted model for Jalpaiguri and Uttar Dinajpur districts. It was concluded that the machine learning models outperformed multiple linear regression model for district-wise yield prediction of rapeseed-mustard crop.