Forecasting of Sugarcane Productivity Estimation in India - A Comparative Study with Advanced Non-Parametric Regression Models

Purpose: In recent times, sugarcane production and area under cultivation have been fluctuating from year to year depending on climate and price policy, adversely affecting sugarcane growers' decisions to invest in cultivation and their livelihood. The declining trend of productivity may affect the future competitiveness, and therefore it needs to be investigated. Design/Methodology/Approach: In this study, prediction of sugarcane yielding through regression analysis is performed with the help of Multivariate Adaptive Regression Splines (MARS), Support Vector Regression (SVR), Partial Least Square Regression (PLSR), Elastic-Net Regression, and Multiple Linear Regression (MLR) on the basis of the historical data of sugarcane cultivation from 1971-72 to 2018-19. The prediction is done by training all the regression models with 80% of the data, by taking the overall Indian sugarcane productivity as a dependent variable and other major sugarcane producing states as independent variables. Findings: As a main result, the non-parametric regression model MARS is found to be much better than other well-fitted models. All of these models' performances are cross-validated using the Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and the Wilcoxon Signed-Rank test. Also, the MARS model is found to be a more flexible and accurate model in predicting the behavior of sugarcane yielding in India. Practical Implications: The practitioners and farmers facilitate the model comparisons to achieve more profits through accurate estimation The research outcome implicates the agricultural industry to improve the sugarcane cultivation and productivity under uncertain environments. Originality/Value: The study suggests best management practices can be developed to increase the large potential of sugarcane production in India towards greater sustainability and food security modelling.