Crop yield prediction in cotton for regional level using random forest approach

Abstract

Early and precise estimation of crop yield plays a crucial role in quantitative and financial assessment at field level, to lay down strategic plans for import–export policies for agricultural commodities and to doubling the income of farmers. In this study, a machine learning based random forest (RF) algorithm was used to predicate cotton yield at three distinct times before the actual harvest in the state of Maharashtra in India using R package. Long-term agromet-spectral variables derived from multi-sensor satellites with actual crop yield from 2001 to 2017, were used to generate co-linearity of predictor variables and further, calibrate and validate the RF model. The performance of the RF model was found reliable and faster in predicting the crop yield with the most influencing variables with 69%, 60% and 39% of coefficient of determination (R2) in the final yield for September, December and February months, respectively using CART decision tree and recursive feature elimination method in R programming. Results showed as RF algorithm has the capability to integrate and process a large number of inputs as derived from different satellite modalities, unscaled and non-uniform ground based information, expert knowledge, etc. With high precision and avoid over fitting of the model.