Rice Yield Prediction Using On-Farm Data Sets and Machine Learning

Abstract

In this paper a rice crop prediction performance analysis of five machine learning and two multilinear regression algorithms is presented. A five hectares rice plot was selected. For the database, in the plot, 72 sampling points spatially distributed were defined. For each sampling point, physicochemical, biomass and leaf chlorophyll content measurement were taken at vegetative stage. Additionally, the plot was flown with a quadcopter to take multispectral images in order to calculate vegetation indices maps. As output variable, the crop yield was defined. The machine learning (ML) algorithms used in this analysis were: Random Forest, eXtreme Gradient Boosting, Support Vector Regression Machines, Multilayer Perceptron Regression Neural Networks, and K-Nearest Neighbors; the multilinear algorithms were Partial Least Squares and Multiple Linear regression (MLR). The results show the best performance for K-Nearest Neighbors with an average absolute error for the testing point of 10.74%. The worst case was the MLR with a root mean square error (RMSE) of 2712.26 kg-ha\(^{-1}\) in the testing dataset, while KNN regression was the best with 1029.69 kg-ha\(^{-1}\).