Ensemble Machine Learning Methods to Estimate the Sugarcane Yield Based on Remote Sensing Information

Received: 8 October 2020 Accepted: 12 December 2020 The purpose of this study is to investigate the computing capabilities of machine learning algorithms and remotely sensed signals to extract the agricultural information. Many techniques and models have been developed to extract information from the remotely sensed observations, but it remains an exigent problem due to the accuracy, reliability and timeliness parameters. Sugarcane yield estimation based on the temporal profile of multispectral Landsat-8 data has been explored in the proposed work. An initial attempt has been made in this study to select important parameters to be used as input to the machine learning method. Mean Decrease Accuracy and Mean Decrease Gini measures of random forest algorithm have been used to select the important parameters for predictive modelling. The results of the study revealed that Green Normalized Vegetation Index, Normalized Difference Vegetation Index and Land Surface Water Index performed best among other indices. Bands B2, B3, B6 and B7 of Landsat-8 recorded as top scorers. The proposed work focused on ensemble machine learning methods to optimize the correlation of historical crop yield values with spectral information. The Random Forest method exhibits a significant performance (RMSE= 1.51 t/ha and R = 0.94) as compared with other methods such as Classification and Regression Tree, Support Vector Regression and K-Nearest Neighbor. The proposed model based on random forest algorithm is best among all the scenarios and growth stages, whereas model based on classification and regression tree performs worst in all the cases. The proposed study indicates that the numerical value of a single spectral parameter and single-date data is not sufficient for the reliable yield estimation because it is difficult to discriminate some of the crops due to similar phenology in a particular growth period.