Exploring machine learning algorithms for mapping crop types in a heterogeneous agriculture landscape using Sentinel-2 data. A case study of Free State Province, South Africa

Abstract

Accurate and detailed studies in crop mapping are crucial in precision agriculture, yield estimations, and crop monitoring. This study focused on exploring the utility of Sentinel-2 data in mapping of crop types and testing the two machine learning algorithms which are Random Forest and Support Vector Machine performance in classifying crop types in a heterogeneous agriculture landscape in Free state province, South Africa. Nine crop types were successfully classified. The utility and contribution of different bands for classification were evaluated using RF mean decrease GINI for variable importance. Validation of results was done using a confusion matrix which produced overall accuracy, errors and prediction measures. The best performance was attained by SVM with an overall accuracy of 95% and a kappa value of 94%. RF also performed fairly well with 85% of overall accuracy and kappa value of 83%. It was concluded that Sentinel-2 data performs better using the SVM classifier compared to RF classifier.