Improving machine learning prediction of peatlands fire occurrence for unbalanced data using SMOTE approach

From our previous study, we have known that only a small number of literatures have studied peatlands fire modeling in Indonesia. It is including our recent study on the prediction of the forest fire occurrence in the peatlands area using some machine learning classification techniques. In the previous empirical study using data from South Kalimantan Province in Indonesia, we found that the datasets are unbalanced between the two classes of data, i.e., the occurrence of fire hotspots and the nonoccurrence of fire hotspots areas. In this paper, the performance of the classification method is improved, by balancing the data using what so called Synthetic Minority Over-sampling Technique (SMOTE). In the empirical results, we show the performance of the classification results on the balanced data are mixed. It is found that only using the ensemble AdaBoost with SMOTE balanced data the performance of the methods has always been improved over unbalanced data, either for in-sample or for out-sample cases. The open-source software R is used for implementation of the methods.