Search

The literature review found that only a few studies in Indonesia have been applied to modeling peatlands fire occurrences. In our previous studies, various machine learning classification approaches are applied to predict forest fire occurrence in the peatlands area. Our previous empirical study found that the datasets of fire hotspot areas, especially for data from the South Kalimantan area, need to be more balanced. In this paper, we consider ADAptive SYNthetic (ADASYN) method for the sampling approach to balance the datasets. We found that by balancing the data, the performance of the classification method can improve about 2-4%. [ABSTRACT FROM AUTHOR]

It is known that the studies of peatlands fire occurrences in Indonesia are less studied before. In our previous study, the prediction of the peatlands fire occurrence was modeled using various machine learning classification approaches. It is found that using South Kalimantan Province data, in the empirical study we previously found that the datasets are unbalanced, i.e., the occurrence and the nonoccurrence of fire hotspots areas. In the study presented in this paper, to improve the classification performance, we consider Adaptive Neighbor Synthetic Majority Oversampling Technique (ANS-SMOTE) approach to balance the data. Using the considered empirical data, we found that this method did not always gives improvement in the classification results. [ABSTRACT FROM AUTHOR]

Various classification methods have been considered to predict the occurrence of the forest fire, including the recent ensemble methods, such as bootstrap aggregating (bagging) method and its extension. Here we consider a hybrid approach between fuzzy c-means clustering and bagging. As the weak learner in the bagging approach, we consider multinomial logistic (multilogit) regression and Support Vector Machines (SVM) classification approaches. To see their empirical performance, the proposed approaches are applied to the public data set, where the bagging approach has not been considered as the classification method in the other studies yet. From the empirical results, we obtain the hybrid fuzzy c-means clustering - the bagging multilogit regression approach has the best accuracy for classifying the size of forest fire data. [ABSTRACT FROM AUTHOR]