Improve performance of extreme learning machine in classification of patchouli varieties with imbalanced class

Abstract

Patchouli has various varieties with almost the same physical characteristics. This often makes it difficult to recognize varieties with a high PA (Patchouli Alcohol) content. In this study an improvisation was introduced in the identification of patchouli varieties using leaf images using Extreme Learning Machine (ELM). However, problems occur in ELM if the data used is not balanced where the training process can not able to recognize data in the minority class well. Therefore, this study conducted a process to balance the composition of the data using the Synthetic Minority Over-sampling Technique (SMOTE) method. The test results of 93 data on the imbalanced composition with a comparison of 70% of training data and 30% of test data obtained an average accuracy of 93.57%. After implementing SMOTE in the Tetraploid, Patchoulina and Sidikalang classes where the amount of data in each class becomes 58 data, an average accuracy of 96.00% is achieved. This shows the existence of an increase in the process of identification with ELM when new data generation with SMOTE is carried out to balance the composition of the data.