Optimization of Random Forest Algorithm for Paddy Growth Stage Classification Using Bayesian Optimization and Oversampling

Remote sensing and machine learning (ML) technologies that have developed very rapidly allow data retrieval and processing for monitoring paddy field conditions with minimal human intervention. One of the machine learning algorithms that is widely used for paddy growth stage classification is Random Forest. This study rebuilt the Random Forest model with datasets derived from a combination of Area Sample Framework (KSA) data and Sentinel-1A imagery attributes. The highest accuracy result by this model is still <60%. Therefore, this study aims to examine how to optimize the Random Forest algorithm in classifying paddy growth stages. The optimization method used is to perform hyperparameter tuning with the Bayesian Optimization method and handle imbalanced data with oversampling techniques. As a result, the f1-score of Random Forest Optimization Model with Bayesian Optimization and oversampling increase in average of 6% and has a trend similar to the Random Forest baseline model with and produces a stable f1-score under varying dataset conditions.