Machine Learning Approaches using Satellite Data for Oil Palm Area Detection in Pekanbaru City, Riau

Abstract

Palm oil is a commodity that plays an important role in economic activity. The oil palm tree is capable of producing palm oil and is the most widely consumed vegetable oil in the world. Indonesia is the world's largest producer and exporter of palm oil. The huge potential of the palm oil industry in Indonesia demands the availability of accurate and up-to-date data sources. The latest remote sensing methods have now been widely used in detecting oil palm. We focus on modeling for oil palm detection as well as identifying features that affect oil palm to differentiate it from other land covers. This study compares the performance of the machine learning model with the Random Forest (RF), Xtreme Gradient Boosting (XGBoost), and Classification and Regression Tree (CART) methods. Grid Search is used to perform hyperparameter tuning. The results showed that the XGBoost model gave the best results with an F1 score of 0.90 and an accuracy of 90.97%. The most influential features on the model are B3 (blue). In addition, B3 is also mostly used by the palm oil class. The estimated area of oil palm based on the best model is 14,390.65 Ha, which is 13.18 percent higher than the official data.