COMPARATIVE ACCURACIES USING MACHINE LEARNING MODELS FOR MAPPING OF SUGARCANE PLANTATION BASED ON SENTINEL-2A IMAGERY IN KEDIRI AREA, EAST JAVA

Abstract

Data collection in smallholder sugarcane plantations is still very sensitive to the subjectivity of informants and data collectors. In the meantime, the problem with data collection on sugarcane plantation companies is a low response rate. This situation can reduce the precision of the estimates that are produced. Consequently, the goal of this research is to recognize sugarcane fields using the machine learning models on Sentinel-2A satellite imagery in Kediri Area that covering Kediri Regency and Kediri Municipality, East Java. Along with developing machine learning algorithms, this research will evaluate how well LightGBM performs when compared to other algorithms, including CART, SVM, Random Forest, and XGBoost. Each model employed hyperparameter tuning with random search and stratified 10-fold cross validation to avoid overfitting. The process of labelling satellite imagery using images from Google Street View, then predictor variables used are NDVI, NDWI, NDBI, EVI, and elevation. The most accurate classification model obtained was LightGBM, with a 98% accuracy and a cohen’s kappa of 97.7%. The estimated area of sugarcane plantations in the Kediri Regency and Kediri Municipality in September 2022 is 18,897.6 ha and 571.87 ha.