Mapping Oil Palm Plantations Using WorldView-2 Satellite Imagery and Machine Learning Algorithms

Abstract

Currently, remote sensing has been used extensively in the agriculture industry for oil palm monitoring due to their large plantation area. Oil palm monitoring can be done by performing land cover classification using various classification methods and machine learning algorithms. This study was conducted to perform oil palm mapping using WorldView-2 satellite imagery and classify land cover features using machine learning algorithms such as Random Forest (RF) and Linear Support Vector Classifier (LSVC). A total of 58609 sampling points were classified into six classes which are water, built-up, bare soil, forest, mature oil palm (≥9 years) and young oil palm (3-8 years). The training and testing samples were split using 3-fold cross-validation. 67% of the total sampling points were used for training samples whereas the other 33% were used for testing samples. The methods used to validate the data in this study is by using spectral reflectance and Google Earth Pro. Accuracy assessment was conducted after obtaining the classification output such as Overall Accuracy (OA), Kappa Accuracy (KA), Precision, Recall and F1-score. As a result, the oil palm mapping using RF has a higher accuracy than LSVC which is 72.49% for OA and 62.98% for KA. The p-value obtained from the McNemar’s test conducted in this study is 0.683 (>0.05) which concludes that the predictive performance of the two models are equal.