Analisis Klasifikasi Tutupan Lahan di Kabupaten Jember Berbasis Citra Satelit Landsat-9 Menggunakan Algoritma Xgboost dan Maximum Likelihood

Abstract

Change in land availability driven by population growth and economic development demand thorough spatial planning. This study aims to implement and analyze the performance comparison between the traditional Maximum Likelihood Classifier (MLC) method and an ensemble learning-based method, Extreme Gradient Boosting (XGBoost), for mapping land cover in Jember District using Landsat-9 satellite imagery from 2023. The features utilized for classification consist of 6 original spectral bands (Blue, Red, Green, NIR, SWIR-1, SWIR-2) and 3 derived spectral indices (NDVI, MNDWI, NDBI). The study area was classified into 6 main classes: water body, open land, built-up land, cultivated vegetation, natural vegetation, and shadow. Model accuracy was evaluated using a confusion matrix with an 80% training and 20% testing data split from a total of 3.953 clean sample pixels. The results demonstrate that the XGBoost algorithm is superior in terms of effectiveness, yielding an overall accuracy (OA) of 93,81% and a kappa accuracy of 92,8%. Conversely, the MLC algorithm only achieved an OA of 87,86% and a Kappa Accuracy of 85,9% due to its vulnerability to misclassification in areas with high spectral similarity, such as extreme topographic shadows falsely detected as open land. However, from the computational efficiency perspective, MLC vastly outperformed XGBoost, with a processing time of just 12 seconds for mapping the entire region, compared to XGBoost which required 4 minutes and 15 seconds.