A Comparative Study of Machine Learning Classification Algorithms for Benthic Habitat Mapping in West Bali National Park

The West Bali National Park (Taman Nasional Bali Barat/TNBB) comprises a variety of ecosystems, including coastal regions that support different habitats such as mangroves, seagrass, and coral reefs. This distinctive conservation area, situated in the Jembrana and Buleleng Districts, encompasses roughly 3,415 hectares and fosters the development and preservation of diverse species, including benthic ecosystems. Benthic habitats are vital elements of coastal ecosystems, including seagrass, coral reefs, macroalgae, and diverse substrate types. This study seeks to evaluate the efficacy of two machine learning classification algorithms, Random Forest (RF) and Support Vector Machine (SVM), in delineating benthic habitat composition in the Teluk Terima region of TNBB, utilizing 3-meter resolution PlanetScope imagery encompassing visible and near-infrared (NIR) spectral bands. The classification emphasizes four principal benthic habitat categories: predominant seagrass, predominant coral, predominant macroalgae, and predominant substrate. The RF algorithm achieved an overall accuracy (OA) of 59.1, delineating 27.3 ha of predominant seagrass, 35.9 ha of coral, 0.5 ha of macroalgae, and 40.2 ha of substrate. In contrast, the SVM method achieved a diminished OA of 47.3, delineating 35.4 ha of seagrass, 43.9 ha of coral, 15.5 ha of macroalgae, and 9.2 ha of substrate. Comparative accuracy assessments reveal that Random Forest (RF) exhibits more stability in classifying seagrass and substrate categories, whereas Support Vector Machine (SVM) has superior performance in identifying coral and macroalgae, notwithstanding discrepancies in user and producer accuracy. The results indicate that the selection of classification method substantially influences benthic habitat mapping results, with Random Forest providing more reliable conclusions in shallow water settings characterized by intricate substrate compositions. © ACRS 2025.All rights reserved.