H.N. Salsabila, S.D. Harahap, A.F. Sahitya
Accurate benthic habitat mapping is essential for coastal management and ecosystem monitoring. However, remote sensing-based classification faces challenges due to the small size of benthic objects and their submerged nature, which increases the risk of misclassification. Machine learning algorithms such as Random Forest (RF) and Support Vector Machines (SVM) have been widely used to address these limitations, yet each has its drawbacks-RF may overfit. At the same time, SVM can misclassify and is sensitive to complex samples. This study proposes a hybrid classification method to overcome these limitations by fusing RF and SVM outputs within Google Earth Engine. The study area is located along the coast of Bontang City, East Kalimantan, Indonesia. Sentinel-2 imagery was classified using RF (ntree=50, 100), SVM (gamma=10, cost=10), and a hybrid RF-SVM approach. The fusion rule applied in the hybrid model is as follows: if RF and SVM predictions agree, the shared class label is accepted. If the predictions differ, the final class is determined using a 3×3 majority neighborhood filter to resolve the disagreement and assign the benthic habitat class. Three benthic classes were mapped: coral/macroalgae, seagrass, and bare substrate. RF yielded an overall accuracy range from 78.19 to 78.50%, while SVM reached 72.27 to 81.93%. Based on four scenarios, three of them showed that the hybrid method outperformed both RF and SVM individually. The hybrid method achieved the highest overall accuracy of 82.24% with a Kappa value of 0.70, with producers' and users' accuracy outperforming both individual classifiers in most classes. The use of varied parameters for RRF and SVM demonstrated the effectiveness of the hybrid approach. Spatially, this method reduced salt-and-pepper noise and improved consistency across benthic habitats. These findings indicate that the hybrid fusion method enhances the accuracy of benthic habitat mapping. Moreover, the hybrid RF-SVM approach effectively addressed key limitations encountered in single-algorithm classifications, such as the inability to accurately map coral/macroalgae in SVM and RF's "salt and pepper" effect. © ACRS 2025.All rights reserved.
Department of Geography, Faculty of Mathematics and Natural Sciences, Universitas Negeri Makassar, South Sulawesi, Makassar, 90224, Indonesia; Faculty of Geography, Universitas Gadjah Mada, Sekip Utara, Kab. Sleman, Daerah Istimewa, Yogyakarta, 55281, Indonesia; Blue Carbon Research Group, Faculty of Geography, Universitas Gadjah Mada, Sekip Utara, Kab. Sleman, Daerah Istimewa, Yogyakarta, 55281, Indonesia
