Seabed classification of coral reef environments using 200 kHz single-beam echosounder and machine learning techniques

Abstract

This study focused on seabed classification of coral reef environments in northern Aceh using a single-beam echo sounder (SBES). The area is characterized by rubble and coarse sediment interspersed with fine sediment from estuaries. Additionally, the acoustic features extracted using Sonar5-pro were the first and second bottom peaks (BP1, BP2), attack (Att), and decay (Dec). Target classes for sediment classification were constructed based on six textural categories, including gravel, sand, gravelly sand, muddy sand, sandy gravel, and slightly gravelly sand. The occurrence of gravel and mud fractions within classifications was a proxy for the presence of coral reef and potential fine sediment influx. Meanwhile, machine learning models, such as Random Forest (RF), Support Vector Machine (SVM), and K-Nearest Neighbors (KNN), were applied to classify seabed types. Feature selection significantly affected the model performance and SVM achieved the highest cross-validation accuracy, excelling in mixed sediment zones due to nonlinear boundary capabilities. RF performed with coarser substrates, while KNN reported moderate accuracy with specific feature sets. The results showed that the combination of BP1, BP2, Att, and Dec enhanced classification accuracy of habitat influenced by coral and estuarine environments to capture subtle textural variations often missed by traditional methods. © 2025