Klasifikasi Sedimen Dasar Laut Menggunakan Multibeam Echosounder dan Metode Deep Learning

Abstract

Pemetaan dan klasifikasi sedimen dasar laut merupakan komponen penting dalam pengelolaan wilayah pesisir, perencanaan infrastruktur laut, konservasi ekosistem bentik, serta kajian dinamika oseanografi. Salah satu teknologi yang banyak digunakan untuk tujuan tersebut adalah Multibeam Echosounder (MBES), yang mampu menghasilkan data batimetri dan hambur balik akustik (backscatter) beresolusi tinggi dengan cakupan spasial yang luas. Nilai backscatter MBES dipengaruhi oleh karakteristik fisik sedimen, sehingga berpotensi dimanfaatkan untuk klasifikasi sedimen dasar laut secara otomatis. Penelitian ini bertujuan mengklasifikasikan sedimen dasar laut di Perairan Pulau Pari, Kepulauan Seribu, serta membandingkan performa tiga algoritma deep learning, yaitu Convolutional Neural Network (CNN), Feedforward Neural Network (FNN), dan Tabular Prior-data Fitted Network (TabPFN). Akuisisi data dilakukan menggunakan MBES Teledyne Reson SeaBat T50-R, dilengkapi dengan pengukuran Conductivity Temperature Depth (CTD) dan pasang surut untuk mendukung koreksi batimetri dan backscatter. Sebanyak 42 sampel sedimen dikumpulkan menggunakan Van Veen grab sampler dan dianalisis di laboratorium untuk menentukan fraksi ukuran butir sebagai data ground truth. Hasil pengolahan menunjukkan bahwa sedimen dasar laut di Perairan Pulau Pari didominasi oleh kelas pasir dengan sebaran terbatas sedimen karang/batu/rubble. Pola pasang surut yang bersifat periodik menunjukkan peran pasut yang kuat dalam mengontrol dinamika hidrodinamika dan redistribusi sedimen di wilayah penelitian. Data backscatter MBES dikombinasikan dengan parameter turunan batimetri berupa slope, curvature, rugosity, dan Bathymetric Position Index (BPI) sebagai fitur klasifikasi. Evaluasi kinerja model menunjukkan bahwa CNN dan FNN menghasilkan akurasi rendah (0,444) serta nilai kappa lemah akibat keterbatasan jumlah data training dan ketidakseimbangan kelas. Sebaliknya, TabPFN memberikan performa terbaik dengan nilai akurasi 0,837, F1-score 0,833, dan koefisien kappa 0,673, yang menunjukkan kemampuan generalisasi yang lebih baik pada dataset berukuran kecil.

Seafloor sediment mapping and classification play an essential role in coastal management, marine infrastructure planning, benthic habitat assessment, and oceanographic studies. One of the most widely used technologies for seafloor characterization is the Multibeam Echosounder (MBES), which provides high-resolution bathymetry and acoustic backscatter data with broad spatial coverage. Acoustic backscatter is strongly influenced by the physical properties of seabed sediments, making MBES a reliable tool for automated sediment classification. This study aims to classify seafloor sediments in the waters of Pari Island, Seribu Islands, Indonesia, and to compare the performance of three deep learning algorithms: Convolutional Neural Network (CNN), Feedforward Neural Network (FNN), and Tabular Prior-data Fitted Network (TabPFN). Data acquisition was conducted using a Teledyne Reson SeaBat T50-R MBES system, supported by Conductivity Temperature Depth (CTD) measurements and tidal observations to ensure accurate bathymetric and backscatter corrections. A total of 42 sediment samples were collected using a Van Veen grab sampler and analyzed in the laboratory to determine grain-size composition, which served as ground truth data for model training and validation. The sediment analysis results indicate that the seafloor of Pari Island is predominantly composed of sand, with limited occurrences of coral rubble. Tidal analysis reveals a periodic tidal pattern, indicating that tidal dynamics play a significant role in controlling hydrodynamic processes and sediment redistribution within the study area. Acoustic backscatter data were integrated with bathymetric derivatives, including slope, curvature, rugosity, and Bathymetric Position Index (BPI), to construct the feature dataset for classification. Model performance evaluation shows that CNN and FNN yield relatively low accuracy (0.444) and weak kappa values, reflecting limited generalization capability due to the small and imbalanced training dataset. In contrast, TabPFN demonstrates superior performance with an accuracy of 0.837, an F1-score of 0.833, and a kappa coefficient of 0.673, highlighting its robustness and effectiveness in low-data conditions. These results confirm that TabPFN is the most suitable deep learning model for MBES-based seafloor sediment classification in data-limited, shallow tropical environments.