Pemetaan Habitat Bentik Menggunakan Machine Learning dengan Citra Sentinel-2A di Perairan Dangkal Barat Pulau Satanger, Sulawesi Selatan

Abstract

Perairan dangkal bagian barat Pulau Satanger memiliki keragaman habitat bentik yang berperan penting bagi fungsi ekologis kawasan pesisir, namun informasi spasial mengenai persebarannya masih terbatas. Penelitian ini bertujuan memetakan habitat bentik menggunakan citra Sentinel-2A berbasis Object-Based Image Analysis (OBIA) dengan algoritma machine learning, membandingkan kinerja algoritma Support Vector Machine (SVM), Random Forest (RF), dan K-Nearest Neighbor (KNN), serta mengevaluasi pengaruh koreksi kolom air menggunakan metode Depth Invariant Index (DII). Pengolahan citra meliputi pre-processing, koreksi kolom air, segmentasi citra dengan parameter scale 70 pada level 1 dan scale 10 pada level 2, klasifikasi multiskala, serta uji akurasi. Hasil klasifikasi menghasilkan lima kelas habitat bentik, yaitu rubble, pasir, pasir + lamun, lamun, dan karang hidup + karang mati. Algoritma SVM memberikan performa terbaik dengan Overall Accuracy (OA) sebesar 68,33% dan koefisien Kappa 0,60 pada kondisi tanpa penerapan DII, kemudian meningkat menjadi 70% pada nilai OA dan 0,62 pada nilai Kappa setelah penerapan DII. Peningkatan nilai OA dan Kappa juga ditunjukkan oleh algoritma RF, sedangkan algoritma KNN menunjukkan penurunan.

The shallow waters of western Satanger Island have diverse benthic habitat that play an important role in the ecological functions of coastal ecosystems. However, spatial information regarding their distribution remains limited. This study aimed to map benthic habitat using Sentinel-2A imagery based on an Object-Based Image Analysis (OBIA) approach with machine learning algorithms, compare the performance of the Support Vector Machine (SVM), Random Forest (RF), and K-Nearest Neighbor (KNN) algorithms, and evaluate the effect of water column correction using the Depth Invariant Index (DII) method. Image processing included pre-processing, water column correction, image segmentation using a scale parameter of 70 at Level 1 and 10 at Level 2, multiscale classification, and accuracy assessment. The classification results identified five benthic habitat classes, namely rubble, sand, mixed sand and seagrass, seagrass, and mixed live and dead coral. The SVM algorithm achieved the best performance, with an Overall Accuracy (OA) of 68.33% and a Kappa coefficient of 0.60 without DII application, which increased to an OA of 70% and a Kappa coefficient of 0.62 after DII application. Improvements in OA and Kappa values were also observed for the RF algorithm, whereas the KNN algorithm showed a decline.