Performance of different classifiers for marine habitat mapping using side scan sonar and object-based image analysis

• Main Authors: Rusmadi, Raihanah, Che Hasan, Rozaimi
• Format: Conference or Workshop Item
• Language: English
• Published: 2020
• Subjects: QA75 Electronic computers. Computer science; T58.5-58.64 Information technology
• Online Access: http://eprints.utm.my/id/eprint/93210/1/RozaimiCheHasan2020_PerformanceofDifferentClassifiersforMarine.pdf; http://eprints.utm.my/id/eprint/93210/; http://dx.doi.org/10.1088/1755-1315/540/1/012087

Acoustic sonar techniques have been one of the successful underwater mapping alternatives for identifying the seafloor features. The integration between the technique and classification analysis can produce detail map of the seafloor. Among these sonar technologies, side-scan sonar (SSS) is one of the tools for underwater mapping that can provide high spatial resolution seafloor mosaic which is presented in greyscale level. However, before it can be used for the coral reef marine habitat mapping, it is essential to properly assess its performance and quantify the amount of information that can be extracted. The objective of this study is to determine the accuracy of habitat maps derived using side scan sonar data, Object-based Image Analysis (OBIA) and five different classifier algorithms; Support Vector Machine (SVM), Random Forest (RF), k-Nearest Neighbour (k-NN), Decision Tree, and Bayes. This study utilized side-scan sonar model Klein system 3000 which operated at 100kHz combined with video data that was conducted in shallow water (depth > 10m). First, eight (8) texture layers were derived from side scan sonar mosaic using GLCM technique. Then, the GLCM layers of texture features were reduced using Principal Component Analysis (PCA) and analysed to seek for the most contributed texture layers. A total of 80 samples were derived which consist of four (4) classes; coral, sand, silt and mud. The result shows that the Support Vector Machine (SVM) method produced the highest accuracy which is 81.25% followed by k-Nearest Neighbours (k-NN), Random Forest (RF), Decision Tree and Bayes (68.75%, 66.25%, 57.5% and 45% respectively). The used of OBIA with SSS data offers a promising method to map marine habitats for a better understanding of spatial distribution and monitoring habitat changes in the future.