Evaluation of classification techniques for benthic habitat mapping

Abstract

The coral ecosystem is sensitive to environmental changes thus accurate, up to date information on their status is critical for effective management of these important marine resources. However, environments containing these habitats are challenging to map due to their remoteness, extent and costs of monitoring. In this research, the capabilities of satellite remote sensing techniques combined with in situ data were assessed to generate coral habitat map of Lang Tengah Island, Terengganu, Malaysia. Several classification techniques were utilized in identifying coral distribution to assess their ability to map different type of benthic habitat associated with coral reefs. Five classifiers were used to classify the study area mainly, Parallelepiped, Minimum distance, Maximum likelihood, Fisher and K-Nearest Neighbour. Using the same training data sets to evaluate their effectiveness, results from the classification shows that each method produced different accuracy based on bottom type. Utilizing the strength of each classifier this study was able to increase per class accuracy of the habitat map through several image processing techniques mainly majority voting, simple averaging and mode combination. Results show that by utilizing these ensemble techniques for classifying benthic habitat the accuracy produced was higher than conventional supervised techniques.