Benthic Habitat Mapping using Sentinel 2A: A preliminary Study in Image Classification Approach in An Absence of Training Data

Abstract

Numerous approaches for deriving benthic habitat mapping from visible spectrums of remotely-sensed imagery have been widely used, but image classification without training data for remote benthic habitat remains a few. In many cases, the collection of the needed ground-truth data is often prohibitively expensive or logistically infeasible. This will prevent us from providing training data for image classification purposes. In this paper, we evaluated the accuracy of the classification of benthic habitat from Sentinel 2A imagery in an absence of training data in the optically shallow water of Pari Island, Kepulauan Seribu, Indonesia. Benthic Habitat map was produced from geometrically, radiometrically, and water column corrected Sentinel 2A images. For water column correction, we performed Depth Invariant Index (DII) transformation. It was followed by the classification of Sentinel 2A imagery by applying unsupervised classification, such as IsoData and K-means algorithm. From the experiment, we produced four habitat classes. The analyses result for each unsupervised classification shows that the overall accuracy of IsoData and K-Means was 47.98% and 55.64%. However, the results of the Kappa coefficient show that the IsoData algorithm has slightly better accuracy of benthic habitat mapping (0.39) rather than K-Means (0.30).