Classifying benthic habitats in the shallow water around St. Eustatius using remote sensing

Abstract

This research aims to create a benthic habitat map of the shallow water (<30m) around St. Eustatius using highresolution multispectral WorldView-3 imagery and in situ observations. Various classification algorithms are explored and the effect of sunglint correction and water column correction on classification were examined. Additionally, the research investigates the detectability of changes in benthic habitat composition between 2009 and 2018.

To improve the original imagery, preprocessing steps are performed including sunglint removal and water column correction. Due to limited amount of sunglint, sunglint removal performed poorly, resulting in reduced variation in the deeper part of the imagery. Water column correction however, showed more promising results. When creating a Depth-Invariant Index of the blue (450-510 nm) and green (510-580 nm) spectral band, structures in deeper water showed more clearly than in the original multispectral image. However, despite visible improvements, overall classification accuracy remains lower compared to classification of the original imagery.

After preprocessing one unsupervised and four supervised classification algorithms are compared. Unsupervised classification algorithms can be used to segment the imagery without prior training data. The results of the ISODATA (Iterative Self-Organizing Data Analysis Technique) classification show that its accuracy is comparatively lower than supervised algorithms. However, it could be a valuable tool for initial exploration and identification of habitat patterns. Supervised classification algorithms, such as Maximum Likelihood Classifier (MLC), Random Tree (RT), K-Nearest Neighbor (K-NN), and Support Vector Machine (SVM), were better able to classify benthic habitats with accuracies ranging from 56.9% - 67.3%. SVM performed best visually in classification of the water column corrected image and represents each of the habitat classes more evenly compared to the other classifiers.

Lastly, an attempt at change detection is done by visually comparing imagery of 2009 and 2018. The results show potential shifts in benthic composition, which highlights the importance of regular monitoring. This research indicates that integration of up-to-date in situ data with high-resolution satellite imagery has high potential for identifying and classifying benthic habitats around St. Eustatius.