Area calibration in automated subsurface coral cover counting

Abstract

A machine vision-based system that can classify benthic objects and estimate benthic area cover from video taken by a boat-towed submersible camera has been developed. The camera is submerged up to 0.5 meters beneath the water surface and as the boat moves video of the reef floor is captured. A simple model to convert pixel to physical area from camera optics and depth information is presented which eliminates the need for area calibration using submerged reference objects. Color and texture features derived from normalized chromaticity coordinates (NCC) and Local Binary Patterns (LBP) were used as descriptors for the reef categories and Linear Discriminant Analysis was used as classifier Classification success rates of 86% and 84% as compared with ground truth were obtained for living and nonliving components, respectively from underwater videos of Ngederrak Reef, Palau. Our novel system in automating benthic classification can serve as groundwork for performing faster reef surveys with minimum and less expensive resources and diving manpower.