Image classification for maerl and sandy seabed percentage coverage estimation

Abstract

Analysis of the seabed composition over a large spatial scale is always an interesting yet very challenging task, Apart from the field work involved, hours of video footage captured by cameras mounted on Remote Operated Vehicles (ROVs) have to be reviewed by an expert in order to classify the seabed topology and to identify potential anthropogenic impacts on sensitive benthic assemblages, Apart from being time consuming, such work is highly subjective and, through visual inspection alone, a quantitative analysis is highly unlikely to be made, This current study investigates the applicability of Machine Learning Techniques to analyse images extracted from footage of the seabed, Frames were recorded by a down-facing camera during an ROV survey of maerl beds on a sandy seabed, conducted as part of an Environment Impact Assessment (EIA) on the laying of an electrical cable (interconnector) between the Maltese Islands and Sicily in the Central Mediterranean, A training data set was initially defined by using the recorded intenSity of the three colour channels, Pixels were then assigned to one of two classes representing sand patches or maerL A supervised decision tree algorithm was then used to construct a model that can automatically classify unseen images, Initially, the model was applied on pixels with a known target value, and the Pearson and Spearman correlation values were computed and found to be 0,99, indicating a good agreement between the true and predicted values, To further improve on the results, before classification, image enhancement in Fourier space was applied to remove the periodic noise, Results achieved for images taken over three different seabed stretches of 500m each are presented in this study, By assuming constant speed and depth of the ROV above the ocean fioor, the geographical position of each frame, as well as the area corresponding to each grid cell in the image, was estimated, The total areas covered by sand and maerl were then given for the locations corresponding to each of the processed frames, along with the percentage area covered by each type of seabed assemblage, This study will be repeated after the laying of the electrical cable in question, and the methodology performed should prove useful in making quantitative comparisons with the maerl coverage statistics collected in the baseline study, This will be achieved by estimating any potential regression in maerl coverage as a result of the cable-laying process and might potentially shape the coastal planning processes leading to the laying of similar marine infrastructure in future,