Automated Textural Image Analysis of Seabed Backscatter Mosaics: A Comparison of Four Methods

Four seabed backscatter image classification methods that differ substantially in their approach are compared, using a dataset from Sydney Harbour. The first method is based on GeoAcoustic’s commercial GeoTexture software. GeoTexture relies on the user to extract features from the image to train the system to recognize textures based on a statistical method, before classifying the entire image. GeoTexture achieved classification accuracies for the sand, gravel and mud classes of 47%, 67% and 77%, respectively. The second method characterizes textures in the wavelet-domain utilizing Hidden Markov Models (HMM) and uses a Bayesian method for image segmentation. In this approach, subimages, 512 x 512 pixels large, are segmented into three classes of sediment. The best supervised classification success for this method was relatively poor with 100% for gravel, but only 29% for sand and 36% for mud. The third and fourth methods use a neural network-based approach where two different techniques are compared for feature extraction: (I) a space-domain method based on grey-level run-length features, spatial grey-level dependence matrices and grey-level difference vectors in four directions, and (II) the grey-level co-occurrence iteration algorithm (GLCIA) method, which is far superior in terms of computational speed, but relies on a smaller number of feature vectors. Method I provides accuracies of 92% and 89% for the gravel class, compared with accuracies for the sand, gravel and mud classes of 88%, 77% and 78%, respectively, for method II. Both the GeoTexture and neural network-based approaches were found to be superior to the HMM approach, which is not mature enough for application to seabed images, at this time.