SEABED CLASSIFICATION OF THE SOUTH TASMAN RISE FROM SIMRAD EM12 BACKSCATTER DATA USING ARTIFICIAL NEURAL NETWORKS

We have developed an automated method for sea-floor classification for the South Tasman Rise, based on a SIMRAD EM12-backscatter (13 kHz) mo saic and 47 sea-floor samples. The samples have been divided into 3 di stinct groups representing: (i) thick blankets of foraminiferal ooze; (ii) mixed sediments comprising sand/silt/mud (turbidites/chalk); and (iii) outcrops of metamorphic basement and volcanic rocks. A total of 515 sub-areas, each measuring 32x32 pixels (similar to 4 km(2)) and re presenting the different seabed types, were extracted from the image f rom areas 128x128 pixels large, centred on the sample locations. The t exture of the sub-images was analysed by calculating grey-level run-le ngth features, spatial grey-level dependence matrices, and grey-level difference vectors in four directions. A total of 100 samples for each class and 18 feature statistics were chosen to train an artificial ne ural network to recognise the textural attributes and their variabilit y for each class. The performance of the network was evaluated first b y classifying the image sub-areas used for training, followed by class ification of the remaining sub-areas. Classification accuracies for th e training samples for ooze, sand/silt/mud and basement rocks are 98%, 98% and 91%, respectively, and 91%, 83% and 84% for the test samples. Subsequently, we classified a total of more than 20 000 unknown image sub-areas 4 km(2) large on the northern South Tasman Rise. The result agrees well with a visual geological interpretation of the sidescan m osaic as well as with a facies map of the area based on 3.5 kHz data. The success of a combined textural image and neural network analysis t o classify a sidescan mosaic, in the presence of noise and processing artifacts, suggests a wide range of potential applications including t he recognition of sediment textures and other objects in high-resoluti on, shallow-water backscatter images, and pattern recognition in remot ely sensed geophysical images for mineral exploration.