Seabed Classification Using BP Neural Network Based on GA

Side scan sonar imaging is one of the advanced methods for seabed study.In order to be utilized in other projects,such as ocean engineering,the image needs to be classified according to the distributions of different classes of seabed materials.In this paper,seabed image is classified according to BP neural network,and Genetic Algorithm is adopted in train network in this paper.The feature vectors are average intensity,six statistics of texture and two dimensions of fractal.It considers not only the spatial correlation between different pixels,but also the terrain coarseness.The texture is denoted by the statistics of the co-occurrence matrix.Double Blanket algorithm is used to calculate dimension.Because a uniform fractal may not be sufficient to describe a seafloor,two dimensions are calculated respectively by the upper blanket and the lower blanket.However,in sonar image,fractal has directivity,i.e.there are different dimensions in different direction.Dimensions are different in acrosstrack and alongtrack,so the average of four directions is used to solve this problem.Finally,the real data verify the algorithm.In this paper,one hidden layer including six nodes is adopted.The BP network is rapidly and accurately convergent through GA.Correct classification rate is 92.5% in the result.