Using principal component analysis and canonical discriminant analysis for multibeam seafloor characterisation data

USING PRINCIPAL COMPONENT ANALYSIS AND CANONICAL DISCRIMINANT ANALYSIS FOR MULTIBEAM SEAFLOOR CHARACTERISATION DATA ZBIGNIEW àUBNIEWSKI, ANDRZEJ STEPNOWSKI GdaĔsk University of Technology, Department of Geoinformatics Narutowicza 11/12, 80-233 GdaĔsk, Poland The paper presents the seafloor characterisation based on multibeam sonar data. It relies on using the integrated model and description of three types of multibeam data obtained during seafloor sensing: 1) the grey-level sonar images (echograms) of seabed, 2) the 3D model of the seabed surface which consists of bathymetric data, 3) the set of time domain bottom echo envelopes received in the consecutive sonar beams. The classification is performed by utilisation of several statistical methods applied for analysis of a set of seafloor descriptors derived from multibeam data. In the paper, the use of Principal Component Analysis (PCA), as well as Canonical Discriminant Analysis (CDA) for reduction of the seafloor parameter space dimension is presented along with the obtained results. In addition, the use of the open source World Wind Java SDK tool for implementation of imaging and mapping of seafloor multibeam data, integrated with other elements of a scene and overlaid on rich background data, is also shown. INTRODUCTION Multibeam sonars are widely used in applications like high resolution bathymetry measurements, underwater object detection and imaging, etc. Also, they are the promising tool in seafloor characterisation and classification, having several advantages over conventional single beam echosounders. The proposed approach to seafloor classification relies on the combined use of three different techniques. In each of them, a set of descriptors foreseen to be applied in seabed classification procedure, is calculated using a given type of data obtained from multibeam sonar system: 1) the grey-level sonar images of seabed, 2) the 3D model of the seabed surface which consist of bathymetric (x, y, z) points, 3) the set of time domain echo envelopes received in the consecutive beams. 123 1. MATERIALS AND METHODS The schematic concept of the applied approach was shown in Fig. 1. In the first technique used, i.e. Method 1 in the figure, the grey-level sonar echograms of seabed surface are utilised [1]. Usually, such images are generated by a multibeam sonar firmware. Next, a set of parameters describing the local region of sonar image is calculated for each bottom type. The parameters set include: 1. Basic statistical parameters describing the grey level distribution, i.e. local mean (MEAN) and standard deviation (STD). 2. Slope of the autocorrelation function of a grey level (in along track direction) approximated for a local region of the image (SL_AUTC). 3. Texture analysis parameters based on the Grey-Level CO-occurrence Matrix (GLCM) of a sonar image local region: entropy (ENTR) and local homogeneity (HOMOG). This technique description may be found in [1]. Set of beam echoes Method 3 Feature extraction and averaging Multibeam sonar ... Angular dependence estimation Investigated seafloor Seabed classification Method 2 Seabed surface Feature extraction and averaging Method 1 Seabed imagery Feature extraction and averaging Fig.1. The concept of three combined methods of seafloor classification using multibeam sonar. In the second technique of multibeam sonar data processing (Method 2 in Fig. 1), the 3D “bathymetric” model of seabed surface is utilised [1]. It is constructed as a set of (x, y, z) points obtained from the detected bottom range for each beam, within the multibeam sonar seafloor imaging procedure. The examples of seabed surface model obtained for two bottom 124 types, e.g. mud and coarse grained sand, are presented in Fig. 2. Next, for the local region of the constructed seabed surface, among some others, the following descriptors are calculated, viz.: rms height (SURF_RMS) and the slope of the seabed surface autocorrelation function (SURF_AUTC). In the third technique of multibeam sonar data processing (Method 3 in Fig. 1), the set of echo signal envelopes received in the particular beams is analysed [2]. The data processing procedure in this method is more complex than in two previous ones. Firstly, after detection of a bottom echo in the received signal, the set of echo parameters is calculated for an appropriate part of each beam echo. The parameters include: 1. The normalised moment of inertia I of the echo envelope, with respect to the axis containing its gravity center [3]. 2. Fractal dimension D of an echo envelope, interpreted as a measure of its shape irregularity. It is calculated as a box dimension approximation, as described in [2]. Next, for each seabed type, the dependence of I and D parameter values of the particular beam incident angle is estimated, and then, for the application in seafloor classification procedure, the following parameters are calculated for each sounding (swath): 1) the approximated slope of the angular dependence of the beam echo moment of inertia I(M), for the angle range of [2°, 17°] (I_SLOPE), and 2) the same approximated slope for the beam echo fractal dimension D(M), for the angle range of [4°, 19°] (D_SLOPE). 0.6 0.6 0.4 0.4 0.2 0.2 0 0 -0.2 -0.2 -0.4 -0.4 -0.6 -0.6 80 80 80 70 70 60 60 80 70 70 60 60 50 50 50 50 40 a) 40 b) Fig.2. The examples of seabed surface model obtained for two bottom types: a) mud, b) coarse grained sand. Axes in meters. Finally, using the results obtained by the techniques described above, the 2D plots of calculated values for selected pairs of echo parameters were constructed. The obtained results were reported in [1], [2] and [4]. Sample result is presented in Fig. 3. The field experiment summarizes as follows. The data used in the experimental verification of the proposed approach were acquired using Kongsberg EM 3002 sonar in the Gulf of GdaĔsk region of the Southern Baltic from 2007 to 2009. Several sites of different seafloor types were investigated, but the results of the current investigation refer to 4 selected sites, characterised by the following true seabed types: mud, anthropogenic sand and mud, fine grained sand, and coarse grained sand. The sonar operating frequency was 300 kHz, the beamwidth was 1.5° x 1.5°, the transmitted pulse length: 0.15 ms, the echo sampling rate: 14.3 kHz. The bottom depth was in 125 a range between 10 m and 100 m. Approximately, 1000 swaths from each of four seafloor types were processed. For each swath, 160 beams covered the angle sector from -65° to 65°. In the first – “imaging” technique, the seabed sonar image part corresponding to the beam angle sector between 15° and 30° was selected for further processing. In the estimation of mean, standard deviation, skewness and kurtosis of an image grey level, the size of a local image region was chosen as 11 x 11 pixels. The same local region size was used for entropy and local homogeneity calcu (...truncated)