If the
angle in a bin is φ , then the value α=φ−φ¯/σφ is computed, where φ¯ is the mean angle and σφ its standard deviation in all the bins located at the same depth as the bin considered. Only those angles within two standard deviations around the mean (i.e. |α| < 2) have been taken into account in the analyses. These values were quantised to four values corresponding to the four intervals [− 2, − 1], [− 1, 0], [0, 1] and [1, 2]. The procedures for the echogram loading and the computation of the Haralick variables were implemented in the Octave language and are available on the website http://www.kartenn.es/downloads. Energy-based acoustic classification. Based on the volume backscatter of the sound wave, a mTOR inhibitor classification of the data could be tested using the roughness and hardness acoustic indexes. These indexes are computed from the first and second acoustic bounces respectively, and have been introduced as seabed features (Orłowski 1982). The first echo energy (E1) is computed as the time integral of the received backscattered energy corresponding to the diffuse surface reflection (i.e. without the leading selleck compound increasing power signal). The second echo energy (E2)
is computed as the time integral of the entire second bounce signal. Both energies are normalised by depth applying the correction + 20 log(R), where R is the range. This approach using two variables was introduced for seabed classification by Burns et al. (1989) and is currently used by the commercial system RoxAnn (Sonavision Limited, Aberdeen, UK). Multivariate statistical analysis. The multivariate statistical method used was based on Legendre et al. (2002) and Morris & Ball (2006) and includes dimensional find more reduction, principal component analysis (PCA)
and clustering analysis of the reduced variables. The original variables included in the analysis were the energy variables (E1, E2) and the alongship and athwartship Haralick variables, corresponding to Type 1 and Type 2 textural features. The matrix of Haralick textural features was centred and normalised and the PCA was applied (using singular value decomposition whenever more variables than samples were available) to obtain new uncorrelated variables (independent components). Only those components having eigenvalues larger than 1 were kept for the subsequent hierarchical cluster analysis (known as Kaiser’s rule). This choice removes noise from the analysis retaining only variables having higher variance than the original (normalised) ones. The clustering analysis of these selected principal component variables was performed using an agglomerative nested hierarchical algorithm to generate dendrograms; complete linkage and Euclidean distances were used. Finally, a stability analysis, based on Jaccard’s similarity values (J-values) was used to test the significance of these clusters, i.e.