When training models for %Mud, we discovered that BDT and SVM need only one
secondary variable: Local Moran I for Bathymetry to obtain the best performance. GRNN
used secondary variables of Homogeneity, Local Moran I for Backscatter, Variance, and
Planar Curvature to reach the best performance. For %Sand, the best performing model,
BDT, did not use Slope, Surface Area, TPI and Homogeneity in East direction. The best
performing SVM model used the secondary variables of Homogeneity in South-East
direction, Slope, Variance, Local Moran I for Bathymetry, Surface Area, Homogeneity in
North direction, and Relief. Three secondary variables including Local Moran I for
Bathymetry, Local Moran I for Backscatter, and Homogeneity in South-East direction
were used to obtain the best performing GRNN. The findings indicate the importance of
spatial autocorrelation in mapping seabed sediment parameters.
As an example, Figures 1 and 2 display the prediction maps for Point Cloates, which
show similar spatial patterns among the three models for both sediment parameters.
Percentage mud generally increases with water depth (Figure 1A-C). Figure 1D indicates