Previous studies have shown that seabed sediment parameters such as %Mud, %Sand,
and %Gravel are useful surrogates for predicting the distribution of benthic species (e.g.,
Beaman and Harris 2007; Degraer et al. 2008). Typically, these parameters are derived
from a limited number of widely distributed sediment grab samples. To improve
predictions from these point data, continuous layers of these parameters are needed.
Apart from often used geostatistic techniques, predictive modelling techniques can be
used for large area mapping. In particular, machine learning models offer most potential
because they are able to handle both linear and non-linear relationships.
Multibeam data with high resolution coverage is now routinely collected in marine
surveys. From multibeam bathymetry we can derive a range of terrain and morphometric
variables that have known relationships with sediment distribution patterns. Multibeam
backscatter intensity depends on both acoustic impedance contrast and the roughness of
the seafloor, which are seabed habitat dependent. Various first and second order texture
measures derived from backscatter data may be useful in predicting sediment. Variables
that measure spatial autocorrelation are also considered to be useful.
This paper reports the results of predictive spatial modeling of two seabed sediment
parameters: %Mud and %Sand for a 700 km
2
area of the Carnarvon Shelf, Western
Australia. Multiple machine learning models were applied to create prediction maps and
prediction uncertainty maps.