Combining observations with acoustic swath bathymetry and backscatter to map seabed sediment texture classes: The empirical best linear unbiased predictor

Abstract

Seabed sediment texture can be mapped by geostatistical prediction from limited direct observations such as grab-samples. A geostatistical model can provide local estimates of the probability of each texture class so the most probable sediment class can be identified at any unsampled location, and the uncertainty of this prediction can be quantified. In this paper we show, in a case study off the northeast coast of England, how swath bathymetry and backscatter can be incorporated into a geostatistical linear mixed model (LMM) as fixed effects (covariates).Parameters of the LMM were estimated by maximum likelihood which allowed us to show that both covariates provided useful information. In a cross-validation, each observation was predicted from the rest using the LMMs with (i) no covariates, or (ii) bathymetry and backscatter as covariates. The proportion of cases in which the most probable class according to the prediction corresponded to the observed class was increased (from 58% to 65% of cases) by including the covariates which also increased the information content of the predictions, measured by the entropy of the class probabilities. A qualitative assessment of the geostatistical results shows that the model correctly predicts, for example, the occurrence of coarser sediment over discrete glacial sediment landforms, and muddier sediment in relatively quiescent, localized deep water environments. This demonstrates the potential for assimilating geophysical data with direct observations by the LMM, and could offer a basis for a routine mapping procedure which incorporates these and other ancillary information such as manually-interpreted geological and geomorphological maps.