Reproducibility of single-beam acoustic seabed classification

Single-beam acoustic seabed classification continues to be a popular method for mapping seabeds and their sediments. Modern methods can generate maps of acoustic classes that are useful and reasonably accurate. Research toward improved methods continues. A continuing impediment to this research is ranking the accuracy of maps produced by new methods. Non-acoustic data, or ground truth, is usually sparse compared to the detail of the acoustic survey, which can mean that ranking maps for accuracy can be inconclusive. Here we present a new tool for ranking classification maps, namely the reproducibility of acoustic classes from repeated surveys of the same area on different days. Methods that have high reproducibility achieve that by capturing echo characteristics that are strongly influenced by seabed type while suppressing details that are driven by sea state or the water column. Six surveys, done with a 50 kHz sounder over a pair of transects near Miami, FL, USA, between 1 May and 13 August 2007, were used to evaluate two questions. First, how reproducible were classifications of this dataset using QTC IMPACT¿ (Quester Tangent Corporation)? Second, can classification be improved with adjustments to the standard IMPACT processing? Reproducibility was quantified with the overall accuracy and Kappa statistics, which are both derived from the confusion matrix whose rows and columns are numbers of sites with particular class assignments under distinct circumstances.