groundTruther: A statistical approach to accounting for uncertainty in side-scan sonar classification

Objective Side-scan sonar is a valuable tool for mapping benthic habitat, allowing habitat patches (e.g., substrate and large woody debris) within the sonar swath to be mapped and classified with near-complete coverage over large spatial extents. However, like all remote sensing classifications, classifications derived from sonar imagery have inherent uncertainty. Ground-truthing can help quantify classification accuracy, but positional errors introduce additional uncertainty and may reduce confidence in the estimated classification accuracy. We propose an approach to account for positional errors when estimating classification accuracy and propagate uncertainty in patch classification. We provide a worked example of the process and its application to a Lake Sturgeon Acipenser fulvescens habitat suitability index (HSI). Methods We begin by accounting for the magnitude of positional error associated with ground-truthing geographic coordinates, considering two options: (1) subset to retain ground-truthing data with a high degree of confidence that they are within corresponding patches or (2) propagate uncertainty in ground-truthing locations as point clouds of possible locations. Multinomial models are then fit to the subset data or location realizations to estimate the accuracy of sonar classifications. Next, the estimated classification accuracies are used to reclassify patch types multiple times, propagating uncertainty in the composition of patch types. Last, the sets of reclassified patches are used to calculate multiple realizations of the HSI to estimate the median and probable area of suitable habitat within the study extent and areas of interest. Results Because substrates that are difficult to distinguish in sonar imagery (e.g., sand vs. gravel) had similar suitability scores, uncertainty in the overall HSI scores tended to be lower than uncertainty in the substrate classifications. Nevertheless, incorporating classification uncertainty into the calculation of the HSI allowed uncertainties to be maintained in end products used for decision making. Conclusions To provide a reproducible workflow for propagating classification uncertainties in side-scan sonar imagery and other remote sensing data, our approach is formalized as a set of functions available as an R package, groundTruther. This package provides end users with functions to develop habitat maps while accounting for uncertainty to better inform management decisions.