Improving coral monitoring by reducing variability and bias in cover estimates from seabed images

• Published in: Progress in oceanography Vol. 222; p. 103214
• Main Authors: Curtis, Emma J., Durden, Jennifer M., Bett, Brian J., Huvenne, Veerle A.I., Piechaud, Nils, Walker, Jenny, Albrecht, James, Massot-Campos, Miquel, Yamada, Takaki, Bodenmann, Adrian, Cappelletto, Jose, Strong, James A., Thornton, Blair
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.03.2024
• Subjects: Conservation; Data imputation; Environmental monitoring; Image annotation; Ocean floor; Underwater photography; Underwater photography; Environmental monitoring; Data imputation; Image annotation; Ocean floor; Conservation
• ISSN: 0079-6611; 1873-4472
• DOI: 10.1016/j.pocean.2024.103214

Seabed cover of organisms is an established metric for assessing the status of many vulnerable marine ecosystems. When deriving cover estimates from seafloor imagery, a source of uncertainty in capturing the true distribution of organisms is introduced by the inherent variability and bias of the annotation method used to extract ecological data. We investigated variability and bias in two common annotation methods for estimating organism cover, and the role of size selectivity in this variability. Eleven annotators estimated sparse cold-water coral cover in the same 96 images with both grid-based and manual segmentation annotation methods. The standard deviation between annotators was three times greater in the grid-based method compared to segmentation, and grid-based estimates from annotators tended to overestimate coral cover. Size selectivity biased the manual segmentation; the minimum size of colonies segmented varied between annotators fivefold. Two modelling techniques (based on Richard’s selection curves and Gaussian processes) were used to impute areas where annotators identified colonies too small for segmentation. By imputing small coral sizes in segmentation estimates, the coefficient of variation between annotators was reduced by approximately 10%, and method bias (compared to a reference dataset) was reduced by up to 23%. Therefore, for sparse, low cover organisms, manual segmentation of images is recommended to minimise annotator variability and bias. Uncertainty in cover estimates may be further reduced by addressing size selectivity bias when annotating small organisms in images using a data-driven modelling technique. [Display omitted] •We assess annotator error in cover estimation from images with sparse coral cover.•Annotator variability in grid-based estimates was three times that of drawn areas.•Grid-based results overestimated coral cover by 45% above the reference dataset.•For drawn areas, missing sizes of smallest colonies led to 38% cover underestimation.•Model estimation of missing coral sizes reduced annotator variability by 10%.