Image-Based Estimation of Cubic Artificial Reef Settlement and Projected Remaining-Area Ratio Using Mask R-CNN

Cubic artificial reef (CAR) modules may undergo settlement, inclination, and burial after deployment, requiring non-contact assessment. This study presents an image-based framework combining Mask R-CNN instance segmentation with three geometric methods for estimating vertical settlement and the projected remaining-area ratio. The framework was evaluated using images from 1:20-scale flume experiments under controlled, clear-water, near-frontal conditions. An additional mask-level evaluation used 14 unaugmented held-out images containing 38 annotated instances from source videos reported by the dataset curator as excluded from model development. The retained checkpoint achieved an overall mask AP of 0.896, with AP50 and AP75 both equal to 1.000. For uniform settlement, Method 1 produced valid estimates for seven of ten modules, with a maximum absolute normalized settlement error of 14.90%, whereas Method 2 produced estimates for all ten modules, with a maximum absolute error of 5.71%. Method 3 evaluated nine modules; normalized area errors ranged from −8.25 to −0.67 percentage points, with a mean signed difference of −4.50 percentage points. Method 3 quantifies a two-dimensional projected remaining-area ratio rather than physical inclination or three-dimensional exposed area. These results demonstrate segmentation and geometric-estimation feasibility for a curator-confirmed held-out dataset within a controlled laboratory domain; full-scale field performance remains to be validated.