Assessing the Repeatability of Automated Seafloor Classification Algorithms, with Application in Marine Protected Area Monitoring

Abstract

The number and areal extent of marine protected areas worldwide is rapidly increasing as a result of numerous national targets that aim to see up to 30% of their waters protected by 2030. Automated seabed classification algorithms are arising as faster and objective methods to generate benthic habitat maps to monitor these areas. However, no study has yet systematically compared their repeatability. Here we aim to address that problem by comparing the repeatability of maps derived from acoustic datasets collected on consecutive days using three automated seafloor classification algorithms: (1) Random Forest (RF), (2) K–Nearest Neighbour (KNN) and (3) K means (KMEANS). The most robust and repeatable approach is then used to evaluate the change in seafloor habitats between 2012 and 2015 within the Greater Haig Fras Marine Conservation Zone, Celtic Sea, UK. Our results demonstrate that only RF and KNN provide statistically repeatable maps, with 60.3% and 47.2% agreement between consecutive days. Additionally, this study suggests that in low-relief areas, bathymetric derivatives are non-essential input parameters, while backscatter textural features, in particular Grey Level Co-occurrence Matrices, are substantially more effective in the detection of different habitats. Habitat persistence in the test area between 2012 and 2015 was 48.8%, with swapping of habitats driving the changes in 38.2% of the area. Overall, this study highlights the importance of investigating the repeatability of automated seafloor classification methods before they can be fully used in the monitoring of benthic habitats.