Context
Benthic habitat mapping is crucial for effective marine spatial planning. Despite advancements in multibeam echosounder (MBES) technology, selecting appropriate classification methods to accurately map seafloor habitats remains a challenge.
Objectives
This study aims to provide novel comparisons of large spatial scale habitat classifications using pixel-based (PB) and object-based image analysis (OBIA) methods, applied within a hierarchical random forest framework, to classify benthic biotopes in the northern section of Apollo Marine Park and the adjacent Cape Otway coastline, Victoria, Australia.
Methods
We utilised high-resolution MBES-derived data, implementing a hierarchical random forests algorithm to classify benthic habitats. The PB method treated each pixel independently, allowing for high spatial detail, while the OBIA method grouped pixels into meaningful segments for classification. Prior to segmentation, backscatter data from two different MBES systems were harmonised using a bulk shift method (Misiuk et al., 2020) to ensure consistency across datasets. We then applied the Supercells segmentation technique (Nowosad 2022) to the harmonised backscatter data, forming the foundation for the OBIA-based classification. Both methods were evaluated using accuracy, F1 scores, and uncertainty maps were generated to assess classification reliability.
Results
Both classification methods demonstrated strong performance, with no statistically significant differences in overall accuracy. However, the complexity of the habitat maps varied: the PB approach excelled in capturing fine-scale habitat details, beneficial for management and conservation efforts requiring high detail. Conversely, the OBIA method produced more interpretable and less complex maps, suitable for general spatial analyses, though it resulted in the omission of some minority classes.
Conclusion
This study emphasises the importance of defining the desired level of complexity in habitat maps before analysis, ensuring that chosen methods yield maps suitable for specific applications—particularly in datasets with strong class imbalances. Future advancements in machine learning and emerging technologies have the potential to further refine habitat mapping techniques and enhance classification accuracy.
