Comparing pixel-and object-based approaches for classifying benthic habitats

Benthic habitat mapping is essential for effective marine spatial planning, yet choosing appropriate classification methods remains challenging. This study compares pixel-based (PB) and object-based image analysis (OBIA) approaches within a hierarchical random forest framework to classify benthic biotopes in Apollo Marine Park and along the Cape Otway coastline, Victoria, Australia. High-resolution multibeam echosounder (MBES)-derived bathymetry and backscatter data were used, with backscatter harmonised across systems via a bulk shift method. The Supercells segmentation technique supported OBIA-based classifications. We assessed performance across three hierarchical biotope levels (BC2–BC4) using accuracy, F1 scores, and prediction certainty. Both PB and OBIA approaches showed strong classification performance with no statistically significant differences in accuracy (all p > 0.14). OBIA models achieved higher F1-scores at the intermediate classification level (BC3: F1 = 0.88) compared to PB (F1 = 0.82), whereas PB outperformed OBIA at the coarsest level (BC2: PB F1 = 0.86 vs. OBIA F1 = 0.79) and at the most detailed level (BC4: PB F1 = 0.73 vs. OBIA F1 = 0.68). Accuracy differences across levels were minimal (1–1.5 %). Spatially, both approaches identified consistent biotope patterns, but differed in map texture: PB models captured fine-scale detail with noticeable fragmentation, while OBIA maps were more generalised and spatially coherent. Certainty maps indicated comparable model confidence. These results highlight the need to balance spatial detail with interpretability when selecting classification methods. In contexts with complex seafloor structure and class imbalance, aligning the classification approach with the intended map use is critical for supporting marine spatial planning stakeholders, including conservation managers, policymakers, and environmental consultants, in making informed decisions. • We compare Pixel-Based (PB) and Object-Based Image Analysis (OBIA) methods for benthic habitat classification. • A hierarchical Random Forest approach was applied using multiscale bathymetric and backscatter features. • Supercells segmentation was used to generate OBIA backscatter datasets for habitat mapping. • Pixel-based and object-based models achieved similar accuracy, with OBIA clustering resulting in more generalised habitat maps and reduced ground-truth sample independence. • Differences between pixel- and object-based classifications highlight important trade-offs that management and policy makers should consider when selecting habitat mapping approaches.