Marine habitat maps are essential for marine spatial planning, providing crucial information for conservation and resource management. Accurate classification supports sustainable use and identifies key areas for protection. While machine learning tools are widely used in spatial ecology, deep learning remains less explored. Convolutional Neural Networks (CNNs) have shown promise for habitat classification, while Random Forest (RF) remains a robust and interpretable approach.
This study compares CNN and RF models for high-resolution benthic habitat mapping in Apollo Marine Park, Australia. Both models were trained to classify three distinct habitat types using bathymetric, multibeam backscatter, environmental datasets, and the RF received additional multiscale bathymetric derivatives. Model performance was evaluated using accuracy, precision, recall, F1-scores, and uncertainty metrics, highlighting each method’s strengths and classification challenges.
CNNs excel at automatically extracting spatial features and multiple scales, while RF provides high interpretability with structured inputs. Uncertainty mapping revealed areas requiring refinement, particularly in habitat transition zones with similar acoustic properties. By comparing these approaches, we assess their potential to enhance marine habitat classification and inform marine spatial planning. This study underscores the value of integrating machine learning and deep learning techniques for improved habitat mapping in complex environments.