Comparing convolutional neural network and random forest for benthic habitat mapping in Apollo Marine Park

Abstract Marine habitat maps are essential tools for marine spatial planning, providing information for decision‐making in conservation and resource management. Accurate classification of benthic habitats supports their sustainable use and identifies key areas for protection. Convolutional neural networks (CNNs) are powerful deep learning algorithms that have shown promise for advancing habitat classification tasks and mapping complex marine environments. This study compares the performance of a CNN and a Random Forest (RF) model in classifying benthic habitats within Apollo Marine Park, Victoria, Australia. Models were trained to classify three distinct habitat types using bathymetry, multibeam backscatter, wave height and positioning data; however, the RF model had access to 100 additional bathymetric derivatives, of which 10 were selected as predictors. The CNN achieved an overall accuracy of 67.32%, while the RF model achieved 62.57%. For individual habitats, the CNN obtained F1‐scores of 0.664 for high energy circalittoral rock with seabed‐covering sponges, 0.538 for low complexity circalittoral rock with non‐crowded erect sponges and 0.774 for infralittoral sand and shell mixes. The corresponding RF scores were 0.598, 0.506 and 0.739. Both models encountered challenges in classifying transitional habitat zones, where diffuse boundaries between habitat types led to overlaps and shared acoustic properties. However, the CNN demonstrated an advantage due to its ability to automatically analyse spatial patterns across multiple scales. In contrast, while the RF model incorporated terrain attributes that capture local variation, its ability to utilize spatial context was constrained to predefined scales of the derived features. The CNN's ability to leverage spatial relationships resulted in clearer and more coherent habitat maps, reducing the salt‐and‐pepper effect commonly observed in pixel‐based classifications. This study highlights the potential of CNNs for marine habitat mapping through their ability to classify data derived from multibeam bathymetry, while also identifying avenues for further refinement to enhance their utility in marine spatial planning tasks.