Satellite remote sensing and bathymetry co-driven deep neural network for coral reef shallow water benthic habitat classification

Shallow-water benthic habitat classification of coral reefs based on satellite remote sensing is an important part of coral reef monitoring. Leveraging its potent capacity for feature learning, and generalization, deep learning emerges as a robust method for coral reef benthic habitat classification. Due to the complexity of the marine environment, it is difficult to produce high-quality pixel-by-pixel labels for deep learning-based methods, which makes it challenging to recover structural details of coral reef benthic habitats. Bathymetry data can provide spatial contextual information and geometric features, serving as auxiliary features to provide abundant structural information for benthic classification models. However, how to use bathymetry and what kind of bathymetry features to employ for assisting model learning remains to be explored. Therefore, a bathymetry feature fusion-weakly supervised coral reef benthic habitat classification model (BFFBHCM) is proposed. BFFBHCM is supervised by sparse benthic habitat samples with bathymetry and can generate dense, multi-scale bathymetry features. With the robust bathymetry-benthic feature fusion module (B-BFFM), BFFBHCM can consider both semantic and structural details of the benthic habitats, thus generating highly accurate benthic habitat classification results. Experiments were conducted using the NJUReef + dataset containing 10 coral reefs in the Spratly Islands, China, constructed based on in-situ data. Comprehensive experimental results demonstrate that the proposed BFFBHCM is insensitive to the vertical error in bathymetry, with an average mIoU 22.54 % higher than state-of-the-art methods. Furthermore, it outperforms the weakly-supervised method that excludes bathymetry by 10.14 %, and still exhibits generalization to coral reefs in different regions around the world.