Seafloor classification with optimized texture convolutional neural network: application to complex benthic habitats on seamount and volcanic island slopes

International audience Seafloor heterogeneity is a key parameter structuring seamount biodiversity at multiple spatial scales. However, substrates characterization at very small (meter) scale from manual processing of optical images is time-consuming and difficult to be objectively applied on a large dataset. Consequently, quantification of several substrate types per image is usually lacking.To automize seafloor classification, we proposed to use a supervised machine learning algorithm. First, a substrates dataset with 772 annotations composed of four classes was produced (volcanic, carbonate, gravel, and soft) and splitted randomly into two parts: 90% for training, 10% for validation. A biogenic class was also extracted but set aside because of its small size towards others. Note that several annotations refinements were made to respect computer vision standards.To classify theses substrates, we used a texture specialized convolutional neural network based on wavelets [1]. For the training step we used a pretrained VGG19 network as basis and we applied a transfer learning approach to tune our network for our substrates dataset. We artificially increased the number of annotations by applying a data augmentation technique. The wavelets network was improved, and we obtained a classification rate close to 84% on validation set. Finally, due to the seafloor heterogeneity, an image segmentation process based on sliding windows was applied to extract statistics over tests images with a 224x224 patch size.This substrate classification algorithm was used to assess percent cover of the four substrate classes per image from nine towed camera surveys (i.e. 9000 images) on seven seamounts and volcanic island slopes along the Mozambique Channel. From this dataset, the variability of substrate composition along and among these seamounts was tested as structuring factors of benthic megafaunal assemblages (Hanafi-Portier et al., this meeting).