Fish Recognition in Underwater Environments using Deep Learning and Audio Data

Environmental conservation is a field where AI can provide significant help for many types of tasks. Oil, plastic, anthropogenic noise, overfishing and global warming are known to affect marine ecosystems (flora, fauna) inducing a drastic decrease of marine biodiversity and ecosystem services. The assessment of marine animals' distribution could benefit from automatic recognition of the presence of a species in a specific location. For this purpose, the passive acoustics monitoring can use underwater audio recordings and try to recognize the sound produced by the species. This work compares the performance of classical computer vision algorithms and modern deep learning methods for the task of identifying if a spectrogram contains the characteristic sound produced by the brown meagre. An accuracy of 95% was achieved using a deep convolutional neural network based on a recent architecture that was partially pretrained, outperforming classical computer vision algorithms.