This paper describes the application of convolutional neural networks (CNNs) to the identification and classification of ten classes of benthic macrofauna in high-resolution photomosaics captured on the Pacific continental shelf by an ROV. Each photomosaic was previously hand-annotated with the location and classification of each animal, providing a training set for the machine learning algorithms. These annotations are used to extract image patches around each contact, resulting in approximately 5000 image samples, which are supplemented with randomly selected image patches representing the background. The resulting corpus of data is used to train a series of convolutional neural networks in the Nvidia DIGITS and Google Tensorflow environments. Due to the relatively sparse nature of the training data set, a number of data augmentation approaches are used to increase the diversity of training data. The performance of the resulting algorithm is evaluated in three problem scenarios: first, classification of fauna in an image patch known to contain a target; second, classification of a given image patch as either background or non-background; and third, a single-pass combination of the two problems. The presented networks prove highly accurate at background/non-background segmentation with ∼ 96% accuracy. Fauna identification is less reliable at ∼ 89% accuracy, and unified segmentation and identification proves to be the most challenging at ∼88% accuracy.