Automatic identification of seamounts using neural networks trained for bathymetric, gravity and VGG data

Abstract

Building on the method of Valentine et al. [2013], we explore the influence of data types and pa- rameters on a learning algorithm that can be used to automatically identify seamounts. Knowl- edge about the distribution of these undersea volcanoes is important for several disciplines, such as oceanography, marine biology, geology and in the future possibly for the economy. A special kind of neural network (an ‘autoencoder’) is trained to learn the characteristics of seamounts in particular datasets: global bathymetry, free-air gravity anomaly and vertical grav- ity gradient (VGG) datasets. There are lots of possibilities for the pre-processing, network design and post-processing. In this study we investigate the influence of four different window functions used to preprocess the data. We found that the network output is sensitive to the window function used. Secondly, we vary the compression rate of the encoding, for encoding dimensions between 128 and 32. Furthermore, we explored the influence of several post-processing filter parameters, which highly influence the resulting set of candidate seamounts. The automatically identified seamounts of all three networks are compared to handpicked sea- mounts in two demonstration regions, one in the Pacific Ocean and one in the Atlantic Ocean. These results are quantified by two ratios: success rate (SR) and false positive rate (FPR). All three networks have comparable performance for the Pacific region, although the best results are obtained using the bathymetric data (SR/FPR: 77.0%/8.7%). The performance in the more complex Atlantic region is best for bathymetry, 77.0%/26.8%, then for VGG, 71.8%/40.3%, and least favourable for gravity, 74.1%/59.2%. The results of all three networks in the Atlantic region compares favourably with existing catalogues because of the lower FPRs.