Sea-Floor Basalt Type Classification from Machine Learning
College of Arts and Sciences
Department of Geosciences
Dr. Dylan Mikesell
Machine learning is becoming an increasingly useful tool across many scientific disciplines; however, adoption in the geologic sciences is just beginning. Here we present a new MATLAB-based image classification algorithm used to characterize basalt morphology on the sea-floor from underwater vehicle imagery. In 2016, the submersible Alvin collected seafloor imagery data with both downward and forward looking cameras in the region of 14°N on the Mid-Atlantic Ridge. This resulted in 64 GB of geo-referenced imagery. We will present the current results of our effort to automate the basalt classification process. We will discuss the labeling procedures, as well as the regional convolutional neural network (R-CNN) used for image segmentation and classification. The accuracy of current machine is limited by the number of different basalt types used during the training, but overall has high accuracy (e.g. 96% for pillow basalts). Finally, we will present preliminary results of biological organism classification in the images as well. The ultimate goal of this undergraduate research project is to develop a fully automated classification scheme for seafloor images that not only classifies basalt type, but also things like bio-diversity, sediment structures, and water column turbidity. Using the geo-reference information, this will eventually lead to an improve capabilities to map the oceans and seafloor in an automated way beyond simple bathymetry.
Paustian, John; McCully, Emma; and Wanless, Dorsey, "Sea-Floor Basalt Type Classification from Machine Learning" (2019). 2019 Undergraduate Research and Scholarship Conference. 128.