ALAN: A LEARNING ALGORITHM FOR IDENTIFYING REVERSAL BOUNDARY CROSSINGS IN MARINE MAGNETIC DATA

Crossings of geomagnetic reversal boundaries extracted from marine magnetic profile data are fundamental to reconstructing motions of the Earth’s tectonic plates over the past ~200 Myr. Manually identifying these crossings is a laborious and time-consuming process, but the existence of numerous other sources of seafloor magnetic variation makes automation difficult. We are investigating whether a machine learning algorithm can be trained to reliably identify the signature of reversal crossings in magnetic profile data, which has the potential to more rapidly and reproducibly increase the spatial coverage and resolution of identified reversal crossings.

We have constructed a training dataset of >1000 magnetic reversal boundary signatures, by associating identified magnetic reversals in the Global Seafloor Fabric and Magnetic Lineation Database with the magnetic data from which they were originally derived, and a similar number of “non-reversal” signatures such as those associated with seamounts and fracture zones. Feature vectors generated from half of these data are used to train 3 different types of supervised machine learning classifiers, using the scikit-learn open-source python library: Support Vector Machine, Random Forest Decision Tree, and Multi-Layer Perceptron (neural network). The other half of the training dataset is then used to test the absolute and relative performance of the classification algorithms, by determining the fractions of successfully identified reversal boundaries and false positives. Using a small and minimally processed training dataset, preliminary success rates of >70% have been achieved for all 3 model types. This indicates that a robustly trained machine learning model can indeed be used to rapidly and accurately identify reversal boundary crossings in magnetic profile data; however, further optimization is required.