Congratulations to Lucy Dyer for a successful MS thesis defense!

Congratulations to Lucy Dyer for successful defending her MS thesis, ‘Identifying marine magnetic anomalies using machine learning‘.

Abstract

Magnetic reversal boundaries identified from marine magnetic surveys are used to date the oceanic lithosphere and are a key source of information for reconstructing oceanic ridge spreading rates and past plate motion. However, the identification process is tedious and time consuming, which results in incomplete and inconsistent reversal identification along ridge systems and between ocean basins. This study investigates the feasibility of using machine learning to automatically identify the signature of magnetic field reversals in marine survey data. A training dataset of ~10,000 known reversal boundaries, seamounts, fracture zones, and other sources of magnetic variation on the ocean floor was generated by reassociating the geographic locations of these features with the magnetic and (where available) topographic data on their source tracklines. From this main dataset, 8 filtered training datasets with different constraints on data resolution, bathymetric coverage and validation status were also produced, and used as the inputs for 3 separate machine learning algorithms: a Support Vector Machine, a Random Forest Decision Tree, and a Neural Network. The composition of the input feature vectors was varied by changing the sampling resolution and adding or omitting the latitude, longitude, and azimuth of the ship trackline. Parameter tuning with cross-validation optimized the performance of the machine learning algorithms with each combination of training dataset and feature vector configuration. Overall, Random Forest Decision Tree models produced the best classification performance, with a maximum balanced accuracy of ~81% and a maximum precision of ~92%. Neural Network models achieved a maximum balanced accuracy of ~77% and a maximum precision of ~88%, and Support Vector Machine models achieved a maximum balanced accuracy of ~73% and a maximum precision of ~87%. Comparison of test and training dataset scores found little evidence of overtraining. Parameter tuning generally had a small effect on model performances, with the exception of kernel type in Support Vector Machines. Much more important was the selection of training dataset, where inclusion of bathymetry and exclusion of low-resolution data, and manual validation, improved classification performance despite sometimes substantially reducing the number of included features. Including bathymetry or geographic information in the feature vector also significantly improved classification performance. These results confirm that machine learning can be used to identify geomagnetic reversal boundaries in oceanic lithosphere, and the training datasets and processing code developed over the course of this project establish a solid foundation for doing so. Further work would include validating all points within the datasets, adding more non-reversal anomalies, and possibly exploring other aspects of machine learning such as Deep Learning.

Lucy has worked really hard over the past two and a half years to start amassing a robust training dataset containing the magnetic and bathymetric signatures of known reversal crossings and other ocean floor features such as seamounts and fracture zones, and showing that they can be used to train machine learning models to identity new features on the seafloor.

The defense had to take place virtually, so unfortunately celebratory photos are thin on the ground. But congratulations, Lucy!