New methods for the interpolation and interpretation of lineaments in aeromagnetic data

Abstract

Aeromagnetic data is one of the most widely collected types of geophysical data. In mineral exploration it can assist in mapping geological features, as well as indicate potential locations of economic interest. Due to the method in which aeromagnetic surveys are flown, an interpolation process must be completed before any map-based interpretation can be accomplished. One artifact common to many interpolation methods is that of “beading”, which is a discontinuous sequence of circular magnetic features that are at acute angles to the traverses, often caused by thin, linear geologic features such as dykes. Developing an interpolation method that “trends” or images these beads as continuous features on magnetic images would allow automatic and reliable quantitative methods to be used for interpretation by geologists and geophysicists. First, a new interpolation method was developed for aeromagnetic data. Utilizing a Taylor derivative expansion and structure tensors, it iteratively enhances trends evident across flight lines to manifest as linear features on the interpolated grid. When applied to both synthetic data and field data, the new method showed improvement over standard bidirectional gridding, minimum curvature, and kriging methods for interpolating thin, linear features at acute angles to the flight lines .Following this, a machine-learning interpolation approach was developed for aeromagnetic data using support vector machines and random forests. By using multiple standard interpolation methods as input to the machine-learning models, a filter-like approach was developed. These models could produce aeromagnetic maps that were overall more accurate than any single interpolation method, but not as effective as the Taylor derivative expansion method on lineament features. Finally, convolution neural networks were applied to estimate the source parameters characterizing lineament anomalies. A synthetic aeromagnetic data modeler was used to vary relevant physical parameters, and a representative dataset of approximately 1.4 million images was developed. These were then used for training convolution neural networks to estimate the strike and depth of sources. Applying the trained networks to a real-world dataset that was interpolated by the Taylor derivative expansion method, they located a dyke and estimated a depth consistent with a previous borehole investigation.