The importance of realistic noise characteristics in neural network inversion of airborne electromagnetic data

Summary Deep learning methods such as neural networks offer a rapid alternative to geophysical inversion and are increasingly applied to airborne electromagnetic (AEM) data for a range of earth science problems. Such networks rely on both the network architecture as well as suitable training datasets to produce realistic earth resistivity models. Here we develop a Convolutional Neural Network (CNN) to invert AEM data and pay particular attention to the characteristics of the training dataset. As well as covering a realistic range of subsurface property distributions, the training data must emulate the noise and data gap characteristics of the target survey data. We find neural networks trained on noise free data, while performing very well on validation data, perform poorly on real world data, often inducing geologically unrealistic artifacts in the resistivity model. Conversely, training data conditioned with noise and data gap characteristics appropriate for the survey of interest, fit the validation data less well, but produce fewer artifacts when applied to survey data. Choice of loss function metric and data normalisation method is important and may vary between surveys. Despite strong advances in neural network design, developing training data with broad enough characteristics to be generalised across many surveys and data acquisition platforms remains an important challenge.