A deep learning benchmark for first break detection from hardrock seismic reflection data

Deep learning techniques are used to tackle a variety of tasks related to seismic data processing and interpretation. Although many works have shown the benefits of deep learning, assessing the generalization capabilities of proposed methods for data acquired in different conditions and geologic environments remains challenging. This is especially true for applications in hardrock environments. The primary factors that impede the adoption of machine learning in geosciences include the lack of publicly available and labeled data sets and the use of inadequate evaluation methodologies. Because machine learning models are prone to overfit and underperform when the data used to train them are site specific, the applicability of these models on new survey data that could be considered "out -of -distribution" is rarely addressed. This is unfortunate, as evaluating predictive models in out -of -distribution settings can provide a good insight into their usefulness in real -world use cases. To tackle these issues, we develop a simple benchmarking methodology for first break picking to evaluate the transferability of deep learning models that are trained across different environments and acquisition conditions. For this, we consider a reflection seismic survey data set acquired at five distinct hardrock mining sites combined with annotations for first break picking. We train and evaluate a baseline deep learning solution based on a U -Net for future comparisons and discuss potential improvements to this approach.