Insights on domain adaptation for fault identification

Mariana Lume, Marcelo Guarido, David J. Emery, Kristopher A. Innanen

Conventional deep learning has proved to be successful for fault identification tasks if the dataset selected for training and testing a model belong to the same domain. When it does not occur, the classifier usually degrades its performance. This study focuses on explaining the successes and failures of applying a model, previously trained with synthetic seismic images, on two different datasets of synthetic and real seismic images. When considering synthetic datasets, the classifier was very accurate; however, in the second scenario, a significant number of misinterpreted faults appeared. These outcomes are a direct consequence of the similitudes or discrepancies between both datasets used, hence Domain Adaptation techniques are usually applied to overcome the encountered challenges.