Application of deep learning in geophysics: localization of seismic wave sources by means of transfer learning

This paper discusses and analyses the application of time reversal imaging (TRI) and transfer learning methods for the spatial localization of seismic sources. The study makes use of the ResNet50 model, which is pretrained on ImageNet images. The effectiveness of the proposed localization method was evaluated using accuracy, recall, and F<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</inf> score, calculated for the localization of multiple seismic sources in the Marmousi composite velocity model. The results show that a combined TRI and transfer learning approach ensures a high degree of precision in the classification of seismic source areas. Notably, six different input combinations achieved F<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</inf> scores above 75%. The best performance was obtained using information from combination of three wavefield components, yielding an F<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</inf> score of 80%, a recall of 75%, and a precision of 86%. The study highlights the potential for further work by adjusting a residual convolutional neural network based on a set of images of wavefield component values characteristic of the source regions.