On the Performance Evaluation of Deep Learning Models for Seismic Facies Segmentation

ABSTRACT The transformative impact of deep‐learning architectures on machine learning has been substantial. Recently, a wide range of studies have successfully applied these methods to seismic facies segmentation using well‐established public datasets, such as F3 and SEAM AI. However, many of these works lack detailed descriptions of their methodologies and implementation details, including dataset partitioning, hyperparameter settings and other critical aspects. The lack of reproducibility information makes fair comparison between studies quite difficult, as methodological details can heavily affect the results obtained. In this work, we discuss this problem and present a fair comparison between five state‐of‐the‐art models commonly used in the literature: DeepLab V3, DeepLab V3+, Segmenter, SegFormer and SETR. We found that the SETR model has promising performance on both the F3 and SEAM AI datasets and convolutional neural network models offer a higher performance to parameter count ratio compared to the transformer models.