A Comparative Study on Calibration of Deep Segmentation Models Over Distribution Shifts in Seismic Interpretation

In recent years, deep learning approaches have been proposed to help geoscientists in interpreting tasks that are often labor intensive and subjective. When deployed in a real situation, those models can present distinct results from the ones seen in their evaluation process. In this work, we compare the segmentation and calibration performance over-distribution shifts of popular segmentation models on the problem of lithology segmentation. Our findings show that in a situation of data abundance during training, fine-tuned visual transformers (ViTs) models perform best over their convolutional-based competitors. We also noticed that when the training data are greatly limited, a common situation in the geology interpretation field, the fine-tuned models show equally or worse calibration and segmentation performances than other evaluated models.