Aprendizado autossupervisionado aplicado a imagens sísmicas: análise de desempenho de treinamento in/off domain

Abstract

In the oil industry, seismic data is of fundamental importance as it is used, for example, in reservoir discovery, oil production, and well abandonment stages. Due to this importance, large volumes of seismic data are generated, but it is impractical to manually label these data completely. Considering the quantity of data produced and limited annotation, self-supervised learning emerges as an option to pretrain in-domain models for subsequent application in downstream tasks. To verify the advantage of conducting self-supervised learning in the seismic image domain, a qualitative and quantitative performance comparison of in/off domain models was conducted. Using proprietary data from Petrobras, pretraining was performed using the DINO, DINOv2, and MAE methodologies. The SFM model, trained on public seismic images, was also utilized. DeiT-III, DINOv2, and random initialization were selected as off-domain models. In the qualitative evaluation, the reduced-dimensional representation of the models showed that the in-domain models were better able to separate seismic images from different sources, with the exception of the SFM model. In the quantitative evaluation, various tasks and training data subsets were tested. For seismic facies segmentation, the in-domain model had better performance in transfer learning for all tested data subsets. In fine-tuning, the best results were mostly from seismic models, although the best overall result was off-domain. In the other tasks, the majority of the results were better for in-domain models, except for geobody segmentation. In the conducted experiments, in-domain self-supervised learning proved to be beneficial. However, due to the high computational power required for pretraining, this procedure is restricted to large research groups or corporations.