Improved Hydrocarbon Detection with Self-Supervised Long Short-Term Memory Networks

In the quest for efficient hydrocarbon exploration, seismic reflection plays a pivotal role. However, the inherent ambiguity in geological features within seismic poses significant interpretation challenges, often necessitating costly analysis. We introduce a novel self-supervised approach utilizing Self Supervised Long Short-Term Memory (LSTM) networks. Our method significantly reduces the reliance on extensive annotated data, a common bottleneck in deep learning applications. We demonstrate that our sequence reconstruction task, employed as a pre-training method, leads to a marked improvement in model performance. Notably, our results exhibit a 5%-10% increase in mean Intersection over Union (mIoU) and an 8%-15% enhancement in mean precision compared to existing models. Our method demonstrates the potential of self-supervised techniques for advancing seismic data interpretation, leading to cost-effective and accurate geological mapping.