The interpretation of seismic images is a complex and ambiguous task because it depends on the experience of the professional analyzing them. Seismic images have a fundamental use in the geosciences and are used in various tasks, including hydrocarbon exploration. Automated deep learning methods have recently been applied to seismic analysis tasks, such as seismic interpretation and event prediction. This work proposes a deep learning method for gas reservoir segmentation in 2D seismic data using a Vision-LSTM enhanced U-Net model. The method was applied to seismic data collected by Eneva, a Brazilian private energy company, from the Paleozoic Basin in Brazil. The data include seismic traces collected over time and two labels: confirmed gas reservoirs and the region of interest in each seismic image. The architecture was trained on clustered seismic datasets based on their features to create better generalization models, achieving an average F1-Score of 50% with variations between different clusters. The trained model showed excellent results in some clusters, with more than 80% in the evaluated metric. In contrast, other clusters had much lower results due to variations in seismic features, demonstrating that the proposed method has potential for automated gas reservoir identification.