Seismic reflection is the most widely used geophysical method in the oil and gas industry to study subsurface layers. Based on the reflection patterns of seismic waves, geoscientists can infer the structure and composition of geo logical layers beneath the surface, identifying potential oil and gas reservoirs. However, interpreting this information is challenging due to the inherent ambi guity of the data, meaning distinct events can have similar seismic responses. In order to guide and assist this process, experts often employ a large set of seismic attributes. However, the use of more information in a machine learning context does not guarantee improvement in results, and in some cases, many of the features may not be utilized by the model. Therefore, selecting which features are most relevant becomes essential. However, manual selection among hundreds of attributes can pose an exponential challenge. This work proposes an approach that incorporates the use of a customized attention layer to han dle multiple features in conjunction with a Long Short-Term Memory (LSTM) model. This approach aims to automatically weigh the seismic attributes, pre selected by domain experts, to evaluate which ones are most important for the model in the natural gas detection process. To evaluate the methodology, 2D and 3D onshore seismic surveys were employed, and the K-fold technique was applied. For quantitative results, the F1-score metric was evaluated, achieving an improvement of up to 13,94 percent.