| Natural gas segmentation in seismic data has been widely addressed using machine learning models. However, these models predominantly explore a single form of representation, either by treating seismic sections as images or by performing temporal analysis of seismic traces. Even so, such approaches utilize the extracted features in isolation, without considering the possible complementary aspects between these representations. Within this context, multimodal models have enabled the integration of multiple complementary representations of the same data. This paper proposes a multimodal deep learning architecture for natural gas segmentation in 2D seismic data that combines convolutional encoders for spatial interpretation and temporal encoders for the analysis of seismic traces. The extracted representations are combined at the bottleneck level and subsequently employed in the natural gas segmentation process. The experiments were conducted using a proprietary seismic dataset containing gas-horizon annotations provided by experts. For each conducted experiment, we performed a qualitative and quantitative evaluation, considering different regions. Our multimodal architecture demonstrated improvements in metrics such as F1-Score and Intersection over Union (IoU). The results of the F1-Score improved from 52.13% to 82.07% and the IoU from 36.20% to 70.28% in comparison to conventional methods. These results demonstrate that integrating spatial and temporal representations allowed substantially more accurate gas segmentation. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.