Supervised Seismic Image Enhancement from Structurally Conditioned Synthetic Data

Post-stack seismic images often present noise and structural deformations that hinder the interpretation of sub-surface features critical for hydrocarbon exploration. This work introduces a framework for seismic image enhancement using deep generative models, conditioned by structural geological information, within a supervised learning scheme. Five structural complexity categories-planar, inclined, folded, diapiric, and faulted-were identified from emerging Colombian basins to guide dataset compilation and model conditioning. A conditional generative network was trained to synthesize structurally realistic seismic images, reaching Fréchet Inception Distance (FID) scores up to 97.11, with visual evaluation confirming structural fidelity. To simulate realistic degradation, four types of noise (e.g., Gaussian, Poisson, speckle) were randomly applied to generate corrupted-clean image pairs. These pairs were then used to train a supervised denoising network, where clean images produced by the generator were degraded on-the-fly and immediately processed by the denoising model. This setup established a dynamic training cycle, allowing the network to learn enhancement strategies directly from continuously generated examples. When applied to real post-stack seismic data, the framework reduced noise and improved the continuity of key structural features. These results underscore the potential of deep generative modeling to improve seismic interpretability in structurally environments.