The transfer learning strategy for 3D post-stack seismic data reconstruction using SR-CNN

The advancement of oil and gas exploration into deeper and more complex environments has created challenges in accurately analyzing underground structures using 3D post-stack seismic data. Enhancing the resolution of these seismic data is essential to overcome these challenges. Traditional methods often suffer from noise sensitivity, limited adaptability, and restricted resolution enhancement. Recent developments in deep learning-based super-resolution technologies have shown promise in seismic data reconstruction. However, training high-quality models typically requires substantial high-resolution sample data. Transfer learning addresses this by leveraging existing knowledge to improve model performance with limited data. This paper proposes a novel transfer learning strategy for 3D post-stack seismic data reconstruction using Super-Resolution Convolutional Neural Networks (SR-CNN). We generate 200 high-resolution seismic datasets through simulation, downsample them, and build a pre-trained model using SR-CNN. The model is fine-tuned with 50 pairs of actual high-resolution and low-resolution seismic data to optimize accuracy and robustness. Experimental results demonstrate that the SR-CNN method significantly improves reconstruction quality, with higher Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) compared to traditional methods. The SR-CNN-based transfer learning strategy reduces training time and enhances robustness against noise, providing an effective solution for high-resolution seismic imaging.