<p>Salt body segmentation in seismic images is vital for accurate hydrocarbon exploration. Deep learning models, specifically the U-Net, have shown a significant improvement for this task. However, existing models often lack generalizability when applied to diverse datasets, limiting their practical effectiveness. This limitation may arise due to factors such as data heterogeneity, variations in data preprocessing methods, differences in seismic acquisition parameters and geological settings, or even the choice of machine learning models, which may overfit to specific data characteristics and fail to adapt to unseen scenarios. This study addresses this gap by proposing a robust and generalizable segmentation framework trained on a combination of three diverse and widely used datasets – TGS, SEAM, and F3. The core of the methodology leverages an encoder–decoder architecture, where a pre-trained EfficientNet-B7 model, originally trained on ImageNet, serves as the encoder to significantly enhance feature extraction from complex seismic data. This integration of transfer learning with domain-specific training aims to improve model robustness and adaptability across varied geological settings. The model is evaluated independently on each dataset’s test set, showing significant improvements in Intersection over Union (IoU), accuracy, and F1 score. Specifically, it achieves an IoU of 0.955 and an F1 score of 0.969 on the validation data, demonstrating strong segmentation capability and consistent generalization across varied geological conditions. These findings highlight the potential of multi-dataset training combined with transfer learning to build generalized salt body segmentation models suitable for real-world seismic applications.</p>