<p>Accurate salt body segmentation in seismic imaging is essential for hydrocarbon exploration, as it directly impacts subsurface characterization, drilling safety, and reservoir modeling. However, existing deep learning approaches often struggle with generalization across diverse geological settings, limiting their reliability in real-world applications. A key challenge lies in optimizing segmentation models to effectively capture complex salt structures while maintaining robustness across varying datasets. To address this, we propose a weighted-ensemble transfer learning framework, leveraging EfficientNetB7 as a high-performance encoder within U-Net and LinkNet architectures. Unlike conventional single-model approaches, our method integrates multiple encoder-decoder networks through weighted averaging, enhancing segmentation accuracy and stability. The training scheme incorporates stochastic weight averaging, adaptive learning rate scheduling, and momentum-based stochastic gradient descent (SGD) to further optimize performance. We evaluate our framework on three benchmark seismic datasets, demonstrating that the proposed ensemble-based approach outperforms traditional deep learning models in salt body segmentation. By systematically assessing encoder-decoder pairings and transfer learning techniques, our study provides key insights into optimizing seismic segmentation models. These findings emphasize the potential of ensemble-driven transfer learning in capturing intricate geological structures and improving model robustness. To support reproducibility and further innovation, we have publicly released our full implementation on GitHub. This work marks a significant step toward next-generation AI-powered seismic interpretation, with broad implications for geophysical exploration and reservoir modeling.</p>