Salmon is subject to significant fraudulent activities due to its high economic value, making it crucial to accurately determine its geographical origin for trade, consumers, and food safety. This study aims to explore the spectral and textural features in Hyperspectral Imaging (HSI) data of salmons and investigate the potential of deep learning methods and information fusion strategy in salmon origin traceability. Additionally, an improved dung beetle optimization algorithm optimized Bidirectional Gated Recurrent Unit (BiGRU) deep learning model (CNN-BiGRU) was constructed to handle high-dimensional data accurately in discerning the origin of salmon. On the other hand, an ensemble learning model was built by stacking three base models (Random Forest, LightGBM, and Gradient Boosting Decision Tree), and a comprehensive evaluation of the predictive performance between the ensemble learning model and the base models was conducted. The results indicate that the classification performance of the ensemble learning model consistently outperforms the base models. However, compared with traditional machine learning methods, deep learning models demonstrate superior performance in handling high-dimensional data, particularly exhibiting robustness in handling fused data. Notably, the MSADBO–CNN–BiGRU model stands out among all algorithms, showing an improvement of 0.8%–7% in accuracy on the test set compared to the pre-optimized CNN model, with the best model achieving a test set accuracy of 99.5%. Overall, this study demonstrates that combining deep learning with information fusion strategy can rapidly and accurately identify the geographical origin of salmon, providing methods and insights for future traceability of agricultural products.