In response to the inadequacy of constructing and encapsulating long-distance feature relationships in convolutional neural networks currently utilized for copper ore grade classification, the author proposes a method that combines a SwinTransformer-EfficientNet ensemble model. This methodology fully exploits the SwinTransformer V2-t architecture's capability in summarizing long-range feature associations, as well as the EfficientNet V2-s's strength in discerning subtle local characteristics. By incorporating a linear layer to amalgamate the outputs of both models and adaptively tuning the weights of this linear layer according to the individual model's output, the mapping relationship is optimized, leading to a substantial enhancement in classification performance. Empirical validation indicates that this fused model attains an accuracy of 92.891%, precision of 93.095%, and recall of 92.654% in classification tasks. Relative to the standalone, non-integrated models, the integrated composite model exhibits an increase of 1.30% in accuracy, 1.9% and 2.186% in precision, and 0.474% and 0.237% in recall, respectively, manifesting considerable advancements.