Deep learning (DL) technology using neural networks that reproduce the behavior of human neurons has gained significant traction in various fields. However, the large amount of data required in DL is difficult to obtain and prepare. To address this shortage, data augmentation is used to pseudo-increase the amount of data through targeted transformations. Generally, data augmentation is evaluated based on the accuracy of the model created using data augmentation in relation to the data for validation; the impact on each layer of the model is largely ignored. In this study, the impact of data augmentation on each layer of the model is considered. For this purpose, fusion transfer learning is performed using multiple augmentation methods from the same source.