The applications which aim to classify seafood on the ships or markets are essential. However, when fishers have caught a large amount of seafood on board, they have to manually classify each seafood type for a long time, affecting seafood quality after catching. Moreover, When classifying seafood by manual method, fishers who have prolonged exposure to seafood are prone to skin diseases such as skin infections, fungal infections, cracked fingers. In addition, they also often suffer from spinal diseases due to having to stand in one position and move continuously during the classification process. Therefore, this study focuses on developing a method to effectively detect and classify some popular species of saltwater seafood using MobileNetV2 based on selected hyper-parameters via experiments on 9000 images of nine seafood types. Experimental results have reached 0.9956 in accuracy metric on the test set with the selected hyper-parameters combining considered data augmentation techniques. Also, the work evaluates the effects of eliminating each data augmentation technique to compare their influence on image classification tasks improvement. We can consider removing augmentation techniques that have revealed less influence in the image classification performance of seafood types.