A novel multi-domain fusion network based on sample intercorrelation (SIMFNet) is proposed. The first mulitvariate aquatic human activity recognition(AHAR) dataset is collected. We evaluate our proposed model with the-state-of-the-art models on our AHAR dataset and the experimental results verify the effectiveness of our models.