The integration of artificial intelligence and distributed fiber optic acoustic sensing technology(DAS) has yielded remarkable results in recent years; however, some application scenarios face the challenge of acquiring an adequate amount of data for higher network accuracy. To address this issue, we propose a decoupling parallel convolutional neural network (DPCNN) that relies on multiple feature inputs to achieve higher accuracy while using smaller databases. Our model offers excellent recognition of five events, including background noise, footstep, digging, car passing, and climbing fence, with an accuracy rate of up to 94.9%. The DPCNN is a parallel and lightweight CNN that boasts a short training time of only 3.76s per epoch and a test time of 0.1175s, with superior network convergence. In comparison to a mature single-branch CNN based on mixed images of time-frequency and time-space, the DPCNN accuracy is 6.4% higher. Our model demonstrates excellent performance across various databases and can achieve recognition accuracy of up to 98.7% with larger databases. Finally, we show the broad range of applications available for DPCNN based on multiple feature inputs when using a mature single-branch replacement in each branch of a two-branch network.