A number of classification systems for dolphin whistles are applied to study the relation between dolphin whistles and their behaviors. Traditional approaches require prior knowledge, numerous pre-processing and manually extracting the features by conventional signal processing methods. In this study, deep convolutional neural networks are used to classify whistles and automatically learn the sound characteristics from training data with less pre-processing. The classification system is trained by using a database of 9 sorts of measured dolphin signals with nearly 4000 samples, and whistle contours in testing dataset are divided into six types, including constant frequency, upsweep, downsweep, concave, convex, and multiple. Finally, more than 90% for the classification accuracy rate is reached, and results show insensitivity to background noise. Therefore, the algorithm can be employed to study the potential relation between dolphin whistle signals and behaviors, and then facilitate future studies on dolphin habits.