This letter investigates the problem of anti-jamming communications in dynamic and unknown environment through on-line learning. Different from existing studies which need to know (estimate) the jamming patterns and parameters, we use the spectrum waterfall, i.e., the raw spectrum environment, directly. Firstly, to cope with the challenge of infinite state of raw spectrum information, a deep anti-jamming Q-network is constructed. Then, a deep anti-jamming reinforcement learning algorithm is proposed to obtain the optimal anti-jamming strategies. Finally, simulation results validate the the proposed approach. The proposed approach is relying only on the local observed information and does not need to estimate the jamming patterns and parameters, which implies that it can be widely used various anti-jamming scenarios.