Smoke detection in storage yard has great signification for fire early warning and protecting the safety of personnel and property. To solve the problem of insufficient features extraction, high false positive rates and poor robustness of traditional smoke detection methods, a new method of smoke detection in storage yard based on the parallel deep residual network is proposed. This method builds the parallel deep residual network with R, G, B components of the smoke RGB image and H, S, I components of the HSI transform image to adaptively extract the features. Meanwhile, the discriminant ability for the target like-smoke of the model is enhanced by the strategy including expanding the sample scale and reinforcement learning of the negative samples. The experimental results show that the proposed algorithm can effectively reduce the false positive rate caused by target like smoke and improve the detection rate and robustness of network.