Vessel detection and type recognition is crucial in any maritime surveillance application. This component aims at preventing or investigating unlawful actions present at sea. Modern very high resolution (VHR) optical satellite sensors are able to capture images with spatial resolution up to 0.3m per pixel, which is sufficient to distinguish ship features such as bridge position, cranes, landing pads and many others and thus possible to differentiate ship types. This paper presents a new method for automatic vessel detection and type recognition based on fusion of deep convolutional neural network architectures (CNN), which has potential for near-real time (NRT) applications.