Scene recognition is the key to visual navigation for the agricultural mobile robot in unknown environment. This paper used what-where dual channel theory to build the models of scene perception, scene representation and scene recognition, and proposed a bionic method of scene recognition on the basis of probabilistic framework. This method first computed the bottomup saliency map of scene based on the contrast prior and the center prior, which can be further optimized with the global energy function. Then shifted the visual focus of saliency map to obtain the saccade sequence as the “where information”, and analyzed the content of the visual focus to obtain the “what information” with the experts network comprised of single layer perceptron. Lastly, according to the action recognition regularity of human, built the discrete and observable Markov model using the “what information” and the “where information”. The parameters of the model can be determined by training the frame images from the camera on the mobile robot and can be viewed as the prior knowledge about different scenes, which can be recognized by maximizing the likelihood probability of the Markov recognition model. The whole recognition process is similar to human`s. Experimental results show that this method has good performance for indoor scenes and the recognition accuracy averaged out at 87.3%.