When using the existing methods of depth learning to study the vegetation extraction, there are some problems that the adjacent objects are in the same window, and some useless crushing plots and the salt and pepper phenomenon appear. We propose a method by combining the optimal segmentation scale with the deep belief network to study the vegetation extraction, and comparison experiments are carried out with spectral-texture features and other information. Experimental results show that the overall accuracy of the proposed method is 91.92% and the Kappa coefficient is 0.8677, and the proposed method can effectively improve the classification accuracy compared with the existing deep learning methods. The classification results show that the proposed method can effectively reduce the salt and pepper phenomenon, and clear express the boundaries of objects.