The invention discloses a rice spike segmentation method of a big paddy field based on deep learning and super-pixel segmentation. According to the invention, by use of simple linear iteration clustering method in the super-pixel segmentation technology, adjacent pixels with similar characteristics form image blocks, i.e., super-pixels; based on automatic annotation and selection of large-scale training samples, the type of the super-pixels is discriminated through a convolution neural network in the deep learning technology so that initial segmentation of rice spikes is achieved; and based on the entropy rate-based super-pixel segmentation method, initial segmentation results are optimized. Thus, effects caused by great difference in colors, shapes, sizes, poses and textures of different varieties of rice spikes in different growth periods, serious irregular edges of the rice spikes, color aliasing of spike leaves, and uneven and changeable light, shielding and wind blowing in the field can be overcome; precise segmentation of different varieties of rice spikes in different growth periods is achieved; and the method is applicable to segmentation of rice spikes in the pot plant environment. Compared with the prior art, the method is advantaged by high precision and applicability.