The invention discloses a frequency spectrum sensing method based on dictionary learning in an extremely low signal to noise ratio environment, and relates to a frequency spectrum sensing method in cognitive radio. The objective of the invention is to solve the problem of low spectrum sensing accuracy in an extremely low signal-to-noise ratio environment in the prior art. The method comprises thefollowing steps: 1, establishing a spectrum sensing binary hypothesis model; 2, forming a training set of dictionary learning; 3, training a dictionary; 4, calculating the inner product of each columnof the spectrum sensing signal and the trained dictionary, and finding out the position of the maximum value in the inner product; 5, updating the index set and the atom set, and obtaining a maximumcomponent by using a least square method; 6, calculating the obtained square of the maximum component to obtain energy corresponding to the maximum component; 7, calculating a perception threshold according to a false alarm probability formula; and 8, comparing the obtained energy corresponding to the maximum component with a threshold to judge whether the signal exists, and if the energy is greater than the threshold, determining that the signal exists, otherwise, determining that the signal does not exist. The method is applied to the field of spectrum sensing in cognitive radio.