The invention relates to a fog-degraded image-based PM2.5 estimation method. The method comprises the steps of extracting the transmittance by using a convolutional neural network according to an existing fog-degraded image atmospheric scattering model; estimating an image depth by using a multi-scale depth network; solving an atmospheric scattering coefficient beta, which is directly related to PM2.5 and air quality by using the transmittance and the depth information; and finally determining the relationship between the beta and the PM2.5 by using a related algorithm of data mining. Through the steps, the estimation of the PM2.5 of the scene through a single fog-degraded image is achieved.