Wastewater treatment is an important part to keep water clean in human production and life, and sludge sedimentation ratio is the key index to measure the effectiveness of wastewater treatment. The existing measurement of sludge sedimentation ratio needs to wait for the completion of impurities settling, which is time-consuming and inefficient to detect. In order to achieve real-time, non-contact and low-consumption measurement of sludge sedimentation ratio, the relationship between optical polarization information and sedimentation ratio of homogeneous sludge water samples was studied. The polarization feature parameters were extracted by measuring the polarization image of the mixed solution before sludge sedimentation, the back propagation (BP) neural network was used to train and model the polarization parameters and sedimentation ratio, and the BP neural network prediction model with input layer of I, Q, U, P, θ and hidden layer of 13 nodes, output layer of sludge sedimentation ratio and network topology of "5-13-1" was established. The Levenberg-Marquardt (L-M) algorithm was used for model training, and tansig-purelin was used for the network transfer function. The results show that the average relative error of the prediction model is 4.361%, the average absolute error is 0.008 21, the mean square error was 0.000 14, and the root mean square error was 0.012 13, all of which are within the error tolerance and can be used for rapid prediction of sludge sedimentation ratio.