Aiming at the inversion of suspended particulate matter (SPM) concentration using hyperspectral data, this paper proposes a supervised band selection method based on pre-trained neural networks (PNNs), and employs the random forest and neural network to establish an inversion model of SPM concentration. The PNN method needs to perform multiple repeated experiments to obtain sufficient and low-noise expression of band importance. In each experiment, an appropriate number of bands is selected as the features of input data of neural networks. Then, we train a neural network and obtain weights of the first layer in the last training epoch. Finally, we use the L1 norm, L2 norm, and ReLU (Rectified Linear Unit) function of the weights to represent the importance of the bands. The experiment results show that the PNN method using L1 norm and L2 norm can obtain a more informative band set, and perform better when used for SPM concentration inversion.