In order to identify the parameters of the river water quality model, we adopted the Crank-Nicolson finite difference scheme to discrete the control equations first, and then appended several additional terminal observations, and designed an effective gradient regularization algorithm to reconstruct the model parameters. Numerical results show that for both the constant coefficient model of straight channels and the natural coupled or non-coupled variable coefficient model, the proposed method can quickly and effectively identify the model parameters. In addition, the influences of the initial value and measuring error are insignificant, which further demonstrates that the gradient regularization algorithm is effective to inverse the parameters of the river water quality model.