The invention discloses a manifold learning-based method for monitoring water quality by remote sensing, and particularly relates to a manifold learning-based method for monitoring water quality by multispectral remote sensing, which comprises the following steps of: training a sample point by a manifold learning method to obtain a projection matrix of water quality remote sensing data in low-dimensional embedded space; building a nonlinear model by using a support vector regression algorithm according to a feature matrix of the low-dimensional embedded space of the sample point and corresponding water quality parameter site monitoring data; and inverting an integral water body by utilizing the nonlinear model to obtain a concentration value of a water quality parameter of the integral water quality, and providing different colors for the water body according to different intervals in which the water body is positioned to display the concentration condition of the water quality parameter of the integral water body visually, so that the water quality evaluation and monitoring of the water body are realized. In the method, the essential law hidden in multispectral remote sensing image data is disclosed effectively, and the nonlinear problem of the water quality evaluation is solved.