The invention provides an inversion method of longitudinal discrete coefficients based on adjoint assimilation which is characterized by comprising following steps: giving upstream and downstream pollutant concentration observed values C1obj(t), C2obj(t), river water quality model related parameters flow velocity u and time t; assuming an initial value E0 of a longitudinal discrete coefficient andsetting target function accuracy requirement tol; step 2: substituting the E0 into a river water quality model to obtain a difference J between a calculated value and the observed value as the targetfunction; step 3. performing inversion integration on an adjoint equation to calculate the gradient of the target function about the longitudinal discrete coefficient; according to the gradient, obtaining the search direction [delta] J and step size [alpha] of the control variable by adopting the steepest descent method, and optimizing the longitudinal discrete coefficient variable; step 4. substituting the optimized longitudinal discrete coefficient into the river water quality model equation to calculate the concentration distribution C2(t) of pollutants at location x2 at the downstream observation section; calculating the target function J again to check whether it meets the accuracy requirements; otherwise, returning to step 2., and continuing to cycle until the accuracy requirementsare met.