Abstract: As a comprehensive indicator of flow resistance, Manning’s roughness coefficient significantly affects the accuracy of one-dimensional unsteady flow simulations. The variation of the roughness with the discharge or water level were seldom investigated in previous studies on roughness inversion. To address this issue, a roughness inversion method for river unsteady flow simulations based on long short-term memory neural network is proposed to realize the direct inversion of roughness by data-driven method, in which the roughness is regarded as a continuous piecewise linear function of discharge. Moreover, in view of the characteristics of the long natural rivers with many cross sections and large discharge variation range, a successive-approximation based stepwise inversion strategy is developed to reduce the dimension of inversion solution. The proposed inversion method is evaluated through a case study on the reaches of the Xiangjiaba Reservoir, China. The results show that the water stage hydrographs computed from the roughness inversion values under different discharge strata at the observation stations are in good agreement with the measured data, and the calculation accuracy of the proposed method is obviously higher than the method without considering the variation of the roughness with the discharge. The results verify the effectiveness of the propose method and provide a novel approach for the inverse estimation of roughness in long river flows.