Hyperpectral unmixing is a difficult problem in academia. Sparse hyperspectral unmixing uses priori spectral library, aiming at finding several pure spectral signatures to express hyperspectral images and computing corresponding abundance fractions. This is NP-hard to solve. Convex relaxation for L0 norm as L1 norm is a common approach to solve the sparse unmixing problem, but only approximation results can be achieved. A Pareto optimization based sparse unmixing algorithm was proposed(ParetoSU). ParetoSU firstly transformed sparse unmixing to a bi-objective optimization problem. One of the two objectives was the modelling error and the other one was the sparsity of endmembers. ParetoSU can solve the sparse unmixing problem without any approximation of L0 norm. At last, synthetic data were used to test the performance of ParetoSU.