Due to the existing limitations of hyperspectral image gradient descen t unmixing algorithm based on bilinear mixing model,a novel nonlinear unmixing algorithm based on the neural network (NN) and differential search algorithm (DSA) is proposed.Based on the p-order polynomial model,the NN can be used to efficiently estimate the nonlinearity order of real hyperspectral data pixels.We construct the objective function in order to turn the nonlinear unmixing problem into an optimization problem for hyperspectral da ta unmixing.By introducing the DSA,the proposed algorithm transforms the parameters to be solved in the hyperspectral unmixing problem into the position parameters in the search process of DSA.And the mapping mechan ism is introduced to meet the requirements of the abundance nonnegative constraint and abundance sum-to-one constraint.The expe rimental results show that the proposed framework has promising performances com pared with gradient nonlinear unmixing algorithms.When using 2000samples to train the NN and using the proposed algorithm to unmix the real hyperspectral data,the reconstruction error (RE) of unmixing index can reach about 1.15×10－2 which is better compared with t hose of other existing hyperspectral gradient descent nonlinear un mixing algorithms.