Hyperspectral unmixing (HSU) is an important technique of remote sensing, which estimates the fractional abundances and the mixing matrix of endmembers in each mixed pixel from the hyperspectral image. Over the last years, the linear spectral unmixing problem has been approached as the sparse regression by different algorithms. Nevertheless, the huge solution space for individual pixels makes it difficult to search for the optimal solution in some HSU algorithms. Besides, the mixing relationship between adjacent pixels is not fully utilized as well. To better handle the huge solution space problem and explore the adjacent relationship, this article presents an improved collaborative non-negative matrix factorization and total variation algorithm (ICoNMF-TV) for HSU. The main contributions of this article are threefold: 1) a new framework named ICoNMF-TV based on non-negative matrix factorization method and TV regularization is developed to improve the performance of HSU algorithms; 2) unmixing efficiency is apparently improved; and 3) the robustness is enhanced. Experiment results on simulation dataset and real dataset demonstrate the proposed algorithm outperforms most of the similar sparse regression algorithms.