Hyperspectral　unmixing　is　an　important　technique　for　remote　sensing　image　analysis.　Among　various　unmixing　techniques,　nonnegative　matrix　factorization　(NMF)　shows　unique　advantage　in　providing　a　unified　solution　with　well　physical　interpretation.　In　order　to　explore　the　geometric　information　of　the　hyperspectral　data,　graph　regularization　is　often　used　to　improve　the　NMF　unmixing　performance.　It　groups　neighboring　pixels,　uses　groups　as　graph　vertices,　and　then　assigns　weights　to　connected　vertices.　The　construction　of　neighborhood　and　the　weights　are　normally　determined　by　k-nearest　neighbors　or　heat　kernel　in　a　deterministic　process,　which　do　not　fully　reveal　the　structural　relationships　among　data.　In　this　article,　we　introduce　an　adaptive　graph　to　regularize　a　multilayer　NMF　(AGMLNMF)　model　for　hyperspectral　unmixing.　In　AGMLNMF,　a　graph　is　constructed　based　on　the　probabilities　between　neighbors.　This　enables　the　optimal　neighborhood　be　automatically　determined.　Moreover,　the　weights　of　the　graph　are　assigned　based　on　the　relationships　among　neighbors,　which　reflects　the　intrinsic　structure　of　the　complex　data.　Experiments　on　both　synthetic　and　real　datasets　show　that　this　method　has　outperformed　several　state-of-the-art　unmixing　approaches.