Data acquired from multi-channel sensors is a highly valuable asset tointerpret the environment for a variety of remote sensing applications.However, low spatial resolution is a critical limitation for the sensors andthe constituent materials of a scene can be mixed in different fractions due totheir spatial interactions. Spectral unmixing is a technique that allows us toobtain the material spectral signatures with their fractions from data. In thispaper, we propose a novel hyperspectral unmixing scheme, called EndNet, that isbased on a two-staged autoencoder network. This well-known structure iscompletely enhanced and restructured by introducing additional layers and aprojection metric (i.e spectral angle distance (SAD) instead of inner product)to achieve an optimum solution. Moreover, we present a novel loss function thatis composed of Kullback-Leibler divergence term with SAD similarity andadditional penalty terms to improve the sparsity of the estimates. Thesemodifications enable us to set the common properties of endmembers such asnonlinearity and sparsity for autoencoder networks. Lastly, due to thestochastic-gradient based approach, the method is scalable for large-scale dataand it can be accelerated on Graphical Processing Units (GPUs). To demonstratethe superiority of our method, we conduct extensive experiments on severalwellknown datasets. The obtained results confirm that our method considerablyimproves the performance compared to the state-of-the-art techniques inliterature.
展开▼
