1: \begin{abstract}

2: Hyperspectral unmixing, the process of estimating a common set of

3: spectral bases and their corresponding composite percentages at each

4: pixel, is an important task for hyperspectral analysis, visualization

5: and understanding. From an unsupervised learning perspective, this

6: problem is very challenging---both the spectral bases and their composite

7: percentages are unknown, making the solution space too large. To reduce

8: the solution space, many approaches have been proposed by exploiting

9: various priors. In practice, these priors would easily lead to some

10: unsuitable solution. This is because they are achieved by applying

11: an identical strength of constraints to all the factors, which does

12: not hold in practice. To overcome this limitation, we propose a novel

13: sparsity based method by learning a data-guided map to describe the

14: individual mixed level of each pixel. Through this data-guided map,

15: the $\ell_{p}\left(0<p<1\right)$ constraint is applied in an adaptive

16: manner. Such implementation not only meets the practical situation,

17: but also guides the spectral bases toward the pixels under highly

18: sparse constraint. What's more, an elegant optimization scheme as

19: well as its convergence proof have been provided in this paper. Extensive

20: experiments on several datasets also demonstrate that the data-guided

21: map is feasible, and high quality unmixing results could be obtained

22: by our method.\end{abstract}

23: