我们建立了一个动态的无监督特征学习模型，以充分挖掘高光谱图像的内在特征。首先，我们将图像数据投射到一个低维的子空间上。然后利用低秩矩阵将主要特征提取出来，并同时过滤掉稀疏的噪声。为了保留数据的光谱信息和空间信息，我们在特征空间中嵌入了两个动态的图以进行约束。最后我们考虑了基于移动平均算法的超像素分割所带来的影响。模型的特征学习效果是通过最小化目标函数来达到的。优化过程中，我们采用带自适应罚函数的线性交替方向法来提高收敛速度。实验在一个公开的数据集上展开。我们利用学习到的特征和少量的标签构造了一个半监督的分类器，然后将分类精度和现有的方法比较。实验结果有力地证明了本模型的优越性。

Aiming to exploit intrinsic feature of hyperspectral images (HSI) under an unsupervised setting, we construct an adaptive feature learning Model (referred as AFLM). First, weproject the data into a lower-dimension subspace. Then the low-rank representationis adopted to extract most representative feature and remove the noise. In order to preserve the spectral and spatial structure of the reconstructed data, webuild two adaptive graphsthatare embedded into the feature space. Finally we take super-pixel effect into account based on mean shift algorithms.To optimize the learning effectiveness, the objective function is minimized throughthe linearized alternating direction method with adaptivepenalty.The experiment is implemented on one public data set.We utilizeour so-obtained features and minor labels to construct a semi-supervisedclassifier, and compare the classification accuracy with the existing feature learning methods. The superiority of our proposed method is strongly verified according to the experiment result.