Jun Yue, Dingshun Zhu, Leyuan Fang, Pedram Ghamisi, and Yaowei Wang
A hyperspectral image (HSI) contains two-dimensional spatial information and spectral reflectance information; each pixel represents a vector shaping a spectral curve. The goal of HSI classification is to determine an object category (e.g., river, road, forest) for each pixel of the image. HSI classification plays an important role in many applications in geology, ecology, forestry, agriculture, and other fields. Due to high spectral dimension of HSI, classification remains a challenging task. Deep learning methods have a potential to significantly enhance accuracy of HSI classification. However, they rely on a large number of labeled samples for training, which is limited in HSI.
In this paper, we introduce an HSI classification method based on adaptive spatial pyramid constraint. By using spatial-spectral correlation between labelled and unlabeled samples, the proposed method can improve the generalization ability of the model and reduce the dependence on labelled samples. The proposed method includes several important steps. First, an HSI complexity evaluation method based on edge detection is introduced to assess the homogeneity of the objects in the HSI. Second, an HSI pyramid segmentation method based on spatial pyramid is introduced to generate multi-scale sub-regions, where HSI complexity is used to adaptively determine the scale of the segmentation. Third, a spatial supervised constraint is introduced to generate the loss function of labeled sub-regions. Fourth, a spatial unsupervised constraint is introduced to generate the loss function of unlabeled sub-regions. The effectiveness of the proposed method is evaluated on three hyperspectral benchmark datasets. The proposed method shows better performance compared to the existing state-of-the-art methods.
IEEE Transactions on Geoscience and Remote Sensing, 2021-07-20.