高光谱图像分数域信息提取理论与方法进展

Hyperspectral sensing technology can acquire spectral, spatial, radiation and other information synchronously, which provides the multi-scale, multi-angle, and multi-dimensional features of land covers. However, there are significant challenges in hyperspectral information extraction, e.g., spectral uncertainty, insufficient utilization of spatial information, and incomplete representation of collaborative information, resulting in poor information extraction and scene interpretation. The applications of hyperspectral interpretation, e.g., earth observation, requires multi-domain information extraction theories and methods to breakthrough these problems. In this survey, we firstly present the existing methods for hyperspectral information extraction and their main problems, and then introduce the fractional information extraction theory and methods of hyperspectral image, which consists of spectral dimension, spatial-spectral dimension, and collaborative dimension. Then, the main theories and applications are introduced, including spectral information adjustment, spatial-spectral information enhancement, and information fusion and transferring of multisource remote sensing data. For spectral dimension, the spectral uncertainty phenomenon makes it difficult to distinguish small targets from complex backgrounds. Focusing on this problem, the fractional-domain spectral information extraction method can improve the performance of hyperspectral anomaly detection. For spatial dimension, the complex spatial distribution of hyperspectral scenes and the lack of labeled samples make the scene interpretation challenging. Focusing on this problem, the fractional-domain spatial-spectral feature extraction methods can effectively generate more discriminative training features and improve the diversity of training sets, which contribute to handling small sample size problems. For the collaborative dimension, the fractional-domain multi-source feature extraction and fusion method can realize the joint use of multi-source and multi-domain features, and achieve high-precision classification. Finally, this survey points out the challenges and development trends of fractional information extraction theory for hyperspectral images. To breakthrough the limitations of hyperspectral data, e.g., high-dimension and low-resolution, it is important to improve the data quality at data- and feature-level. To solve the problem of unavailable training samples, transfer learning techniques are in need to fully exploit the spectral, spatial and collaborative information of the massive unlabeled data in hyperspectral remote sensing images. Targeting the global-scale earth observation by remote sensing, focusing on the condition when some modalities are missing, researches on domain generation and cross-scene classification are in need.