高光谱图像相对含水量反演引导的精细分类

The application of deep learning methods in the hyperspectral research field has been developed continuously, and hyperspectral image classification models based on deep learning methods achieve high classification accuracy compared with others. Current classification models mostly use hyperspectral graphic-spectral features but lack diagnostic features and prior information, which makes it difficult to extract spectral-spatial features collaboratively and implement refined classification further. The results are not refined for intraclass classification. To solve the abovementioned issues, we proposed a Symbiotic Neural Network (SNN) for hyperspectral refined classification, which regards a multilabeled data set as input. SNN can fuse spectrum diagnostic features with graphic-spectral features to retrieve relative water content and implement refined classification simultaneously. First, we proposed a new spectral index, Red Edge Sloped (RES), to characterize the relative water content and used RES to label the relative water content on all the objects in data sets via an adaptive grading algorithm. We then built a multilabeled data set with original object labels. Second, we established the SNN architecture and dimensionality-varied feature extraction module to extract fusion information of space, spectra, and relative water content in hyperspectral data, which was able to improve the cooperative expression capability for features and the discrimination capability for the water content of various ground objects. It was also able to reduce the structural complexity and amount of computation of the deep model. Moreover, the dimensionality-varied module was able to extract more accurate and abstract features at a high level and improved the extensibility to build a deeper network. After the aforementioned progress, we implemented hyperspectral image refined classification with the guide of relative water content retrieval by using the SNN and completed interclass and intraclass classification simultaneously. On the basis of the relative water content retrieval in the classification, the interclass and intraclass distances can be further expanded over traditional classification methods. Experiments were conducted using one hyperspectral data set collected in the laboratory and four open hyperspectral data sets, namely, Lopex, Indian Pines, Pavia University, and Salinas, to validate the effectiveness of RES and the superiority of the SNN model. The experimental results demonstrated that the RES index proposed in this paper was able to represent the relative water content of objects in data effectively. The classification accuracy and discrimination capability for the water content of SNN were improved evidently with the guide of relative water content retrieval. In addition, we compared SNN with other state-of-the-art methods, and SNN obtained higher classification accuracy. Therefore, the SNN model proposed in this paper, which implements refined classification with the guide of relative water content retrieval, can enhance the feature extraction capability of the model for hyperspectral image classification and improve the classification performance effectively.