Abstract:
Abstract: Hyperspectral remote sensing technology are typically used to collect data with at least ten spectral bands of relatively narrow bandwidths. The hyperspectral remote sensing images often have extensive interband correlations, and high-dimensional data with uncertainty, which is significant characteristics of hyperspectral remote sensing imagery. As a result, the adjacent bands of the hyperspectral images may contain similar spectral information and spatial structures. Reducing the dimensionality of data therefore becomes one of the most important preprocessing step in hyperspectral image analysis. Since the individual bands of a hyperspectral image share high similarity, statistical data compression tools, such as principal component analysis (PCA), have been widely applied for dimensionality reduction of hyperspectral images. However, it is well known that PCA is sensitive to the presence of outliers and missing data. Meanwhile, the spectral variability from the identification of crop species makes the conventional information extraction methods unsuitable for the crop classification with hyperspectral remote sensing images. During the PCA, the geographical information including remotely sensed data cannot also be accurate, which means that the boundaries between different classes and phenomena are not clear, i.e., fuzzy. The fuzzy approach to statistical analysis rather than other mathematics methods can be quite suitable to process this kind of image data with the fuzziness and statistical property. Recently, a new learning paradigm—adjustment learning (AL) has been studied for the image retrieval as an emerging technology. In the AL scheme, data points can be identified as small group sets, i.e., "chunklets", with equivalence constraints that are known to originate from the same class (but the label is unknown). A non-iterative and efficient metric learning method, called discriminant component analysis (DCA), has been developed based on the AL. In this paper, the fuzzy sets theory and statistics were introduced into the AL-based technique, DCA, and then a hyperspectral imagery feature extraction method called fuzzy-statistics-based discriminant component analysis (FS-DCA) was proposed to effectively distinguish different types of crops via their hyperspectral signal subspace. The statistical characteristics involved the fuzzy mean, as well as fuzzy scatter matrix were defined based on the fuzzy statistics. In the feature space of FS-DCA, the unknown measurement vector (pixel) has membership grade values to describe the distance between the vector and the mean vector. The fuzzy separable subspace was then established to suppress the uncertainty caused by noise pixels. To verify this approach, a widely used hyperspectral imagery data set, i.e., AVIRIS Indian Pines 92AV3C was chosen including nine classes and two types of grow crops. The performance of the proposed FS-DCA method on AVIRIS hyperspectral images classification was compared with those of using all spectral channels and two representative feature extraction methods, the PCA and DCA. Experimental results showed that the proposed FS-DCA provided the better classification performance and lower standard deviation than the previous PCA and DCA. Specifically, the average overall accuracy of FS-DCA was much higher than that using all spectral channels, the PCA and DCA, ranking values in order 7.88, 3.28 and 0.5 percentage points. The producer’s accuracies and user’s accuracies of grow crops were significantly higher than that using the conventional methods, ranging from 1.37~18.47 percentage points. This finding demonstrated that the proposed FS-DCA technique can greatly reduce the dimension of hyperspectral images, and thus provide a promising method for the information extraction and classification of crops using remote sensing data.