Abstract:Spatial autocorrelation information is easily lost in the process of traditional texture information extraction methods of hyperspectral image，leading to low accuracy of vegetation classification. An improved scheme was put forward aiming at the shortcoming of existent methods to form a new classification algorithm (CHISCI) based on spatial autocorrelation information. Firstly, one kind of spatial autocorrelation information of hyperspectral image was extracted by domain transform filter of interpolated convolution（DTFOIC）. Secondly, another kind of spatial autocorrelation information was obtained by the same filter on dimensionality reduced hyperspectral data. Finally, the two kinds of spatial information were combined and then classified by SVM which was not sensitive to high-dimensional data, forming CHISCI classification algorithm of hyperspectral image by spatial autocorrelation information.The CHISCI classification method was implemented on the hyperspectral data of Indian Pines and Salinas Valley. The following results were obtained. In the first place, the overall accuracy (OA) of Indian Pines was 96.16% and the Salinas Valley was 98.67%, which were 12~16 percentage points higher than those of SVM and PCA-SVM, and 4~16 percentage points higher than those of SGB-SVM, SBL-SVM and SGD-SVM by spatial-spectral information, and 4~6 percentage points higher than that of EPF, and 2~3 percentage points higher than that of IFRF. Furthermore, the average accuracy (AA) and Kappa of the CHISCI were also increased substantially, showing very good performance in hyperspectral classification. In the second place, although the training samples were only 6% of Indian Pines and 1% of Salinas Valley, the OA of both can reach 96.16% and 98.67%, which can remove salt and pepper noise in the classification map obviously. When the training samples were reduced to 3% and 0.3%, the OA can be over 90% and 95%, respectively. The effectiveness of CHISCI was fully verified in the hyperspectral classification. In the last place, the classification of some methods for grapes_untrained and vinyard_untrained in Salinas Valley were bad. The reason was that the spectral reflectances of the two vegetables for all bands were very close. However, the classification for the two vegetables of CHISCI can still reach 98.38% and 99.17%. It was showed that the CHISCI had excellent performance on the vegetable classification with close spectra. The experiments showed that the CHISCI algorithm was better than original SVM with pure spectrum information, the dimensionality reduction-based methods, the spatial-spectral information-based methods, and the methods based on edge-preserving filtering and recursive filtering. With the spatial autocorrelation information extracted by the DTFOIC, the performance of the classification of hyperspectral image with CHISCI algorithm was greatly improved, and the effectiveness of CHISCI was fully verified in the classification of hyperspectral vegetables, especially of those with close spectra. The method can be applied to the field of crop growing, diseases and pests monitoring, accurate classification and identification. It would also have potential significance for precision agriculture and agricultural modernization.
