A comparative analysis on hyperspectral land-cover classification based on frequency spectrum and spectral characteristics

ObjectiveFrequency spectrum as an energy feature of matter can be used to differentiate ground object recognition. However, the study for classification by frequency spectrum is scanty at home and abroad, especially for forest vegetation classification research. This study explored the potential of frequency spectrum in identifying the objects by converting optical spectrum of image into frequency spectrum.

MethodFor exploring frequency spectrum in land-cover classification, the optical spectrum was converted to frequency spectrum to classify based on the fused hyperspectral images from domestic HJ-1A HSI and CCD data by CN Spectral Sharpening in farmland of Jiangle County, Sanming City of Fujian Province, southern China. Then the distance was developed to differentiate the objects based on the obtained frequency spectrum. The classification result based on frequency spectrum was compared with spectral angle mapping (SAM) based on spectrum space.

ResultThe study result showed that the separability was clear among frequency spectrum of different ground objects. The frequency spectrum of the vegetation was significantly different with frequency spectrum of non-vegetation classed, and the spectrum of different tree species can also be differentiated in the low frequency region. The precisions for Pinus massoniana, Cunninghamia lanceolata and broadleaved forest were improved based on the FSSM method. The non-vegetation classes can be distinguished using frequency spectrum of first order harmonic. The vegetation classes can be distinguished using frequency spectrum from one to seven harmonic. However, the frequency spectrum for all objects tends to be similar with the increase of frequency. And the energy accumulated speed was faster for non-vegetation classes. Compared with SAM result, the overall classification precision from frequency spectrum method increased by 0.7%, which was 84.19%.

ConclusionThe comparison results from frequency spectrum of different land types and classification results indicated that the frequency spectrum can be efficiently applied to identify objects. This method can reduce the influence of the noise from spectrum curve and keep the important distinction information between the classes. So, the frequency spectrum can be used for object identification.