Abstract:

The existence of mixed pixels not only affects land cover type recognition and classification accuracy based on hyperspectral images, but also has become a major obstacle to the quantitative development of remote sensing science. Taking the Zhalong Nature Reserve as a study area, the current study compared the performance of the sparse constrained least squares spectral unmixing algorithm (SUFCLS) and the fully constrained least squares spectral unmixing algorithm (FCLS) for wetland remote sensing classification. The classification accuracy and errors of the two algorithms were evaluated and analyzed. The results show that the SUFCLS algorithm adaptively selected the highest percentage endmember combination from the spectral library, and integrated the selected endmembers into the FCLS algorithm to conduct the abundance inversion. Having considered the spatial heterogeneity of endmembers, the SUFCLS algorithm overcomes the shortcoming of the FCLS algorithm during the process of endmembers selection. Compared with the FCLS, higher correlation was observed between the classification results of SUFCLS and the abundance of the wetland communities (reed swamp, cattail marsh, leymus chinensis meadow, and weed meadow) visually interpreted from the high-resolution imagery. In addition, the root mean square error (RMSE) decreased, which indicates that the SUFCLS algorithm has an important significance in improving wetland unmixing accuracies and implementing wetland high precision classification.

Key words: hyperspectral image, sparse unmixing, linear unmixing, wetland classification, Zhalong Natural Reserve