Acta Geodaetica et Cartographica Sinica ›› 2016, Vol. 45 ›› Issue (1): 80-86.doi: 10.11947/j.AGCS.2016.20140520

Previous Articles     Next Articles

Nonlinear Spectral Unmixing for Optimizing Per-pixel Endmember Sets

LI Hui, ZHANG Jinqu, CAO Yang, WANG Xingfang   

  1. Computer School, South China Normal University, Guangzhou 510630, China
  • Received:2014-10-20 Revised:2015-07-31 Online:2016-01-20 Published:2016-01-28
  • Supported by:
    The National Natural Science Foundation of China (No.41171288);The National Natural Science Foundation of Guangdong Province of China(Nos. S2013040016473;S2013010014097)

Abstract: For a given pixel, fractional abundances predicted by spectral mixture analysis (SMA) are most accurate when only the endmembers that comprise it are used. This paper presents a support vector machines (SVM) method to achieve land use/land cover fractions of remote sensing image using two steps: ①defining the optimal per-pixel endmember set, which removes endmembers based on negative fractional abundances generated by SVM method; ②using SVM extended with pairwise coupling (PWC) to output probabilities as the abundance of landscape fractions. The performances of the proposed method were evaluated with the multiple endmember spectral mixture analysis (MESMA) method, which has been widely applied to map land cover for the goodness of the model fitting. The results obtained in this study were validated by real fractions generated from SPOT high resolution geometric (HRG) image. The best classification results were obtained by the proposed method indicated by the lower total mean absolute error, the higher overall accuracy, and the higher kappa. From this study, the proposed method is proved to be effective in obtaining abundance fractions that are physically realistic (sum close to one and nonnegative), and providing valuable application in selecting endmembers that occur within a pixel.

Key words: remote sensing image, pixel unmixing, selective endmember, support vector machines (SVM), nonlinear unmixing

CLC Number: