Resources Science ›› 2017, Vol. 39 ›› Issue (3): 557-565.doi: 10.18402/resci.2017.03.17

• Orginal Article • Previous Articles     Next Articles

Land use and cover classification based on multi-feature soft probability cascading

Bin ZHANG(), Yueyan LIU, Linyu WANG   

  1. China University of Geosciences,Department of Land Resources Management,Wuhan 430074,China
  • Received:2016-09-20 Revised:2017-02-05 Online:2017-03-20 Published:2017-03-20

Abstract:

In order to realize the effective organization and optimization of low-level features for high resolution remote sensing images,and make feature descriptors more discriminative,we focused on mid-level feature learning based on sparse coding and support vector machine (SVM)classification,and propose a land use / land cover (LULC)classification method based on soft probability cascading and mid-level feature learning model. First,the gray level co-occurrence matrix (GLCM),Dense Scale-Invariant Feature Transform (DSIFT)and spectral feature descriptors are extracted as low-level feature descriptors. Second,sparse coding is adopted to obtain the sparse coefficients of GLCM,DSIFT and spectral features,and then max-pooling methods are used for learning mid-level feature descriptors. Combined with SVM classification with training samples of LULC class types the soft probabilities of different LULC class types are calculated. Three different kinds of soft probabilities belong to each LULC class type,which takes GLCM,DSIFT and spectral features as low-level feature descriptors respectively,and are cascaded for the construction of final feature descriptors. The cascaded feature descriptors are more discriminative than unsupervised mid-level feature descriptors,because it is learned by a supervised way. This method incorporates different low-level feature descriptors effectively. Finally,using the cascaded feature descriptors,the LULC classification map is achieved by the SVM classifier in a supervised way. Taking rural residents in the district area of Wuhan City as an experimental area,our proposed method was verified by aerial high resolution remote sensing images. Experimental results show that the overall accuracy is 88%. Compared with the extraction of a single low-level feature classification method,the algorithm in this paper can effectively improve LULC classification accuracy.

Key words: high-resolution, remote sensing images, image classification, land use and land cover, sparse coding, support vector machine