A novel method of feature extraction and fusion and its application in satellite images classification

摘要

Diverse and complementary feature descriptors extracted from the same satellite scene always reflect different characteristics of data. Consequently, it is worthwhile to implement the extraction and fusion on multiple feature descriptors. In this letter, we resolve the problem of multiset correlation feature extraction and fusion in multiple feature descriptors. To this end, a new approach is suggested to perform multiset correlation feature extraction and fusion for classification named global geometric multiset canonical correlation analysis (GGMCCA), which couples global geometric nature of data in the transformed space of low dimensionality and correlational characteristics of any pair of feature sets. In this regard, GGMCCA possesses superiority in discriminant capability in contrast to a previous approach named multiset integrated canonical correlation analysis (MICCA), which merely takes correlations into account for classification. Extensive experimental results on the satellite scenes demonstrate that GGMCCA surpasses multiset canonical correlation analysis and MICCA in computational efficiency and classification accuracy.