Hyperspectral adaptive band selection method through nonlinear transform and information adjacency correlation

Through nonlinear functional transform of hyperspectral remote sensing data, the modified Pearson correlation analysis can effectively identify comprehensive correlation coefficient (rcl), correlation type, and statistical significance level between spectrums. In this paper, nonlinear correlation the main correlation relationship type between hyperspectral bands was proved. Based on correlation coefficient, the adjacent bands' correlation coefficient (rac) of adaptive band selection (ABS) is to express band independence, but rac of ABS algorithm cannot effectively express such independence. Herein, a kind of information adjacency/equivalent bands' correlation coefficient (riac), and via this index, the modified ABS (MABS) were proposed. Using public data and collected private data, the experiments of ABS, MABS(rl) based on linear correlation coefficient(rl), and MABS(rcl) based on rcl were carried out. These two case studies demonstrate that MABS is superior to ABS on spectral range, algorithm validity and accuracy. MABS can take both large amount of information and strong independence into consideration effectively. The spectral range of MABS's bands selection result is more than ABS's obviously, and MABS (rcl)'s is a little more than MABS (rl)'s. The ranking both overall classification accuracy and Kappa coefficient of those three kinds of algorithms are MABS(rcl)MABS(rl)ABS.