Analysis of WorldView-2 band importance in tree species classification based on recursive feature elimination

Abstract

In tree species classifications, different spectral bands feature different importance, and the manner of determining the importance of one band is a problem that needs to be solved. In this study, eight bands of the WorldView-2 fusion data were used as information sources, and a recursive feature elimination based on maximum likelihood (MLC-RFE) was used to sort the importance of these bands. According to the results, the importance of the eight bands was sorted as follows (from important to unimportant): nearinfrared 2 > red edge > yellow > red > near-infrared 1 > coastal blue > green > blue. The poorest band combination yielded the lowest overall accuracy (OA) and Kappa coefficient (40.9153%; 0.3080), whereas the optimal band combination presented the highest OA and Kappa coefficient (74.5479%; 0.7029), indicating the large difference in accuracies between the optimal and poorest band combinations. Therefore, selecting important bands bears significance in tree species classifications. The MLC-RFE method significantly solved the band selection problem. Thus, this method should be extended to more complex feature selections.