Hyperspectral image ground-object identification method based on spectral segment fusion combination and depth residual network

Abstract

An algorithm based on the spectral segment fusion combination and deep residual network is proposed to improve the recognition accuracy of the objects of interest in the WHU-Hi dataset, particularly for cruciferous plants. The accuracy of the objects of interest was effectively improved, as well as the recognition accuracy of other ground objects, and the time efficiency was improved as well. The optimal combination of spectral segments was determined, and spatial and spectral information was extracted from the deep residual network for ground object recognition research. Experimental results showed that the classification accuracy of the cruciferous plants of interest, namely, pak choi, Brassica chinensis, and small Brassica chinensis, increased from 81.36%, 84.2%, and 83.8% to 98.32%, 99.22%, and 98.35%, respectively. In addition, the accuracy of interested trees and grass also increased from 77.6% and 89.09% to 99.12% and 98.33%, respectively, and the overall accuracy, KAPPA, and average accuracy of the three datasets were all improved. The time efficiency was also improved by an order of magnitude.