Selection of Landsat8 Image Classification Bands Based on MLC–RFE

Abstract

In remote sensing image classifications, a reasonable selection of bands can help improve classification accuracy. At present, problems occur in artificial subjective selection of bands in image classifications, thereby resulting in ineffective expression of classification accuracy. In this study, 10 bands of Landsat8 fusion data were used as information sources, and recursive feature elimination based on maximum likelihood classification (MLC–RFE) was conducted to select the most important bands and analyse the classification results of the various band combinations. Results showed that green, red and blue were the three unimportant bands, which had the lowest classification accuracy (92.2312%). Coastal blue, near infrared and short-wave infrared 1 were the three important bands, which had a high classification accuracy (97.5000%). The coastal blue, near infrared, short-wave infrared 1, cirrus, short-wave infrared 2, thermal infrared 1 and thermal infrared 2 constituted the recursive optimal band set in all multi-band combinations and had the highest classification accuracy (99.1667%). The classifications of the 2–7 bands and all 10 bands had an overall accuracy of 95.3495% and 96.9624%, respectively. Additional experiments showed that the Landsat8 band elimination order and optimal band set have certain differences, but the optimal band combinations with high classification accuracies can be coupled from all of them. Our findings confirmed that various band combinations had different classification accuracies, which indicated that selecting the participating bands in a classification was crucial. Furthermore, the MLC–RFE method used in this study selected the optimal classification bands and played an important role in solving the band selection problem.