Classification of plantation types based on WFV multispectral imagery of the GF-6 satellite

Abstract

The high-spatial-resolution GF-6 satellite has been successfully launched for less than one year, and the application of its imagery has just started. GF-6 is a satellite of a high-resolution earth observation system, which is a major scientific and technological project in China. The Wide Field-of-View (WFV) images of the GF-6 wide-format camera add two red-edged bands in comparison with similar images of the GF-1; thus, the monitoring capacity of agriculture, forestry, and grassland is improved. To analyze the ability of GF-6 WFV imagery in plantation classification and promote the further application of GF-6 data in forestry field, this study provides a hierarchical classification method to extract plantation types, such as eucalyptus and fir, by using the latest WFV images of GF-6 in the Nanning forest farm in Guangxi Province. Furthermore, the classification process was combined with ground measured data. The random forest classification method was adopted, and the steps are as follows: First, the six vegetation indices were calculated and optimized using the random forest feature selection method, and then the classification scheme of the four datasets was determined. The schemes are as follows: (1) first four bands of WFV image without red edge, (2) eight bands with red edge, (3) eight bands plus unoptimized vegetation index features, and (4) eight bands plus optimized vegetation index features. Then, the random forest classifier was utilized in four datasets. Random forest is an effective classification method that uses the classification regression tree algorithm to generate classification trees and combines the advantages of bagging and the random selection of feature variables. Lastly, the classification results of four schemes were compared; in addition, accuracy assessment was performed in accordance with field survey and forestry inventory data.Results showed that Scheme 2 has an accuracy of 4.99% higher than that of Scheme 1, and the kappa coefficient increased by 0.058. These results indicate that the accuracy of eight-band data with red edges is significantly improved in comparison with four-band data. The classification accuracy of the eight bands plus optimized vegetation index features was the highest among the schemes, reaching 85.38%. Compared with the bands with and without red edge, the accuracy improved by 3.98% and 8.97%, respectively; moreover, the kappa coefficient increased by 0.046 and 0.104. WFV images are added with the red-edge index, thereby enhancing vegetation information. This dataset can accurately reflect the differences of plantation type characteristics and significantly improve the classification accuracy of the plantation. Therefore, this study can effectively improve the information extraction effect of plantation types in Guangxi Province and provide a scientific reference for the evaluation of GF-6 image quality and its forestry application potential.