Combination of decision tree and mixed pixel decomposition for extracting bamboo forest information in China

Abstract

A bamboo forest is a special and important forest resource that is distributed in subtropics of China. However, present methods cannot quickly and accurately extract spatiotemporal distribution information of a bamboo forest at the national scale. This study presents a method for extracting bamboo forest information in China by combining decision tree and linear spectral unmixing methods using MODIS NDVI, reflectance product, and provincial Landsat classification data in 2003, 2008, and 2014. The steps are as follows: First, forest information was extracted based on the maximum likelihood method; second, a decision tree model was constructed to extract bamboo forest distribution information based on a forest information map; finally, the linear spectral unmixing method was applied to obtain the abundance map of a bamboo forest in China, and the bamboo forest area was calculated. The main experimental results are as follows: (1) The forest information in China was extracted through the maximum likelihood method. The accuracy of a producer and a user was higher than 90%, and the Kappa coefficient was 0.93, thereby establishing the extraction of bamboo forest information. (2) The decision tree constructed by using a C5.0 algorithm satisfactorily extracted the bamboo forest spatiotemporal distribution information in China, with an average classification accuracy of 80%. (3) The estimated bamboo forest area of each province in China was highly correlated with the observations, with the R2 values of 0.98, 0.97, and 0.95, correspondingly. The RMSE ranged from 3.92×104 to 9.58×104 ha, thus indicating that the estimated bamboo forest area was essentially consistent with the actual situation. In this study, the C5.0 algorithm decision tree, which is based on the MODIS remote sensing data, combined with the mixed pixel decomposition presents accurate extraction of bamboo forest spatiotemporal distribution information in China. This method provides the technical approach and data support for the dynamic monitoring and management of bamboo forest resources.