Spectral Unmixing Approach in Remotely Sensed Forest Cover Estimation: A Study of Subtropical Forest in Southeast China

AbstractAddressing the problem of spectral mixing in remotely sensed forest cover mapping, a linear spectral unmixing approach was employed in the study to assess if sub-pixel method would improve forest cover estimation accuracy in the context of complex subtropical forest ecosystem. After masking out water bodies using Modified Normalized Difference Water Index (MNDWI), the TM imagery of Pingnan County, Guangxi Zhuang Autonomous Region, China, was processed with Minimum Noise Fraction (MNF) Rotation transform and Pixel Purity Index (PPI), thus “pure” spectral endmembers of woody cover, herbaceous vegetation and bare ground were extracted as input into the spectral unmixing algorithm and produced forest map. The forest percentage is 55.7%, overestimated by 0.8% when compared with the National Continuous Forest Inventory 2004 statistics, reporting a fair agreement with ground truth. The approach also shows a better performance than Spectral Angle Mapper (SAM) classification (overall RMSE of 9.2 compared with 10.7 for latter)