Comparison of fusion algorithms for GF-2 data from extracted forestland information

Abstract

To obtain an optimal method for image enhancement of GF-2 forestry area data, six frequently-used methods were analyzed:Brovey transformation; hue, saturation, and value (HSV) transformation; Principle Component (PC) spectral sharpening; high pass filter (HPF) spectral sharpening; Gram-Schmidt spectral sharpening; and Pansharp transformation. Qualitative and quantitative analyses were used to assess the effect and quality of the fusion images. Indexes include mean, average gradient, high-frequency information integration, correlation index, entropy index and second moment index. Among them, correlation index and second moment index were calculate by ENVI, other indexes were all by Matlab. Furthermore, to access an appropriate fusion method for GF-2 forestland data extraction, fusion images were classified by performance of fusion methods at two information extraction levels based on an object-oriented classification method. All the transformations used the same parameter and methods on each level, and use the same samples to classify and accuracy check. Results showed that correlation index and high-frequency information integration of HSV transformation could reach 0.823 and 0.570, respectively. In addition, the entropy index and second moment index could improved 25% and 50% compared to original multiple image, respectively. It had a better visual effect with obvious enhanced clarity and texture features. For classification experiments, HSV and Brovey transformations had their own superiority for the extraction of different classes with the HSV transformation having the highest overall classification accuracy of 85.1% and the Brovey transformation having the highest accuracy on the second level of 75.7%. The other four methods had different advantages for quality and information extraction of the fusion images. Thus, the final selection of fusion methods should consider practical forestry application and image information which could provide a reference for GF-2 images to be applied on a large scale in forested areas.