Performance Analysis on Image Fusion Methods

Transferring spatial details of a high-resolution image into a low-resolution one is called image fusion. There are some different fusion methods introduced. Due to the nature of fusion process, these methods may damage the spectral quality of the low-resolution multispectral image to a certain extent. In the literature, there are some metrics that are used to evaluate the quality of the fused images. Depending on their mathematical algorithms, these quality metrics may result in misleading results in terms of spectral quality in the fused images. If fusion process is successful, the classification result of the fused image should not be worse than the result acquired from raw multispectral image. In this study, Worldview-2, Landsat ETM+ and Ikonos multispectral images are fused with their own panchromatic bands and another Ikonos image is fused with Quickbird pan-sharpened image using IHS, CN, HPF, PCA, Multiplicative, Ehlers, Brovey, Wavelet, Gram-Schmidt and Criteria Based fusion methods. The fused images are then classified with Minimum Distance, Binary Encoding, Support Vector Machines, Random Forest, Maximum Likelihood and Artificial Neural Network classification methods by using exactly the same signatures. 450 points are used to calculate the post-classification accuracies of classified images. Some commonlyused metrics (Spectral and Spatial ERGAS, Spectral and Spatial RMSE, SID, SAM, RASE) are also calculated for all fused images. To determine the best fusion algorithm, the consistency between classification results and fusion metric results are evaluated together. HPF fusion method is found out to be the most successful fusion algorithm in terms of preserving the spectral quality and Support Vector Machines classifier is found out to be the best classification algorithm for these data sets.