A Novel Image Classification Algorithm Using Overcomplete Wavelet Transforms

A novel frequency-based classification framework and new wavelet algorithm (Wave-CLASS) is proposed using an overcomplete decomposition procedure. This approach omits the downsampling procedure and produces four-texture information with the same dimension of the original image or window at infinite scale. Three image subsets of QuickBird data (i.e., park, commercial, and rural) over a central region in the city of Phoenix were used to examine the effectiveness of the new wavelet overcomplete algorithm in comparison with a widely used classical approach (i.e., maximum likelihood). While the maximum-likelihood classifier produced < 78.29% overall accuracies for all three image subsets, the Wave-CLASS algorithm achieved high overall accuracies - 95.05% for the commercial subset (Kappa = 0.94), 93.71% for the park subset (Kappa = 0.93), and 89.33% for the rural subset (Kappa = 0.86). Results from this study demonstrate that the proposed method is effective in identifying detailed urban land cover types in high spatial resolution data.