A new classification algorithm for high spatial resolution remotely sensed imagery is proposed

which integrates neighborhood information of multiscale such as 2×2

4×4

8×8 and 16×16 window sizes around the central pixel.In order to compress the information of the multiscale spatial features

a wavelet coefficients fusion algorithm is employed to reduce the dimension but retain the spatial information at the same time.After the stage of multiscale neighborhood feature extraction

a good tool of pattern recognition: SVM is employed to process the multiscale features

in this algorithm

four groups of spatial features based on four scales produce four classification maps.And then

these maps

which represent multiscale classification results

are fused by a scale selection parameter.The final fusion map is the result of multiscale features classification and shows an obvious adaptability to objects of different scales.Experiments of QuickBird and Ikonos show that the proposed classification algorithm of multiscale features fusion can achieve better results and better accuracies than the conventional per-pixel multispectral method.