Multi-resolution decomposition in relation to characteristic scales and local window sizes using an operational wavelet algorithm

Abstract

Data from an IKONOS image acquired over Dallas were used to demonstrate the use of an operational wavelet-based algorithm to examine the performance of different texture measures and window sizes at various resolutions in connection with characteristic scales. It was found that a 63 × 63 window was the optimal size, and energy measure produced the highest accuracy. Results from this study suggest that the choice of window size in wavelet-based classification affects the accuracy. Larger window sizes significantly improve the overall accuracy when using homogeneous samples. In the real-world situation, a larger window may not necessarily produce higher accuracy since a larger window tends to cover more land-use and land-cover classes and therefore may miss smaller regions of classes that could lead to poorer accuracy. On the other hand, a smaller window tends to be incomplete in its coverage of texture features that represent a complex class. The classification accuracy can be improved by using more combinations of sub-images at different scales. However, smaller sub-images at the last two levels may lower the classification accuracy. The characteristic scale of the most complex feature among all selected classes could be the optimal local window size necessary to achieve the highest accuracy.

Acknowledgment

This research has been supported by the National Science Foundation (grant no. 0351899). The author wishes to thank Gautam Gudhar and Liran Ma for their assistance in coding the algorithms.