|
Flood-induced changes over a wide area can be detected and assessed using bitemporal high-resolution satellite images in a timely and cost-effective manner [1]. Multivariate alteration detection-based method [2] is designed to use twice of the standard deviation of a variate as a threshold to separate change from no-change pixels in the variate and the final detection result is the union of individual variates results. However, the threshold in this method is significantly affected by the number of pixels of no-change to the number of pixels with change which leads to high omission error. Moreover, using all the whole set of variates increases the commission error due to including the variates with high correlation. In this paper, we introduced a novel method to decide between change and no-change, by using Otsu technique which is an optimal technique to separate between classes in gray level images [3]. Also, we proposed using the variate of least correlation instead of using the whole set of variates in detecting the induced change. Using this variate is justified since a variate with least correlation includes the induced change between bitemporal input images. The proposed algorithm is summarized as follows: Firstly, we have adjusted radiometric reflectance of the images that acquired at different times by using normalization which is a critical step in the application of flood-induced change detection. Multivariate alteration detection [2,4] is implemented as an approach for normalization. Secondly, we employed the variate with least correlation for induced flood detection by using Otsu method. To evaluate the proposed algorithm, real bitemporal-images of flood events are employed in this study. The experimental results on these images show that the proposed method quantitatively and qualitatively has outperformed the traditional MAD-based method for flood induced-detection and simultaneously reduced omission and commission errors.
|
