Determination of Threshold in Change Detection Based on Histogram Approximation using Expectation Maximization Algorithm

摘要

One of the key issues of land use/cover change detection using remote sensing images is the threshold determination, which is discriminate of changed and unchanged pixels in difference images . This paper introduces the histogram approximation method based on Expectation Maximization (EM) algorithm into unsupervised change detection. Firstly, obtained co-registered multi-spectral satellite images covering the same geographic area (Fig.l) and applied the change vector analysis (CVA) technique to acquire the "difference image". Namely, for each pair of corresponding pixels, a "spectral vector" is computed as the difference between the feature vectors in two times, the pixel's value of the difference image is the modules of the "spectral change vectors". Consequently, the change magnitude is in direct ratio to the pixels' values, show in Fig.2. The larger values indicate the higher change magnitude.