A Probability Model-based Method for Land Cover Change Detection Using Multi-Spectral Remotely Sensed Images

Change detection is one of the main research areas in remotely sensed image processing. Image differencing methods have been widely used to quantify changed pixels by labeling such pixels with differencing images. There is room, however, to further develop the approach by enhancing the change detection reliability method by reducing the index sensitivity to seasonal variations. Using the information provided by image differencing results, a probability model-based change detection method is proposed in this study. A Chi-square distribution model is built using multiple index images based on the assumption that the pixels in the differencing image follow a normal distribution. By means of Chi-square distribution percentiles, different probability contours can be found to differentiate the changed pixels from all pixels in the feature space. The pixels located outside the probability contour will then, be identified as the changed pixels with a certain probability level. Tasseled Cap transformation components can be utilized to construct the Chi-square distribution, thus obtaining a higher accuracy of change detection. Due to the availability of multiple index images such as NDVI and Tasseled Cap transformation components, ETM+ images of Hong Kong on Aug. 20, 1999 and Sep. 17, 2002 were used as experimental data to test the performance of the proposed method. The experiments showed that the combination of NDVI and Brightness indices produced the highest overall accuracy and Kappa coefficient values.