Using Combined Linear Regression and Principal Component Analysis for Unsupervised Change Detection of Forest Fire

In this paper, an unsupervised change detection technique is used for obtaining the change mask of forest fire depending on bi-temporal images, pre-event, and post-event. While there are many techniques for change detection, they don't give the same result accuracy for different data types. Here, two mid-resolution remotely sensed RGB images are used with combined methods to estimate the change that resulted from a burn scar. Lack of enough resolution is a key factor in this scenario as it could be the only data to deal with. Linear regression is utilized to find a linear relationship between the two images and use the linear parameters for constructing the background scene of the post-event image. The post-event image, the background image and the difference between them, in addition to greenness index images of pre-event and post-event with their difference, are processed by Principal Component Analyses (PCA) to obtain the most descriptive principal component of the change. A simple greenness function is proposed here depending on segmentation in HSV (hue, saturation, and value or intensity) color space to indicate the vegetation index. The results of this combined method confirm that it is more accurate in comparison with the methods it combines each alone.