Burn Scar Delineation from Landsat Imagery Using Modified Chan-Vese Model

摘要

In this paper, we propose a novel semi-automatic method for burn scar delineation from Landsat imagery using modified Chan-Vese model. Burn scars show reddish-brown from Landsat's band 742 combination. This feature can be used for semi-automatic delineation of burn scars. Usual segmentation algorithms detect all objects in the image, while our goal is to detect burn scars from a given Landsat image, regardless of other objects. Our algorithm is based on segmenting objects of interest (OOI) using modified Chan-Vese model. Firstly, we choose sample pixels visually from the burn scar. From these sample pixels, a discrimination function for whether a pixel belongs to burn scar or not is determined by the principal component analysis and interval estimation. Then we define a modified Chan-Vese functional. In order to minimize this modified Chan-Vese functional, the evolution equation corresponding to the functional is given. The initial contour of the evolution is a function of the discrimination function, and automatically located near the boundary of OOI. An improved stopping function is used in our evolution equation. Our stopping function depends on image gradient and the discrimination function. This stopping function helps the evolving curve reaches the exterior and interior boundaries of burn scar fast and accurately. Landsat imagery of burn scars over Russia is used to test our algorithm. The result shows that the algorithm is very accurate for delineating exterior and interior burn scar boundaries.