Multitemporal Remote Sensing Image Change Detection Using Image Structure Information

Multitemporal remote sensing image change detection has been used in civil economy and national defense widely. Through change detection between multitemporal remote sensing images of the same region, one can obtain the land change information that can be used for updating data of resource and environment, analyzing the situation of battlefield, assessing the battle damage and so on. So far, many methods have been developed for change detection. These methods can be classified as pixel-based, feature-based and object-based methods according to the image information level used in detection. Image structure information, such as edges and texture, is robust to the illumination change and is the main characteristic of man-made object in high resolution remote sensing imagery. Change detection based on structure information is robust to the illumination change, the noise and misregistration error. This paper focuses on how to develop change detection methods based on the edge structure information and differential structure information.The edge structure information, which can be described by Edge Tag (ET) based on biological vision principle, is a kind of structure information. Changes are decided by calculating the correlation between the ET vectors. This paper proposed an improved ET method by using more proper visual receptive field model,reducing the components of ET vectors and introducing the region growing technology. The experiment results show that the improved method has better performance than the ET method, and can detect more continuous edge and is more efficient in calculation.The image differential structure information, which can be described by using the differential invariants, is an important method to characterize the image structure information. The differential invariants are robust to the noise and also invariant to the translation and rotation. This paper proposed a new change detection method using the differential invariants. The differential invariants are firstly used to form local invariant feature, then the invariant features from two images is compared by calculating the Mahalanobis distance between two features, and finally a threshold of the distance is used to decide whether or not changes take place. The experiment results show this method is robust to noise and misregistration error.