In this paper we present a technique for the efficient representation of the change information available in multitemporal and multispectral remote sensing datasets. When dealing with multispectral images, the detection of changes requires the solution of an N-dimensional (ND) problem. Due to the complexity of such a problem, the most common practice is to exploit only 2 (or few) spectral bands, by using prior knowledge about possible changes occurred on the ground for bands selection. However, when prior information is not available each spectral band potentially includes useful information for change detection and cannot be neglected a priori. Nevertheless, the resulting multidimensional problem is difficult to be approached in an unsupervised way. In order to reduce both the complexity of the change detection problem with a limited loss of change information and to simplify the visualization of the change information, we propose an effective procedure for dimension reduction (from N to 2). A new feature space is proposed defined by the magnitude of spectral change vectors obtained subtracting multitemporal images, and an angle measure evaluated between a proper reference vector and the multidimensional spectral change vector. In the new feature space classes of unchanged and changed pixels (and different kinds of changes) can be separated according to thresholding procedures. The effectiveness of the proposed technique was tested on two multitemporal and multispectral datasets: one acquired by the Thematic Mapper sensor mounted on the Landsat 5 satellite, and one acquired by the very high geometrical resolution sensor mounted on the Quickbird satellite.
展开▼
