Abstract
Recently, remotely sensed images have had very high spatial resolution. It is difficult for the most of conventional classification methods to obtain land use maps by using such high resolution images, because these methods are too sensitive to variations of pixel intensity in the same land use area. The reasons are as follows : 1) these methods are only based on spectral (color) and spatial (shape) information ; 2) these methods process the target images in pixel-by-pixel. Therefore, a new image classification method based on not only spectral and spatial information but also texture information is required to obtain land use maps similar to visual interpretation results.
In this paper, a new texture quantification method for remotely sensed images is proposed. This method has invariant property to translation and rotation of the texture patterns in the target images. In the method, a target image is transformed into spatial amplitude spectra by using 2D-DFT to reduce the influence of image translation, followed by quantifying these spectra by a set of complex Zernike moments to reduce the influence of image rotation.Then, the vector which consists of these moments is normalized to reduce the influence of image size. The normalized Zernike moment vector (NZMV) can describe the texture as a unique vector from the origin to a point on the surface of unit hypersphere in the space spanned by Zernike moments, because the elements of the NZMV (Zernike moments) are mutually orthogonal.
To evaluate the texture discrimination capability of the proposed method, some experiments by using 15 images from Brodatz's photo album and their rotated images with every 30-degree rotation angle are conducted. For discrimination of two texture patterns, the angle between two NZMVs obtained from each texture is employed. The experimental results show that the proposed method can identify the same texture images exactly, even though the target images contain various translations and rotations. The proposed method has been also applied to supervised texture classification of airborne multispectral scanner images. The results obtained by the proposed method are closer to those by visual interpretation in comparison with conventional maximum likelihood method.
