Quality Analysis of Synthesized High Resolution multispectral Imagery
Yang-cheng Liao,Tengh-yih Wang and Wei-teng Zheng
Department of Surveying Engineering National Cheng Kung university
University Road 1,Tainan Taiwan
Tel (886)-6-2370876 Fax: (886)-6-237564
E-mail : ycliao@mail.ncku.edu.tw
Key words: image registration, image fusion, linear feature, quality analsis
Abstract
In Taiwan there exist significant natural environment changes. To aid environment management, planning and monitoring, the environment changes detection on the land must be done firstly. For that, the SPOT and LANDSAT imagery could be available, but their spatial and temporal resolution is not yet fine enough for practical application in Taiwan. It might be applicable to integrate high resolution panchromatic images and low resolution multispectral imagery into a synthesized high resolution multispectral imagery. The quality of synthesized imagery might be different from one of real high resolution . In this paper, the indices of correlation coefficients, signal-noise-ratio, momentum of differences pixel value and entropy value are used to evaluate the difference of both kinds of imagery
1. Introduction
The multispectral images provide necessary information for land cover interpretation, but don't have sufficient spatial resolution. Therefore, there are different methods to integrate high resolution BW (=black-white) image and low resolution multispectral images into synthesized high resolution images.(Yun Zhang 1997) analyzed the HIS method (=intensity, Hue, saturation), the PCS method (=principal component substitution), the RVS method (=Regression Variable Substitution ) and the SVR method (=Synthetic Variable Ratio), and verified that HIS and PCS yield finer spatial resolution, SVR provided good spectral information and RVS gives worst spatial and spectral data. (Shue and Tseng 1998) tries to utilize the hyperspectral imagery and the image registration and fusion techniques ti improve the data quality of SPOT images for land cover interpretation (Shieh 1994) studied four integration methods for SPOT multispectral and panchromatic imagery, namely colour space transformation, principal component analysis, high pass filter methods and radiometric method. In these method, only spectral data quality of the synthesized imagery are analyzed. To exam not only spectral but also spatial data quality of the synthesized imagery, this paper uses also the techniques of linear feature extraction,e.g. Foerstner Operator and canny operator (Chio and Wang 1996) investigated the accuracy of the Foerstner Operator on feature extraction. Canny Operator can be used to extract linear features (canny 1986). One can utilize the synthesized imagery
2. Image Synthesis
There are in general two steps to integrate high resolution panchromatic and low resolution multispectral imagery :image registration and image fusion
2.1 Image Registration
Image registration aims to overlay two images and unify their resolution and coordinate system. It contains generally there processes:
-
Selection and matching of GCP (=ground control points):In practice, GCP are often high contract feature, e.g.house corner points area are used to get an accurate registration. One can use area-based matching techniques or feature based matching techniques ( Lemmens 1988) to homologous GCP points in stereo image pairs .(Tang 1995) approved that a better matching accuracy can be reached if both matching techniques are combined together.
- Coordinate Transform :The high resolution panchromatic image is often used as a reference one to unify the image coordinate system. The GCP points on low resolution multispectral images are then transformed into the reference image system. The simple affine transformation is available e.g.for satellite (Jensen 1986). The other polynomial transformation of higher order might be also suitable, if the land surface is relatively flat and the angle of FOV (=field of view) is not large (Schowengerdt 1007).
-
Resampling: image value on a new transformed position in multispectral image can be determined by different interpolation methods, e.g.cubic convolution, bilinear interpolation and nearst neighbor method (pen and pan 1991).
The experimental images used in this paper have the same coordinate system, so that the coordinate transformation is not necessary. Moreover, the cubic convolution method is used to retain a best spatial resolution.
