Quality Analysis of Synthesized High Resolution multispectral Imagery
2.2 Image fusion
Image fusion integrates both spatial and spectral data to hold superior characteristics of high spatial and spectral resolution and no improve our knowledge of a scene. The fused image should improve the image classification accuracy ( Munechika 1993) and helps feature extraction and recognition (vrable 1996). The image fusion methods can be divided into two classes: spatial domain method and spectral domain method.(Chavez 1986) proposed to extract firstly the high frequency component from the high spatial resolution image and add it to the low resolution image. This is one of the spatial domain methods. The spectral domain methods to be used in most application are color space transformation and principal component analysis. Therein the color space used in most image processing are the RGB (=Red-Green-Blue) color space, YIQ (=Y-signal, Inphase, Quadrature phase) color space, and HIS(=Intensity-Hue-Saturation) color space. In this paper, the HSV(=Hue-Saturation-value) model is used as a color space and the image fusion done as follows:
-
The RGB color space is transformed to the HSV model:
-
The gray value g' of a pixel in the black-white image is used as the Value in the related color image, i.e. in the above equation V=g'.
-
The HSV model is transformed inversely to the RGB space.
Thus, we complete image fusion and get high resolution color image.
It is feasible the value in the HSV model for a image is replaced in our experiments by the gray value g' of the corresponding black-white image, because the RGB bands overlap with the panchromatic band. If the multispectral imagery covers the infrared band, the above -mentioned method for image fusion is not usable any more, because the intensity of infrared is not correlated with the one of visible light. To solve the problem, the fusion method proposed by carper is utilized in this paper to merge panchromatic and multispectral images of SPOT data. Table 1 shows the SPOT spectral bands. In the color space transformation , the vector (R,G,B,) is replaced by the vector (XS3,XS2,XS1).since the panchromatic band don't cover the infrared one, the improved fromula
is used as the intensity in the IHS model, namely the value in the above-mentioned HSV model.
4. Quality Analysis
A. Four Methods for Spectral Data Quality Analysis
Four methods are used here to study spectral differences between synthesized and real high-resolution images. They are correlation coefficient analysis, momentum analysis, entropy analysis and signal -to- noise analysis
3.1 Correlation coefficient
Where x
1,y
1 are the gray value of homologous pixel synthesized image and rreal high-resolution image, and
are the mean gray values of both images. A larger absolute value r means a higher correlation and vice versa.
3.2 Moment
This method computer the differences of gray value of homologous pixel, and their mean value (1
st order momentum) and standard deviation (2
nd order momentum) as well to estimate spectral variation of synthesized and real images.
3.3 Entropy
The following first-order entropy is used as an index to measure the amount of image information:
where p(i) is the probability of the i-th gray value level
3.4 Signal-to Noise Ratio
Hereby, only white noise is considered and it is supposed to be an independent variable of a Gaussian distribution with the mean of zero. Thereupon, image variance equals the summation of signal variance and noises one. The standard deviation of gray value in a homogeneous image area is regarded as the estimation of noise variance. Then, image variance minus noise one is equal to signal variance. Signal variance divided by noise one yield the SNR value.
Additionally, image feature extraction is done for both kinds of images. The number and quality of extracted features are used to measure their spatial data quality.
