GISdevelopment.net ---> AARS ---> ACRS 1990 ---> Poster Session Forest vegetation information of multispectral image from space and it's false color display tradeoff Zhu Qijiang, Liu Jinying Beijing Normal University, Beijing, China Abstract The vegetation response to environment is very sensitive .We need to concern both the quality and the temporal of imagers when selecting the images. In order to compare the performance of land sat MSS TM images and SPOT HRV images with infrared color photographs we select three images windows for processing in Pingquan county, Hebei province. The problem of selecting a subset of a multi image for enhancement by false color compositing rationing or differencing is generally difficult Usually an intuitive selecting based on physical characteristics of the scenes and on experiences is made multispectral images often exhibit high correlations between spectral bands therefore the redundancy between the components of such multi images may be significant. It's common knowledge that there are similarities between some two image bands of land sat TM. A false-color composite image generated by three different bands which are dependent on each other may give more information of physical landscape one way is to choose three difference image bands in which exhibition of lower correlation exists for enhancement by false color composite the correlation matrix of Dawopu digital image window are shown in table 1. Table 1: Correlation matrix of TM digital image (expect band 6) Dawopu image window Pingquan county 1. 1.000 2. 0.876 1.000 3. 0.833 0.964 1.000 4. 0.108 0.216 0.043 1.000 5. 0.730 0.842 0.773 0.531 1.000 7. 0.845 0.922 0.913 0.223 0.918 1.00 Based on the independence of image data we can distinguish the original six image bands as three image data subset visible image subset (TM1,TM2,TM3,) near infrared image subset (TM4) and mid infrared image subset (TM5,TM7) thus all of possible optional color display subsets of TM images are: (TM4,TM5,TM3),(TM4,TM5,TM2),(TM4,Tm3 TM1); (TM4,TM2,TM7),(TM4,TM3,TM7) and (TM4,TM5,TM7). Color plate 1 (TM4,TM3,TM2) and color plate 2 (TM,TM5,TM3) show two land sat TM false-color composite images of Weizhangzi image window (Pingquan county ) by linear stretch at the same time in different way. The color plate 1 corresponds to standard false color composite The space geometric features in color pears in color plate 2 are almost the same in accuracy but more spectral information appears in color plate 2 than color plate 1 the plate 2 can show the difference not only in crops category but also in water condition and immaturity of crops. Orthogonal Transform Based on statistical Features of image data. The K-L transform to principal components provides a new set of component images that are in correlated and ranked so that each components ha variance less than the previous component. Thus, the K-L transform can be used to reduce the number of spectral components to fewer principal components that account for all but a negligible part of the variance in the original multi spectral image The principal components images may be enhanced combined in to false color composites. The K-L transform image F° is obtained p components of a multispectral image X by the transformation. F° = t(X-Mx) Where X is a vector whose elements are the components at a given location (j, K) in the original multispectral image, Mx is the mean vector of X the components of vector f are the principal components at the location (j,k), T is the P by p unitary matrix whose rows are the normalized eigenvectors tp of the spectral covariance matrix Cx of X arranged in a descending order according to the magnitude of their corresponding eigenvalues, the covariance matrix cx is computed as : Cx = E [ (X-Mx) (X-M)T ] The eigenvector tp from the basis of a space in which the covariance matrix is diagonal therefore the principal components are uncorrelated. The means and various of six Tm bands (expect the thermal infrared band 6) of Dawopu image window Pinquan county are shown in table 2. The table 2 shows that larger variance appear in three reflectance infrared bands of TM than is visible bands and the most abundant information is given in TM5 band. Table 2: means and variances of TM digital image (expect band 6) Dawopu image window Pingquan county Channel Wavelength (um) Mean Variance 1 0.42-0.52 60.252 41.226 2 0.52-0.60 26.578 26.436 3 0.63-0.69 24.592 74.716 4 0.76-0.90 78.801 283.478 5 1.55-1.75 71.921 416.664 6 2.08-2.35 25.107 114.679 The eigenvalues the percentage variances and the cumulative percentage age variances are calculated and shown in table 3. Table 3. eignvalues and cumulative percent variance. Parameter Principal Component 1 2 3 4 5 7 Eigenvalues 691.098 234.514 18.187 7.658 2.872 1.904 Percent variance 0.772 0.245 0.019 0.0008 0.003 0.002 Cumulative percent variance 0.772 0.967 0.986 0.994 0.997 0.99 The first principal component image contains 77.2 percent of the original data variances the first three principal component images contain 98.6 of K-L transform and contain obvious physical shown in color plate 3 there is very abundant vegetation information in color plate 3 that almost is a fine vegetation distribution map. The interpretation results on vegetation from color plate 3 are shown in table 4. Table 4. interpretation list for false display of components PCI ( red ) PC2 (green) PC3 (Blue). Tone Land use / land Cover Purple bare land or spared withered grass land yellow shrub (down edge of forest ) or farmer land with irrigation condition Red farmer land with nature crops Black Chinese pine forest Blue black larch forest Dark blue mixed forest of birch and Chinese pine Light blue birch forest (young growth ) Azure meadow Grass green birch forest. Trade-off in false-color composite of the rationing images. Pingquan county is located in high mountain region with heavy shadows that hinder the interpretation of vegetation We need to develop a technique to remove the shadows and extract the vegetation information. According to the concept of the optimum index factor (OIF), Eq. Where si standard deviation of rationing image for order i, | R | is the absolute value of correlation coefficient for color order j. By random test we get the rationing image subset [TM/TM3,TM3/TM2,TM2/TM1] whose OIF value is the maximized and make false-color display for this rationing image subset (appropriate color red , green, blue )in which the mountain shadows were removed and the land cover information was extracted successfully in order to compare the effects between the false color composition of rationing image (TM4/TM3,red,TM3/TM2) green TM2/TM1blue and original image we also given the false color composition of original images ( TM5,red, TM3, Green, TM2 , blue) which are shown in color plate 4. A discussion on mixed image subset of false color composite for feature enhancement . Three image subsets are mentioned above i.e. subset 1 ( TM1,TM2,Tm3…TM7) formed by the original six base data of landsat TM images (expect the thermal infrared band 6) made up by principal components of K-L transformation and subset 3 composed of 30 independent rationing images generated from the same TM image data . we call it can mixed subset which is made up of one or two subsets above Therefore according to the principle for us to from the new subsets how should we form the mixed subsets in order to extract or enhance certain physical landscape information we follow we follow a principle that is called physical landscape in formation weighting and complement each other adopting such a principle we can effectively enhance any features we are interested in. We select Dawopu image window to discuss the effectiveness of mixed subset in feature enhancement comparing mixed subset image (K1,K2,T4) (color plate 5) with K-L transformation image (K1,K2,K3) (color plate 3) we can discover that the false color composite image formed by mixed subset has a higher ability in vegetation classification for example a cutover a kilometer north of Lujuanzi displayed as a unitary pattern in the standard false color image but the same area can be divided into three patterns of different colors in the image formed by (K1,K2, T4) it illustrates that the region can be further classified in vegetation for example according to 1:35000 color infrared airphotos, there is brush around Majiazi village a km south east to Lujuanzi but in the orthogonal transform image the pattern is divided in to yellow and light blue standing for buch and grass in the mixed subset image the same patter is cut into four separate parts by four color similarity the wide flood land near Shihu village in color plate 5 displays as a uniform yellow color but in color image of (T5,T4,K1) (color plate 6) it shows as four colored pattern which illustrate the difference of water condition of various land cover types if analyzed with great concentration variable can be obtained. An approach to external forest vegetation information from multi temporal NOAA-AVHRR image. Although it is obvious unrealistic for using above mentioned image processing methods to study global scale or continent sized forest vegetation it is very important to illustrate changes of global environment .People are forced to concern on finding a way to extract vegetation information from NOAA-AVHRR digital images .such vegetation cover types as force Brush grassland or meadow have their own life rhythm which can be reflected by vegetation index vegetation index mentioned here is defined as the ratio of observed values of the second channel to the first channel of AVHRR the ratio is a function of time which values reflects of vegetation and strength of photosynthesis moreover it's deeply influenced by background (siol) as a matter of fact VI is the index of vegetation landscape. The researched AVHRR image is a 512*512 pixel image window of Dalainor Inner Mongolia to the North West lie Great Xingan Mts to the south there stands mountain QiLaoTu with Silamuren river passes through and dalai nor lake in the center .Te time that the four images were received is the vegetation gouts season in 1989 and respectively the date is May 4 and June 8 July 2 and August 13 . The original images are strictly matched by mercator projection transform after computing the VI of four original images separately we co responsibility get four temporary VI images to separate the vegetation types from each other the four VI images are K-L transformed and the temporal dimension coordinate is completed. The result K-L transformation is shown on Table 5. Table 5. eigenvalues and eigen vectors of K-L transform for vegetation index Temporal number 1(5.04) 2,(6,08 ) 3(7,03) 4(8,13) Mean 189.98 232.27 154.07 153.85 Variance 25.029 33.752 77.434 77.507 Correlation matrix 1 626.48 211.07 1415.0 1428.5 2 211.07 1139.2 1056.4 1059.4 3 1415.0 1 1056.4 5996.0 5825.6 4 1428.5 1059.4 5825.6 6007.3 principal component 1 2 3 4 eigen values 0.899 0.068 0.020 0.013   1 -0.169 0.094 0.977 0.086 2 -0.133 -0.988 0.072 -0.007 3 -0.690 0.084 -0.189 -0.694 4 -0.691 0.083 -0.065 0.715 Table 5 shows that after the principal transformation the information of vegetation or vegetation index basically gather to the first principal component the first principal component image has greatly reveled the difference between various types of vegetation which can be effectively distinguished by the image segment technology in processing the pixels with the same characteristics get the incontinently the pixel on the region should be similar to each other and some variable characteristics of the pixels vary from one region to another therefore the edged can be made out. The segment to some honogeneouse attribute PK of a two dimensional image point matrix X(1,J) is to divide X into some non subsets X11,X2,X3…,Xk. They meet the conditions follows: Eq. The method to give a threshold is used in image segment the value of thres hold operator Tk determined according to the histogram map : Eq. In order to draw the outline of the edges horizontal and vertical direction should be examed at the same time. Eq. The first principal image segment is completed by the new developed function SEGMT thresholds divide the image DN value region in to 9 parts the computer scans and searches one by one and in the meantime prepared color values are given by this way we successfully get color images according to threshold segment checked with the vegetation map of Chifen inner Mongolia the types of land cover which the various colors in the segment image stand for are shown on table 6. Table 6. Image segmentation for the first principle component of multi-temporal vegetation index VI. D N 0-32 32-64 64-96 96-128 128-160 160-192 192-224 224-254 254-255 Color dark green apple green blue green pea green light yellow brown yellow dark yellow red magenta Vegetation cover-type forest or shrub alpine grass land meadow grass land deyenerate grass land crop land bare land Segment of single spectral image is equal to the classification of multi spectral image . Especially for forest and grass land region with background more unitary and objects largely continuous the segmentation result of the principal component image has obvious classification significances segment image are the authentic accords of the distribution of forest or brush in the section of great Xinggan Mts. and Mountain QiLaoTu. References Nesdis, NASA,global vegetation index user's hand book, aug., 1983. Tucker,c.j.et al., African land-cover classification using satellite data , SCIENCE, vol, 227, No. 4685, 1985. MOik,J.G., Digital processing of remotely sensed images ( NASA SP; 431) 1980. Simonett, D.S., et al., Manual of remote sensing and ed. Virginia, vol. I ,1983.