Vegetation Spectral Feature Extraction Model

Printer Friendly Format

Page 2 of 3
| Previous | Next |

Vegetation Spectral Feature Extraction Model

Net height of infrared platform-H_I: the difference between the averaged reflectance of NIR platform and the reflectance of point R. It can be substituted by the reflectance difference of R and the midpoint of line I₁I and reflects the feature of platform I₁I:
H_I=(å((Rⁱ+R_i+1)×(l_i+1-l_i)/2))/Dl-R_R»(R_I1+R_I)/2-R_R,
where l_iÎl_I1-930nm, Dl is the width of NIR=930-l_I1.
FWFH of green peak-l_wG: generally represents the width of green peak. It is approximately calculated as the horizontal interval between points B and Y:
l_wG »l_Y-l_B
FWFH of red absorption peak-l_wR: the reflection of the width of red "valley". It is calculated as the half width of continuum-removed red "valley" and can be replaced as the horizontal interval of line YV:
l_wR»l_V-l_B
Averaged reflectance of NIR platform-R_Ia: can be substituted by the reflectance average of I₁ and I:
R_Ia=(å((R_i+R_i+1)×(l_i+1-l_i)/2))/Dl»(R_I1+R_I)/2=H_I+R_R
where l_iÎl_I1-930nm, Dl is the width as 930-l_I1.
Area of green peak-A_G: the integrated area under curve MBGYR. It is the embodiment of green peak intensity and can be approximately substituted by the area under multiline MBGYR:
A_G=(å((R_i+R_i+1)×(_li+1-l_i)/2))
»[(å(R_p+R_p+1)×(l_p+1-l_p)/2)),p=M,B,G,Y]
where l_iÎl_M-l_R.
Pure area of green peak-A_G': the integrated area enveloped by curve MBGYR and line MR. It is the net intensity of green peak and obviously can be replaced by the area of polygon MBGYRM:
A_G'=A_G-((R_M+R_R)×(l_R-l_M)/2)
»[(å(R_p+R_p+1)×(l_p+1-l_p)/2)],p=M,B,G,Y]
-((R_M+R_R)×(l_R-l_M)/2)
where l_iÎl_M-l_R
Net area of red absorption peak-A_R: the area enclosed by curve GYRVI₁ and line GI₁. It can estimate as the area of polygon GYRVI1G:
A_R=S_GYRVI1G

Other parameters can also be gotten if necessary. For instance, after removing the continuum, the 5 most intensive peaks can be extracted from the continuum-removed(named asldi with i increasing with intensity decrease) and normalized continuum-removed results(lci).

It should be pointed out that the usually applied parameters in vegetation study such as NDVI, red edge l_re and red edge slope dr_re could also be obtained from these parameters as follows:

NDVI=(R_I1-R_R)/(R_I1+R_R) , l_re=l_V , dr_re»S_V

Therefore, the definition of eight feature positions not only reflects the general feature of the reflectance data, but also can get many high-information-content feature parameters that may have good relationships with some property parameters of vegetation such as chlorophyll concentration.

3.2 Analysis for Effectiveness of VSFEM---Relative Stability of Position
To study characteristics of the above feature positions and parameters , especially to study relationships between feature positions and vegetation types , as well as action of feature parameters for reflecting vegetation parameters . Based on above definitions and algorithms , more than 100 spectral data of about 20 types of vegetation in Chang Zhou were analyzed to get spectral feature positions and parameters.

Table 1 gives principle results , from which we can see that ,
under the research conditions of the experiment , all feature positions , especially 8 feature positions are stable , these positions are : M:404, B:525, G:556,Y:573,R:671,V:723,I1:758,I :900nm, take I position as example , which has the largest change range , for 62 samples at different time and place , its confidence width is 6.4nm , others less than 3nm , even for standard deviation , are less than 6nm generally . This kind of confidence interval has been super than spectral resolution of many instruments .

Table 1 feature position statistics result based on 62 samples

Feature position	samp les	aver age(nm)	Stand ard devia tion	Relat ive error(%)	95%confid ence(nm)	Range of 95%confidence(nm)	Width of 95%confi dence(nm)
l_B	62	524.7	1.04	0.20	0.3	524.5 - 525.0	0.5
l_G	62	556.2	3.67	0.66	0.9	555.3 - 557.1	1.8
l_I	62	900.7	12.77	1.42	3.2	897.5 - 903.8	6.4
l_I1	62	758.3	7.01	0.92	1.7	756.5 - 760.0	3.5
l_M	62	403.9	2.72	0.67	0.7	403.3 - 404.6	1.4
l_R	62	671.4	2.40	0.36	0.6	670.8 - 672.0	1.2
l_V	62	723.4	9.80	1.35	2.4	720.9 - 725.8	4.9
l_Y	59	573.2	0.81	0.14	0.2	572.9 - 573.4	0.4

3.3 Rediscussion for VSFEM---Reduction of Some Feature Positions
In addition , we calculate the correlation coefficients between different bands for all the spectral vegetation data measured by Se590 to show the independence of the bands and to select the most information-containing band group in order to indicate more efficiently the difference among different vegetation species , through which we can prove the effectiveness of VSFEM

where S is the correlation coefficient matrix of all bands, r_ij is the absolute value of correlation coefficient between band i and band j. Obviously:
r_ij=r_ji=|L_ij/SQRT(L_iix L_jj)|
where L_ij is the covariance between band i and band j.

Fig. 2 is the simulated image of the correlation coefficient matrix among 187 bands that is calculated on the base of 71 vegetation . Curves in Fig. 2 are isolines , from diagonal line to outside , the values are 0.9999,0.999,0.99,0.95,0.9,0.5,0.3 and 0.1 in the order .

From Fig. 2 we can see that ,
Except band 100(band number of Se590) to 126 , for vegetation , correlation coefficient are all very high (more than 99.99% generally) , that is to say, it's practical for band reduction .It shows two high-correlative platforms of 400-670nm and 760-950nm respectively, which means within these two regions , a few bands are enough to extract vegetation information . Moreover , around 550nm and 670nm , there are relatively wider areas . It's unnecessary to subdivide bands in these regions . but between 675nm---775nm and around 522nm and around 573nm , correlation coefficient are low generally . This shows more information here and bands should be subdivided . compared to 8 feature positions referred before , we can get that , for vegetation research , those regions that should be subdivided are just blue edge B , yellow edge Y and red edge V, while those regions that needn't be subdivided are blue absorption valley M , green peak G , red absorption valley R , and NIR platform I . In Fig. 2 , position B , Y and V are located in the center of narrow regions of isolines , while M , G , R , I in the center of broad regions . The arrows in Fig. 2 explain this clearly .

Page 2 of 3
| Previous | Next |