Vegetation Spectral Feature Extraction Model
Qian Tan, Hui Lin
Dept. of Geography & Joint Lab. For Geoinformation Science ,
The Chinese University of Hong Kong , Hong Kong
Tel: (+852)-26098105
E-mail: tanqian@cuhk.edu.hk
Yongchao Zhao, Tong Qingxi, Zhen Lanfeng
Lab. Of Remote sensing Information Sciences ,
Institute of Remote Sensing Applications ,CAS ,
Beijing , 100101 , China
Keywords:vegetation, hyperspectral, spectral feature extraction
Abstract
A new spectral feature selection and extraction model(for vegetation only!)--- Vegetation Spectral Feature Extraction Model (VSFEM) is presented . A lot of vegetation field spectrum analyzed , 8 vegetation spectral feature positions are acquired , through which a series of feature parameters are achieved .
1. Introduction
Spectral feature extraction were mostly from or for target classification .
Principle Component Analysis (PCA) is used by most people . This method produces a new series of images , put in order by information content (or variance ) . Relationships among images are essentially removed . With forward principle components , most information content can be seen , which is the optimal result for minimum mean square error . Green (1988) developed PCA , who applied MNF (Minimum Noise Fraction) so as to make every component after MNF transform in order by signal-to-noise ratio (S/N) from large to small , instead of variance . Jia (1999) developed PCA to segment PCA to make feature extraction , whose classification and display result has made some progress . Though PCA can compress and extract information with minimum mean square error , information of principle component images is hard to explain directly according to spectrum , moreover , sample distribution not considered , it's uncertain to get optimal classification result . After realizing this problem , Fukunaga (1970) proposed a new transform method , that was , to find an transform matrix , who satisfied the formula , T(S
1+S
2)T
-T=I , in which Si is the correlation matrix of class i . This method is effective when differences of average values are small and covariance play a key role , but for common situation , is ineffective (Foley , 1975) . Kazakos (1978) put forward a feature extraction algorithm ---linear scalar extraction , making minimum classification error probability for two classes of multi-dimension normal distribution . This method can find an optimal vector to get minimum classification error probability , but when more than one feature is required , this method has no power . After considering within-class and between-class distance , Richard (1986 ) posed a feature extraction method---Canonical Analysis (CA) , Lee (1993) raised a feature extraction method based on decision boundary .
This paper puts forward VSFEM(Vegetation Spectral Feature Extraction Model) , which is very different from above methods . VSFEM aims at vegetation spectral feature extraction , from or for controlling vegetation spectral curves , through an amount of analysis of field vegetation spectral curves to get some regularity . Compared to above methods , VSFEM pay much attention to target spectral reflection of biological and physical features , not thinking feature extraction just as pattern recognition or information compression of a branch of mathmatics .
2. Study Area and Data Collection
Study Area
A study area near Changzhou city, Jiangsu Province, China was chosen for collecting ground data . The study focused on vegetation , which included principle types of agricultural crops and trees in the study area , such as , rice , maize , peanut , sweet potato , cotton ,soybean , cabbage , carrot , etc .
Data Collection
During the period from late August(middle season in vegetation growing circle) to mid-October(later season in vegetation growing circle), 55 field vegetation spectral data were acquired from 20 sites in Changzhou by SE-590 ---a portable field spectroradiometer . At the same time , some biochemical parameters , such as chlorophll concentration , leaf area index(LAI) were measured . Data were collected in nadir orientation of the radiometer and at about 45° solar zenith angle. Four scans at a time were averaged as the final spectra in each measurement. In addition , the data were collected from 11:00 to 13:00 .
When trees were measured , branches with leaves were picked down and laid on the ground. The specific parameters of SE-590 Portable Field Spectroradiometer are shown as follows:
Wavelength: 400 - 1100 nm
Spectral resolution : 4.0 nm
Sample channels: 252
Field of view:150
3. Methodology
3.1 Vegetation Spectral Feature Extraction Model
There are some special features, such as "green peak", "red valley" and "NIR platform", in the curve of the reflectance, reflectance intensities of these featured positions vary remarkably or regularly with the species or growth periods. So, it is possible that we design special parameters that are good tokens of curve shape of different species or growth stages. Moreover , if we want to discuss correlation between spectra and vegetation biochemical properties, we also need to find some special spectral parameters. For this case, we define eight special positions(feature position) and design many parameters(feature parameter) like NDVI to discuss the species and property(including the growth stage) difference of typical vegetation in Changzhou. All the eight feature positions as M, B, G, Y, R, V, I1 and I, and some feature parameters are shown in fig.1 . This figure shows a typical spectral curve R(
l) .The definition of 8 feature positions as shown in fig.1 and their agorithms are as follows:
- Absorption peak in purple-blue band-M(lM,RM): The position where the reflectance is the minimum in the wavelength range of <500nm:
RM=MIN(R(lĪ380-500nm)),lM,li(R(li)=RM)
-
Absorption edge of blue waveband (blue edge)-B(lB,RB): the turning point of the spectral curve in the range of 500-550nm, defined as the maximum point in the first-order derivative value in the same waveband region:
lB=li(R'(l)=MAX(R'(lĪ450-550nm)),RB=R(lB)
-
Reflectance peak of green band(green peak)-G(lG,RG): the maximum position of reflectance band from 500-600 as:
RG=MAX(R(lĪ500-600nm)),lG=li(R(li)=RG)
-
Absorption edge of yellow waveband(yellow edge)-Y(lY,RY): the turning point of the spectral curve in the range of 550-650nm, defined as the minimum point in the first-order derivative in this range:
lY=li(R'(l)=MIN(R'(lĪ550-650nm)),RY=R(lY)
-
Absorption peak in red band(red "valley")-R(lR,RR): where the reflectance is the minimum in the red band of 600-720nm:
RR=MIN(R(lĪ600-720nm)),lR=li(R(li)=RR)
-
Red edge-V(lV,RV): the turning point of reflectance curve within waveband of red-NIR, the maximum point of the first-order derivative spectral curve in 670-780nm:
lV=li(R ' (l)=MAX(R '(lĪ670-780nm)) , RV=R(lV)
-
Start site of the NIR platform-I1(lI1 , RI1): the transition point between the red slope in wavelength >760nm and the NIR platform. It is also defined as the first joining point of the spectral curve and its continuum curve in range of 670-800nm as shown in fig 5.4.A. Its arithmetic is:
lI1=li(Rcr(lĪ670-800nm)) , RI1=R(lI1)
-
Maximum point of reflectance in NIR of 780-950-I(lI , RI):
RI=MAX(R(lĪ780-950nm)) , lI=li(R(li)=RI)
In order to get the green peak G and the red "valley" R, they are also defined as the zero points of the first-order derivative spectra R'(
l) in the range of 500-600 and 600-720nm, respectively. Thus we get G'(
lG' , R
G') and R'(
lR' , R
R') as:
lG'=
l(R'(
lĪ500-600nm)=0)
lR'=
l(R'(
lĪ600-720nm)=0)
Fig 1. sketch of 8 feature positions and some feature parameters for the green vegetation in Changzhou
The reflectance is about Rice measured in Aug. 31 at Changzhou. The parameters and positions labeled in this sketch are defined in the text.
As shown in fig.1, the eight special positions determine, on the whole, shape and spectral feature of reflectance spectra of vegetation in visible-near infrared band. It is distinct that the multi-line MBGYR determines the feature of green peak while the multi-line GYRVI
1 determines the general shape of red absorption peak. Line I
1I can be looked upon as the representation of NIR platform. As shown in Table 1 , these 8 positions almost keep constant with outer changes while their corresponding reflectivity intensities vary greatly .Thus there is some possibility that we can use a variety of these 8 special positions and their relations to represent spectral change of different vegetation.
In order to figure out the correlative variety of these special position and therefore to show the changing rules of reflectance spectra with vegetation species, we designed some parameters(feature parameters) on the base of the 8 special positions according to the spectral features in fig 1. They are:
- The coordinate of 8 feature position M, B, G, Y, R, V, I1, I and two accessorial positions as G' and R': (lP , RP). where the subscript P is the name of these positions, l is the wavelength and R is the reflectance. Obviously, there have 20 such parameters and in general we have G' » G and R' » R.
-
Slope of blue edge-SB: the slope of line MG.
SB=(RG-RM)/(lG-lM)
Thus it approximately determines the curve of MBG.
-
Slope of yellow edge-SY: the slope of line RG. It approximately determines the character of curve GYR:
SB=(RG-RR)/(lG-lR)
-
Slope of the incline among bands of red-NIR-SV:: the slope of line RI1 , it is a representation of curve RVI1:
SV=(RI1-RR)/(lI1-lR)
-
Slope of the continuum-SC: the slope of line GI1. It generally reflects the background feature of the absorption peak GYRVI1:
SC=(RG-RI1)/(lG-lI1)
-
Net height of green peak-HG: the distance between G and line MR in the dimension of reflectance. It generally equals to the net reflectance of the background-removed green peak and is the reflection of reflectance peak MBGYR:
HG=RG-((RR-RM)/(lR-lM)×(lG-lR)+RR)
-
Net depth of red absorption "valley"-HR: the distance between R and line GI1 in the dimension of reflectance. It's can be looked upon as background-removed depth of red absorption peak and reflects the feature of peak GYRVI1:
HR=(RG-RI1)/(lG-lI1)×(lR-lG)+RG-RR
