Transcript Slide 1

SUB-PIXEL ANALYSIS OF TREE COVER ON SATELLITE IMAGERY
USING A CONTINUOUS FIELD APPROACH
Eva Pantaleoni¥, Randolph Wynne‡, , and John Galbraith*
¥IAgER, Tennessee State University
‡Forestry Department, Virginia Tech
*Crop and Soil Environmental Science, Virginia Tech
1) Using the Advanced Spaceborne Thermal Emission and Reflection
Radiometer (ASTER) sensor for determining the correlation between
pixel values and the proportion of tree cover in wetlands.
2) Determining canopy cover over extremely small plots located in the
Coastal Plain of Virginia.
We randomly selected 300 plots over the study area. Each 225
m2 plot had the area equal to one ASTER pixel in the visible and near
infrared range of the spectrum. We calculated the continuous canopy
cover field using ArcGIS® after manually digitizing the tree or shrub
outline on leaf-off digital orthophotographs obtained from the Virginia
Base Mapping program (VBMP). We used aerial photographs from the
National Agriculture Imagery Program (NAIP) as additional reference.
The resolution of NAIP was 1 m, and VBMP was 1:4800 scale and
1:2400 scale. We used the first three bands of two ASTER scenes, from
March and October 2005, and data obtained from the tasseled cap
transformation and the delta normalized difference vegetation index
calculated from the ASTER scenes. NDVI = (Near Infrared-Red
bands)/(Near infrared + red bands). Delta NDVI = (NDVI from October
– NDVI from March). We then generated a point feature for each pixel
and assigned the canopy cover value to it. We calculated a correlation
matrix to determine which variables were significantly correlated with
canopy cover, and we examined the variable inflation factor (VIF) to
determine if the model was affected by multicollinearity (Table 1). A
VIF higher than 10 is considered evidence of multicollinearity (Belsley
et al. 1980). We performed a best subset regression using Mallows Cp
criterion, and we determined which variable the model required (Table
2). The model is satisfactory when the Mallow Cp is approximately
equal to the number of independent variables used in the model.
Table 1: correlation matrix and VIF for all the variables
(a)
(b)
(c)
Dependent
Variable
Independent
Variable
Corr.
Value
p-Value
VIF
AREA
October Band 1
-0.036
0.533
---------Table 3: results for canopy cover < 15%
AREA
October Band 2
-0.360
< .0001
2.79
AREA
October Band 3
-0.153
0.007
8.54
AREA
March Band 1
0.193
< .0001
137.30
AREA
March Band 2
0.001
0.983
----------
AREA
March Band 3
-0.109
0.052
15.91
AREA
DNDVI
< .0001
1.95
0.371
AREA
March Brightness
-0.211
< .0001
5322.49
AREA
March Greenness
-0.235
< .0001
2269.00
AREA
March Wetness
0.254
< .0001
8886.99
AREA
October Brightness
-0.263
< .0001
7.46
AREA
October Greenness
AREA
October Wetness
0.730
----------
0.092
0.109
----------
80
70
60
50
40
30
20
10
0
0
20
40
60
80
100
Observed value of canopy cover (%)
Figure 2: Observed vs. predicted canopy cover, one group
80
70
60
50
40
30
20
10
0
0
20
40
60
80
100
Observed value of canopy cover (%)
Figure 3:
Observed vs. predicted canopy cover, two groups
Table 4: results for canopy cover > 16%
Variables
Parameter
estimate
VIF
Variables
Parameter
estimate
VIF
Intercept
81.71
0.00
Intercept
151.75
0.00
OB2
-91.39
1.57
OB2
-119.07
1.23
MB3
51.04
1.31
MB3
-141.26
1.19
2.21
1.37
DNDVI
1.64
1.09
0.04
0.03
DNDVI
-0.020
90
Predicted value of canopy cover (%)
Objectives:
Methodology:
Our results show that our
model requires only three variables:
Delta NDVI, October Band 2, and
March Band 3. Examining the
predicted values of canopy cover
against the observed values of
canopy cover (Fig. 2), the model
appears to be affected by a saturation
effect. In order to address this effect,
we separated the group of points into
two parts, and we ran two separate
models using the same variables
selected by the best subset
regression. The first group had
canopy cover between zero and 15%,
and the second group had a canopy
cover between 16% and 100%. Table
3 shows that the first group has an
adjusted-R2 of 69.0% and an R2 of
72.0%, with RMSE = 2.7% and VIF
values lower than 10. The VIF values
are lower than 10 also for the second
group, but the RMSE value is higher
(13.7%), the adjusted-R2 is 3.0% and
the R2 is 4.0% (Table 4). Fig. 3
shows the plot of the predicted
values canopy cover and the
observed values for the two groups.
At canopies above 15%, the pixel
signatures became saturated with
color and it was not possible to
accurately differentiate canopy
cover%.
Predicted value of canopy cover (%)
Introduction:
Wetlands are important and complex ecosystems that provide a wide
range of services vital to the environment. Wetlands control water storage
and indirectly runoff, slowing the velocity of water flow and trapping
sediments and nutrients, preserving the quality of the water (Mitsch and
Gosselink, 1993). Cooper et al. (1987) demonstrated that 85% to 90% of
the sediment transported by runoff events is trapped in wooded areas and
never reaches major streams. Even though there is still a debate concerning
relative effectiveness of grass and forested wetlands, it is evident that
forested
wetlands
do provide
resistance
to
sediment
transport.
Consequently, it is important to determine the density of trees within
forested wetlands. DeFries et al. (2000) generated the first global
continuous field product for tree cover, fitting a linear mixture model to a
classification output, using AVHRR and MODIS. A continuous field represents
the proportion of vegetation cover per pixel, and it is an improvement over
discrete land-cover classifications for some applications (Hansen and
DeFries, 2004). These data are extremely useful for global analysis and
regional studies (Franklin et al. 2000), they lose power at scale at which
most land management occurs.
Results:
R2
0.72
R2
Adjusted- R2
0.69
Adjusted- R2
RMSE
2.73
RMSE
19.79
Conclusions:
Table 2: Results from the best subset regression analysis
(e)
(d)
Variables
in model
R2
Adj-R2
Cp
RMSE
DNDVI
0.13
0.13
37.89
23.60
DNDVI, October Brightness
0.23
0.22
6.84
22.33
DNDVI, March Band 3, October Band 2
0.23
0.22
4.32
22.31
DNDVI, March Band 3, October Band 2,
October Brightness
0.24
0.23
4.34
22.27
DNDVI, March Band 3, October Band 2,
October Band 3, October Brightness
0.24
0.22
6.00
22.29
Study area:
Figure 1: study area
The study area is located in the Coastal Plain of Virginia. Representative species are Bald Cypress (Taxodium
distichum), Swamp Tupelo (Nyssa sylvatica var. biflora), Yellow Poplar (Liriodendrom tupilifera), Water Oak (Quercus
nigra), and Sweet Gum (Liquidamber styraciflua). Fig. 1 shows: (a) map of the east coast of the U.S.A. with Virginia in
black; (b) county map of Virginia with boundaries of ASTER scene in red; (c) ASTER granule covering a 60 x 60-km area;
(d) example of digital orthophotos from VBMP; (e) example of aerial photos from National Agriculture Imagery Program.
Contact info: [email protected]
In this study, we found a relationship between canopy cover and the spectral characteristics of
the VNIR ASTER bands and derived indices. The literature had already highlighted DNDVI as a strong
indicator of vegetation characteristics. In our study, this measure was selected as one of the
important variables by the model selection criteria, and it also had the highest correlation with
canopy cover.
After we addressed the multicollinearity problem, and we reduced the model to its simplest
form, we found that DNDVI, the red band from October (OB2) and the NIR band (MB3) from the
March scene were the most important variables. The combination of these three variables produces
good results in plots that have a canopy cover lower than 16%. When canopy cover is higher than
15%, there appears to be no relationship between the ASTER-derived variables and canopy cover.
This result is similar, albeit at a lower threshold, to the off-observed saturation effect between
vegetation indices such as NDVI and leaf area index (Wang et al., 2005).
Our conclusion is that ASTER imagery has potential for use in estimating canopy cover within
forested and scrub-shrub wetlands, even though further research will be required to address the
water absorption and saturation effect problems.
References:
COOPER, J.R., GILLIAM, J.W., DANIELS, R.B., and ROBARGE, W.P., 1987, Riparian areas as filters for agricultural sediment. Proceedings, Soil
Science Society of America, 51, 416-420.
DEFRIES, RS, HANSEN, M.C., TOWNSHEND, J.R.G., 2000, Global continuous fields of vegetation characteristics: a linear mixture model applied to
multi-year 8 km AVHRR data. International Journal of Remote Sensing, 21, 1389–414.
FRANKLIN, J., WOODCOCK, C.E., AND WARBINGTON, R., 2000, Multi attribute Vegetation Maps of Forest Service Lands in California Supporting
Resource Management Decisions. Photogrammetric Engineering & Remote Sensing, 66, 1209-1217.
HANSEN, M.C., AND DEFRIES, R.S., 2004, Detecting Long-term Global Forest Change Using Continuous Fields of Tree-Cover Maps from 8-km
Advanced Very High Resolution Radiometer (AVHRR) Data for the Years 1982-99. Ecosystems, 7, 695-716.
MITSCH, W.J., and GOSSELINK, J.G., 1993, Wetlands. 2nd ed. (New York, NY: Van Nostrand Reinhold).
WANG, Y., and SUN, D., 2005, The ASTER tasseled cap interactive transformation using Gramm-Schmidt method. MIPPR 2005: SAR and
Multispectral Image processing, In Proceedings of the SPIE, 6043.