14.3 Vegetation Change Tracker Algorithm
14.3.2 Individual Image Masking and Normalization
In this step, each image is analyzed individually to create initial masks for water, cloud, and shadow and to normalize the image using known forest samples� This step has the following major processes: creation of a land–water mask, identification of forest samples, calculation of forest indices, and masking of cloud and cloud shadow�
14.3.2.1 Land–Water Masking
Based on the known spectral properties of typical water bodies (Jensen 1996), a pixel is flagged as a water pixel if it has a low reflectance value in the shortwave infrared band (band 5) and satisfies at least one of the following two conditions:
1� It has a decreasing trend of reflectance values from the visible to the infrared bands�
2� It has a low normalized difference vegetation index (NDVI) value, where NDVI is calculated using the reflectance value of the red (Rred) and near-infrared (RNIR) bands:
NDVI NIR red
NIR red
= −
+
R R
R R (14�3)
14.3.2.2 Identification of Forest Samples
Although the LTSS images have been corrected to achieve high levels of radiometric integ- rity, VCT uses forest samples to further normalize image radiometry and to calculate
Masking and l
Masks
Indices
LTSS 12
108
IFZ 6
42
085 87 89 91 Year (19xx)
Time-series analysis
Disturbance products
93 95 97 99 Masking and
Masking and Masking and Masking and normalization
FIgure 14.3
Overall data flow and processes of the vegetation change tracker algorithm� (From Huang, C� et al� Remote Sens Environ, 114, 1, 2010� With permission�)
forest likelihood measures� Such forest samples are identified based on the known spec- tral properties of forest� Specifically, dense, mature forests typically appear dark and green in a true color composite imagery and are among the most easily distinguishable features in remote sensing imagery (Dodge and Bryant 1976)� As such, some of them can be identified reliably using histograms created from local image windows (e�g�, 5 × 5 km)� Because forest pixels are typically the darkest vegetated pixels, they are gener- ally located toward the lower end of each histogram� When a local image window has a significant portion of forest pixels, those pixels form a peak called a “forest peak” in the histogram� In the absence of water, dark soil, and other dark nonvegetated surfaces, which are masked out using appropriately defined NDVI and brightness threshold val- ues, forest pixels are delineated using threshold values defined by the forest peak� Huang et al� (2008) described this approach in detail in their study�
14.3.2.3 Calculation of Forest Indices
The identified forest samples are used to calculate a number of indices that are indicative of the likelihood of each pixel being a forest pixel� Suppose the mean and standard devia- tion of the band i spectral values of forest samples within an image are bi and SDi, respec- tively, then, for any pixel with a band i value of bi, a forest z-score (FZi) value for the band can be calculated as follows:
FZi SD
i i
= −b b (14�4)
For multispectral satellite images, an integrated forest z-score (IFZ) value for that pixel is defined by integrating FZi over the spectral bands, as follows:
IFZ= NBNB FZ
∑=
1 2
1
( i)
l
(14�5) where NB is the number of bands used� For Landsat TM and ETM+ images, bands 3, 5, and 7 are used to calculate the IFZ� Bands 1 and 2 are not used because they are highly correlated with band 3� The near-infrared band is not included in the IFZ calculation because (1) it is less sensitive to logging and other nonfire disturbances than the other spectral bands and (2) spectral changes in this band do not always correlate with distur- bance events�
A major problem with using SDi calculated from forest samples within each individual image is that the value can vary greatly as a function of the forest type composition in that Landsat image� The SDi calculated this way will be low for images consisting of forest pixels that are spectrally similar but can be very high for images consisting of both open canopy forests with bright backgrounds and closed canopy forests� Such a dependency of the SDi and hence the IFZ on the composition of forest types within each image makes it difficult to develop generic change detection algorithms for use over a wide range of for- est biomes� To mitigate this problem, the average standard deviation values derived using images acquired in different years from different places of the United States are used in Equations 14�4 and 14�5 (Table 14�2)� The FZi and IFZ indices calculated using Equations 14�4 and 14�5 have a number of appealing properties:
IFZ is an inverse measure of the likelihood of a pixel being a forest pixel� Pixels
•
having a low IFZ value near 0 are close to the spectral center of forest samples, while those having high IFZ values are likely nonforest pixels (Figure 14�4)�
Assuming forest pixels have a normal distribution in the spectral space, FZ
• i
could be directly related to the probability of a pixel being a forest pixel using the standardized normal distribution table (SDST; Davis 1986)� As the root mean square of FZi, IFZ can be interpreted similarly� Specifically, over 99% of forest pixels likely have IFZ values less than 3� Although in reality forests may not have a rigorous normal distribution, and the standard deviation values used here are not calculated from the image of interest, such an approximate prob- ability interpretation makes it possible to define probability-based threshold values that might be applicable to images acquired on different dates over dif- ferent locations�
While deciduous and coniferous forests often have different spectral characte-
•
ristics, during the growing season they have similar IFZ values that are substan- tially more stable over time and are mostly lower than those of nonforest land-cover types (Figure 14�4)� This observation makes it possible to detect forest changes using the IFZ index without knowing the forest type, although the differences between the IFZ values of different forest types can be greater than those shown in Figure 14�4 (see also Section 14�3�3�2)�
In addition to the FZi and IFZ indices, the VCT also calculates a normalized burn ratio index (NBRI):
NBRI NIR
NIR
= −
+
R R
R R
7 7
(14�6)
1985
0 3
IFZ 69 IFZ
12
15 24
21 18 15 12 9 6 3 0
1987 1989 1991 1993
Year 1995 1997 1999 2001 1984 1986 1988 1990 1992 Year1994 1996 1998 2000 2002 2004 2006
Urban Deciduous 1 Conifer 1
Cropland
Conifer 2
Wetland Water
Deciduous 2
Urban Deciduous 1 Conifer 1
Cropland 1
Conifer 2
Cropland 2 Water
Deciduous 2
FIgure 14.4
IFZ values of different land-cover types in eastern Virginia (left, WRS path 15/row 34) and Oregon (right, WRS path 45/row 29) show that deciduous and conifer forests have IFZ values that are generally below 3 and are dif- ferent from those of other land-cover types (except for some water)� (From Huang, C� et al� Remote Sens Environ, 114, 1, 2010� With permission�)
where RNIR and R7 are reflectance in the near-infrared band (band 4) and band 7, respec- tively� NBRI is correlated with field-measured burn severity indices (Chen et al� 2008;
Escuin, Navarro, and Fernandez 2008) and is used to improve the detection of fire distur- bance events using VCT�
14.3.2.4 Cloud and Shadow Masking
Although the major goal of image selection in LTSS development is to minimize cloud con- tamination due to frequent cloudy conditions in many areas, some LTSS images inevitably contain cloudy pixels� Cloudy pixels generally have high brightness values and low green- ness values� If unflagged, most likely they will be mapped as nonforest regardless of the actual surface conditions beneath the clouds� For forest change analysis, unflagged clouds over forests likely will be mapped as forest disturbance� Cloud shadow over forests may also be mapped as disturbance, because as the spectral signature of forests under shadow can be quite different from that of sunlit forests� The cloud masking algorithm used in the VCT is based on the observation that clouds generally appear bright in reflective bands and cold in thermal bands and can be separated from cloud-free observations using threshold values defined by a set of linear boundaries in a spectral-temperature space� Once a cloud patch is flagged, its height is calculated using its temperature and a normal lapse rate (Smithson, Addison, and Atkinson 2008)� The shadow location of the cloud is then pre- dicted according to solar illumination geometry and the calculated cloud height; the dark pixels at or near the predicted shadow location are flagged as actual shadow� Details on this cloud and shadow algorithm have been provided by Huang et al� (in press)�