TABLE OF CONTENTS DISCLAIMER Illustrations Tables Equations INTRODUCTION  CHAPTER 1: IMAGE  PROCESSING METHODS 1.1    SCENE SELECTION 1.1.1 Season 1.1.2 Water Level 1.1.3 Cloud cover 1.1.4 Simultaneous coverage 1.1.5 Specifications 1.2 SCENE PREPARATION 1.2.1 Download raw imagery 1.2.2 Check header file 1.2.3 Visual check 1.2.4 Line replacement and destriping 1.3 ORTHORECTIFICATION OF IMAGERY 1.3.1 Coordinate System 1.3.2 Ground control in image 1.3.3 Ground control in field 1.3.4 Create model 1.3.5 Satellite image rectification 1.3.6 Accuracy check 1.4 RADIOMETRIC AND ATMOSPHERIC CORRECTIONS 1.4.1 Radiometric correction 1.4.2 Atmospheric correction 1.4.3 Aerosol correction 1.5 INDICES 1.5.1 Vegetation Index 1.5.2 Wetness Index 1.5.3 Temperature 1.5.4 Water Reflectance CHAPTER 2: IMAGE PROCESSING SOFTWARE PROCEDURES 2.0 INTRODUCTION 2.1 SCENE PREPARATION 2.2 ORTHORECTIFICATION OF IMAGERY 2.2.1 GCPWORKS 2.2.2 GCPREP 2.2.3 SMODEL 2.2.4 CIM2 - New file creation 2.2.5 GEOSET - georeference segment 2.2.6 SORTHO - satellite image orthorectification 2.2.7 Inter-scene accuracy check 2.3 RADIOMETRIC AND ATMOSPHERIC CORRECTION 2.3.1 TMRAD 2.3.2 RAYRAD - Rayleigh radiance 2.3.3 Dark object subtraction 2.4 INDICES 2.4.1 Bitmap creation 2.4.2 Vegetation Index 2.4.3 Wetness Index 2.4.4 Thermal band 6 temperature 2.4.5 Water reflectance 2.4.6 Brightness REFERENCES APPENDIX A: Image Pre-Processing Outline APPENDIX B: Sample Header File APPENDIX C: SMODEL Report APPENDIX D: Data Storage

Chapter 1
Image processing methods

1.1 SCENE SELECTION
1.1.1 Season

Figure 1. Big Bend study area on Florida's Gulf coast

Initial imagery selection includes determination of the appropriate dates and time frame for analysis. Leaf-off, or winter aerial photography, has traditionally been used in the determination of wetland habitats. However, satellite imagery does not have the same spectral and textural clues as aerial photography and must be employed in a different manner to identify land cover and change detection. Spring imagery is selected for the Gulf coast sub-tropical environment to maximize distinctions in vegetative health and quantity by providing maximum biomass during a period with a high probability of haze-free images. Biomass would not be adequately identified in winter imagery. The selection of imagery dates must be region and project-specific and should be advised by local ecologists familiar with the territory and the particular research issue.

Our primary project area is in the Big Bend region of the Florida Gulf coast (Figure 1). Two scenes are required to cover the area, and two seasons are required to fully classify the land cover. Winter imagery is used to augment the spring imagery for the distinction and classification of deciduous vegetation. A total of four coordinated scenes were acquired for the base years in the analysis (Table 1). Supplementary analysis in the time series requires contemporary spring scenes for each additional year.

1.1.2 Water Level

Table 1. Landsat TM images and Cedar Key water level at time of overpass
Landsat coordinates path/row	Location	Date	Cedar Key water level (msl)
18/39	Tallahassee, FL	01/23/85	0.40
18/39	Tallahassee, FL	04/25/86	1.05
18/39	Tallahassee, FL	11/29/93	0.71
18/39	Tallahassee, FL	04/09/95	1.46
17/40	Cedar Key, FL	01/16/85	1.18
17/40	Cedar Key, FL	04/25/86	1.29
17/40	Cedar Key, FL	01/12/95	1.26
17/40	Cedar Key, FL	04/02/95	0.54

Tide levels can adversely affect the determination of intertidal environments. It is preferable to select scenes with similar low water levels to optimize the characterization of low elevation intertidal zones. Each additional decimeter of water contributes to the mis-identification of 1-1.5% of the intertidal zone as documented by Jensen et al. (1993), and Raabe and Stumpf (1997). C-CAP recommends selecting tide levels within 0.3 - 0.6 m of mean low water (MLW), which maximizes the utility of the imagery (Dobson et al., 1995). The tide range in the study area is approximately 1 m from MLLW (Mean Lower Low Water) to MHHW (Mean Higher High Water). Although the majority of the intertidal environment along the Big Bend is high marsh, low water level at the time of satellite overpass enhances the discrimination of the low intertidal environment.

Due to a combination of factors such as cloud cover and scene availability, the imagery secured for the Big Bend does not entirely fit the C-CAP standards, but is the best available for the analysis. Some variability in water level is acceptable if two scenes per year are employed in the analysis. In the Big Bend we found that a complementary scene near MLW in the same year can facilitate the correction of differences caused by water level (Table 1). The 1986 south imagery exhibits high water level in both winter and spring, and consequently poses the greatest difficulty in identifying low elevation intertidal habitats.

1.1.3 Cloud cover

Cloud cover is the most limiting parameter in scene selection. Listings may be misleading, and previews of imagery should be made whenever possible to ensure quality data before purchase. Although it is advisable to select entirely cloud-free imagery, in some cases it may be necessary to establish only that the critical study area is cloud-free.

1.1.4 Simultaneous coverage

When a study area is large and/or encompasses more than one scene, simultaneous coverage of the whole area is warranted. It is possible, under ideal conditions, to secure several images within days or at most within 3-4 weeks that cover the entire area. In the instance where a study area falls between two scenes in the same path, it is possible to order an image that straddles the image boundaries, eliminating the need to mosaic scenes together.

1.1.5 Specifications
The CD-ROM Format

Imagery obtained from the EROS Data Center (EDC) and EOSAT is now provided in CD-ROM format. The procedures that follow begin with this format. Other formats may be accommodated with minor adjustments in the initial steps. All imagery for this project is obtained in Space Oblique Mercator (SOM) and with Nearest Neighbor resampling (NN). NN resampling is recommended for studies involving spectral analyses because it minimizes alterations to the count values. Other resampling such as cubic convolution may be preferred in studies involving pattern recognition.

1.2 SCENE PREPARATION
1.2.1 Download raw imagery

The contents of the CDROM will include 7 bands of TM along with summary and header files. All 7 bands of data are extracted from the CD and converted to a readable format for image processing. The header file is also extracted at this point.

The PCI procedures, CDEOSAT and CDNLAPS, read data off the CD-ROM, create a new file on disk, write the band data, and extract header information which is written into an orbital segment attached to the new file. The orbital segment will be used in the rectification of the imagery to the UTM coordinate system. Other coordinate systems may be employed in a similar manner.

1.2.2 Check header file

The header file should accompany the scene and a readable copy of the header file is made to disk. A sample header file may be found in Appendix B. Information from the header file is used in subsequent pre-processing steps. Trailer files may or may not accompany the imagery. Typically, the information desired from this file is the time of overpass in Greenwich Mean Time (GMT), which is noted for use in the radiometric enhancement.

1.2.3 Visual check

All new imagery is previewed, band-by- band, to determine if data lines are missing or other anomalous features exist in the scene. The visual review must be done at full-resolution. Attention is focused on identifying lines or blocks of missing data in each band for subsequent "repair." Haze or cloud cover that may render parts of the scene ineffective for analysis is noted as well. In our case the visual check is conducted in PCI Imageworks with seven image planes in the display window. Each portion of the image is displayed at full- resolution, and each band is viewed separately to identify missing data.

It is useful to examine the range and histogram of count values in each band at this time. Occasionally imagery is purchased with offset or anomalous count values. A quick evaluation now may save time later. Band 6, the thermal band, is especially prone to such errors.

1.2.4 Line replacement and destriping

Missing lines are "repaired," and other problems are resolved in each band before the scene is rectified to the UTM coordinate system. Once rectification has been conducted, errors will no longer coincide with horizontal lines of data and are virtually impossible to fix. Image line replacement is a simple procedure that allows the operator to fill-in missing lines with the line above, below, or with an average of the two.

It also may be necessary to conduct destriping of the image if a linear pattern is prominent in the image. It is usually a more significant problem with Landsat MSS than with TM or SPOT, and affects dark objects such as water in particular. However, we have run several variations of a destriping procedure on scenes with an 8 to 9-line repetitive stripe and have concluded that the output image was not improved over the original. Although we do not conduct this procedure ourselves, it may be advisable in other situations, and, if necessary, must be attempted in this initial phase of pre- processing, before rectification.

1.3 ORTHORECTIFICATION OF IMAGERY

Rectification of the imagery requires several steps: identification of ground control within the imagery, collection of ground control points with a GPS unit, development of a rectification model, reprojection of the imagery using the model, and two accuracy checks.

Three terms are defined:

1. Rectification fixes each feature in the imagery to the correct position on the earth

2. Reprojection involves transforming and rectifying the image to a standard projection such as Universal Transverse Mercator (UTM)

3. Registration involves having features in multiple scenes exactly match each other in location

However, in some projects, a previously rectified scene might be used as reference so that all scenes are registered to the base scene. For instance, a set of Landsat MSS scenes might be registered to a NALC (North American Land Characterization) scene. If the above or other procedures are used where a polynomial transform "rubber sheet" adjustment is employed, it is usually preferable to co-register scenes, then conduct the reprojection and rectification on all of them. Co-registration without rectification is never recommended.

1.3.1 Coordinate System

Rectification and reprojection of satellite imagery to a standard coordinate system is performed on all scenes in the project. Whether you use UTM or another coordinate system, the reprojection allows the determination of geographic coordinates for features identified in the analysis and facilitates integration with other geographic data sets. The approach used in the Big Bend project employs a PCI procedure, SORTHO, which reprojects the image based on satellite orbital information and a set of standard ground control points. The results effectively combine both inter-scene compatibility and coordinate plane rectification. To achieve co-registration without rubber sheeting, a common set of ground control points is used for all images. In this way the original data is rectified and reprojected from the Space Oblique Mercator (SOM) projection of the raw data to the Universal Transverse Mercator (UTM) coordinate system, with corresponding results in each subsequent image rectification.

1.3.2 Ground control in image

Ground control points are identified in the imagery as clearly visible point or right-angle positions, which are also accessible by road. Approximately twice as many points are identified as are ultimately required for the registration for several reasons. Some positions are not retrievable in the field, not all positions are identifiable in all imagery, and some positions are held out as part of the registration accuracy check. The operator examines a full-resolution image display to select the ground control. Preferred locations are corners at small to intermediate right-angle road intersections in rural or low-density developed areas. Highly developed areas tend to give blurred intersections and are difficult to re-locate in subsequent imagery. The identification of visible and distinct locations in the imagery which are also accessible to a field team is a major challenge in many regions. Alternatives to road intersections and right angles may be selected based on other landscape features. For example, the end of a bridge over a stream may be suitable. Consideration is given to the clarity of the location and the possibility of re-identification of the same point in earlier and later imagery. Good judgement and careful selection ensures accurate registration of the study area.

The set of selected positions are marked on the imagery file with a vector which can be displayed on a field laptop computer for identification during the field reconnaissance. If the technology is not available, a hardcopy is printed of each potential ground control point in the imagery, and these are bound with map sheets to aid in field identification.

1.3.3 Ground control in field

Accurate ground control points are collected for each image based on pre-identified locations as described in section 3.2. A hand- held Global Positioning System (GPS) unit is employed with accuracy guaranteed 5-10 m. Positioning of 10 m requires differential, P-code, or comparable receivers for accurate rectification of Landsat TM or SPOT imagery. Standard non-differential units do not produce the necessary sub-pixel accuracy.

Field plans should include 5-10 minutes per station for data collection and site documentation to facilitate identification of the positions in subsequent imagery. We recommend a minimum of 2-3 readings per position, and 24-30 positions per image. The ground control should be well-distributed across the image and at locations which are visible in imagery over a period of 5-10 years. As some GPS readings turn out to be poor quality, we recommend collecting multiple readings at each location. Accurate and complete field notes accompanying the collection of GPS ground control facilitate the identification of the positions in subsequent imagery. A description including county, road, and intersection names, distance from nearby intersections, directional designators, and photographs are helpful in future relocations. We have also concluded that the occupation of an intersection corner, appropriately designated (as in "southwest corner"), is better than occupying the center of a road intersection as has been previously practiced.

Other techniques exist for the collection of ground control points and may be applied where resources are limited. For instance, low-cost, non-differential , non-P-code GPS receivers can be used, providing positional accuracy of 30-100 m. Map accuracy in will be reduced from the 1:25,000 achieved in this project to 1:50,000 or more. Alternative approaches include digitization of positions from mylar 7.5 minute quadrangles, or the selection of similar positions from digital data sets including vectors, digital orthophotographs, and other imagery. Each approach carries with it inherent errors, of which the analyst should be aware. At the very least, the operator should know the error range for the input coordinates prior to the analysis. This will determine an acceptable RMS error range and whether or not the intended and final map accuracy will be achieved. The same caution applies to the selection of alternative adjustment techniques.

The PCI software in use for this project has a field version of GCPWORKS (under our investigation), which incorporates the selection of ground control in the imagery and the collection of GPS ground control simultaneously, on a field laptop computer. Although this does not eliminate the need to carefully select quality and well-distributed control within the image and in the field, it does consolidate the work, and allows the operator to evaluate the quality of the ground control as it is being collected. The on-site development of a rectification model eliminates costly returns to the field in the event some ground control proves useless.

The GPS unit employed in this project has a stated accuracy of < 10 m, which met our mapping needs with TM and MSS. Higher resolution imagery such as SPOT and IRS may benefit from even higher accuracy ground control, available with newer GPS units at 1-5 m. Although 24-30 positions are collected per image, not every position is used in the actual rectification. In addition to the elimination of poor quality positions, other ground control points are not easily relocated in the imagery and are not used. Eight to ten positions are used to rectify the imagery, and another 10-12 are set aside to be used in the accuracy check.

1.3.4 Create model

A model is created with the ground control points. It is preferable to use a low-order polynomial transform to reduce distortion in the final image, particularly at scene edges or over large water bodies. The PCI procedure we use makes the conversion between the two systems without introducing distortion to the resulting image as is normally found in polynomial adjustments. We prefer this approach as it gives consistent spatial positioning across the image. Other registration packages are available and may be applied with varying results depending on the quality of the ground control and the amount of distortion introduced in the polynomial function.

User-entered coordinates

Imagery is brought to display on the monitor for the collection and identification of ground control points. These ground control points are part of the original set selected from the imagery at full-resolution display and subsequently collected in the field. They are spatially well-distributed over the extent of the image and show promise of being stable and easily identifiable in a time series. The GPS coordinates collected at these locations show little variation in repeat readings.

During the identification of ground control, watch the RMS (root mean square) error for each ground control point and the total RMS error in both x and y. Examine the scatter plot of the ground control to help identify positions which are in error. Within reason, it is possible to adjust the location of the ground control in the imagery to achieve optimal positioning and distortion-free rectification.

We have found with 10 m accuracy in the GPS and 30 m resolution in the imagery that the RMS of each individual position and the total RMS need not exceed 0.5 pixels (15 m). At the same time, consider that excessive adjustment of point positions to produce significantly lower RMS readings will introduce errors in the final adjustment. An RMS of 0.5 pixels indicates that the position is + 15 m or 1/2 pixel from the optimal location. Under the circumstances, it is reasonable to expect a total RMS of 0.3-0.6 pixels, or 10 - 20 m, best fits the accuracy achievable with the given data. We have found that substantially forcing the model to lower errors introduces regional distortion into the final image.

Model

The collection of ground control points in the imagery leads to the development of a model to adjust the full image to the new coordinate system. See Chapter 2 for detailed steps in PCI. If not running GCPWORKS, and SMODEL, follow appropriate steps for software application.

1.3.5 Satellite image rectification

Table 2. Scene size and coordinates for georeferencing segment
Landsat TM Big Bend	Scene Location
File Information	Tallahassee	Cedar Key
Path/Row	18/39	17/40
Initial file size	6967 pixels/5965 lines	6967 pixels/5965 lines
Final file size (x/y)	7556 pixels/5412 lines	7916 pixels/7112 lines
Pixel size (x/y in meters)	28.5/28.5	28.5/28.5
Upper left coordinates (UTM)	98560 E, 3408400 N	210287 E, 3296110 N
Lower right coordinates (UTM)	313906 E, 3254158 N	435893 E, 3093418 N

The goal of image rectification is to facilitate the overlay of additional imagery and other geographic data sets. A standard map area, with boundaries set in UTM, is established for each scene, thus all image files for the same region, once rectified, will occupy the same map area. The UTM bounds for the scene are established according to the file size, the 28.5 x 28.5 m pixels, and the minimum/maximum northing and easting required to contain the full scene area. These boundaries, the UTM zone and the ellipsoid are established on each newly- created empty file.

The size and initial boundaries for the two scenes in the study (Figure 2 and Figure 3) are shown in Table 2 with UTM zone 17 coordinates for the upper left and lower right corners. The north image is cut at the Florida/Georgia border, approximately - 1100 lines (Figure 3). A new file of matching size is created for each image in the series. A matching georeference segment is established. Details are provided in Chapter 2.

In the event new regions are processed under these instructions, the designation of an output file size and geographic extent will depend on project needs, although it is recommended that full, or close to full scenes are pre-processed to maximize their subsequent utility. The calculation of file size is determined by the maximum and minimum eastings and northings of the raw image and the pixel size. Pixels in Landsat TM are 28.5 x 28.5 m, in MSS the pixel size is 57 x 57 m. Image rectification is conducted with the model segment created earlier, and the output applied to the newly-created and georeferenced image file discussed in the preceding paragraph.

1.3.6 Accuracy check

Accuracy checks involve both geographic or map accuracy, and co-registration or inter-scene accuracy. It is important to recognize that multiple scenes may meet one criteria and not the other. Two scenes may be co-registered to within a pixel, but have significant geographic error. Conversely, scenes may meet a standard map accuracy in a selected coordinate system, but have systematic errors that result in poor co- registration between scenes. Mis-registration may result from poor models or the use of high order polynomials. Evaluation of both map accuracy and image co-registration is necessary subsequent to image reprojection.

UTM or coordinate system accuracy

Image rectification is not complete without accuracy checks. If check points were not entered and evaluated as in GCPWORKS and SMODEL above, it is necessary at this point to objectively locate 8-12 check points in the rectified image. The evaluation will give a measure of the map accuracy, the percentage of the image within the standard ‘x' m. Obviously the evaluation varies according to the resolution of the imagery, the accuracy of the ground control, and the intended accuracy for the particular study. The measure obtained tells the user the expected positional accuracy of the majority of features in the image.

The check points are part of the larger set of ground control established with GPS as described in 1.3.2 and 1.3.3. Visually locate each ground control check point with the cursor in the rectified image. Note the image coordinates at the location, and evaluate how far the position is from the known x,y (UTM) coordinates. It may be difficult to conduct this step objectively. Try evaluating the portion of a pixel from which your visual position strays from the coordinates at which it visually should be. In our case, the goal was to accurately map 90% or more of the map features within 20 m of their known location, so we evaluate each position relative to our objective.

The accuracy check of the image to the UTM coordinate system meets our needs and is acceptable. Clearly, with only 12 check points, only one position per x/y component can exceed the established tolerance. It is worth conducting a careful check at this point to ensure good geographic correlation of data. If there is a geographic error in the image, it helps if you have indicated in which direction it shifted with a simple ± after each x/y offset.

Inter-image accuracy

The second accuracy check is conducted on imagery which will be used in inter-scene comparisons. It has been a common practice to rectify only one image, and then register all other images to the base scene, or to co-register all scenes to a master and apply the rectification for that scene to all others. While the standard approach provides a solution to multiple-image registration, every image in the present study is rectified individually to UTM coordinate system. According to this approach, each subsequent image is rectified using the same set of ground control points, and is visually checked against the original, or first-rectified base image as follows.

The inter-scene accuracy check consists of selecting 3 to 4 areas within the image to be displayed, one by one, at full resolution with selected bands from each image. Two methods are used to check inter-scene registration: (1) a simultaneous display of the same band from both scenes, using color as a reference to examine horizontal and vertical linear features, (2) a flicker state between the two scenes with the cursor at a fixed, well-defined location. Both methods may be applied regardless of the software in use.

In the first method single bands, usually band 4 or 3, from each scene are displayed simultaneously in a red-green-blue (RGB) display. Offsets will be obvious in either the x or y direction as bands of color on either side of a vertical or horizontal roadway (Plate 1). The first evaluation gives a quick check on overall alignment of the two images.

In the second approach the bands from the base image are displayed simultaneously with the same bands from the new image in a multiple-image plane display and set to flicker state. Three to five right-angle road intersections which display true horizontal and vertical direction within the image are evaluated in each of the selected full-resolution windows. Each intersection is evaluated and recorded in a table (see Chapter 2.2.7).

Plate 2 illustrates the position of the cursor at a right angle road intersection in the base image and evaluation of the cursor position relative to the intersection in the newly-rectified image. The cursor is placed at the center of the intersection in the base image, as if two imaginary lines were drawn N/S and E/W, taking into consideration the mixed pixel effect. The display is flickered to the newly-registered image, and the location of the intersection is compared in half pixel increments. Roads and intersections at non-right angles to the pixel/line orientation are never used in this evaluation. See Plate 3 for examples of intersections preferred for evaluation. The second method provides the operator with a detailed evaluation of inter-scene compatibility and the direction of offsets, if any.

Expected inter-scene registration is ± one pixel. A trend of greater than one pixel in x or y requires re-examination of the ground control point segment and a repeat of the whole rectification process. Consistent offsets or regional trends of greater than one pixel suggest poor quality registration and require re- rectification of the imagery with a thorough evaluation of the ground control point coordinates, positioning, and land cover changes between the imagery dates.

Again, good judgement, and a second opinion may help to eliminate operator errors or unforeseen problems. All imagery in the current project must meet the ± one pixel inter-scene accuracy before additional processing can be conducted. The goal is consistently met and presents no real problem with this particular set of imagery. We encourage other projects to develop similar methods to achieve high inter-scene positional registration.

1.4 RADIOMETRIC AND ATMOSPHERIC CORRECTIONS

Several band enhancements and corrections are applied to the rectified imagery to normalize the dn (digital number) values, facilitating direct spectral comparisons between imagery bands and a comparable set of values as input to indices and clustering programs. Radiometric calibration, conversion to reflectance or solar correction, and atmospheric correction are conducted on every image. The adjustments rely heavily on data contained within the header file, including gain, bias, and solar zenith angle. Programs have been written in-house to facilitate the first two adjustments. The atmospheric adjustment still relies on a visual/manual assessment of dark water values, although it may also be automated eventually.

1.4.1 Radiometric correction

The digital counts in the image are transformed to reflectance using the calibration that comes with the files and the equations and constants of Price (1987), and Markham and Barker (1985):

EQUATION 1. REFLECTANCE CALIBRATION

where is the band; the radiance, L, is determined by:

E_o is the solar constant (Price, 1987; Markham and Barker, 1985), r is the normalized earth-sun distance,   is the solar zenith angle at the image center, N is the digital count, G is the calibration slope, and BIAS is the calibration offset for zero radiance. A scale factor of 500 is applied to bands 1, 2, 3, 4, 5, and 7 to convert to 1 byte per pixel (0-255). A scale factor of 100 is applied to band 6. More sophisticated computations for reflectance exist. These involve additional terms in equation 1 for atmospheric transmission, and are being considered for implementation.

1.4.2 Atmospheric correction

The atmosphere introduces two forms of path radiance into the signal, radiance from Rayleigh or molecular scatter, and radiance from aerosols or haze. These can be removed simultaneously using dark object subtractions. However, improvements in atmospheric correction offers advantages in treating atmospheric and Rayleigh corrections separately. If no dark water exists in the scene, the Rayleigh correction is critical.

Rayleigh radiance is removed before the dark object subtraction. While not critical, when dark water is present, removal of the Rayleigh path radiance permits pixel by pixel correction of aerosols for water pixels and allows better control on the adjustment for aerosols. The Rayleigh term is determined using standard equations and coefficients. Models such as LOWTRAN (Air Force) can be used for the solution.

1.4.3 Aerosol correction

The aerosol correction is performed using subtraction from bands 1-4. Because the Big Bend region has black-water lakes and rivers, water can be found that has negligible reflectance in all bands. A dark-object subtraction is used, with the reflectance of the darkest water being the value subtracted. The correction is either constant or decreases slightly with wavelength (Chavez, 1989). The dn value selected should be that corresponding to the lowest value that has a significant number of pixels. In bands 3 and 4 the dark water area should be the same region. Identifying dark water in the blue and green bands may be difficult. If a suitable area is not present, extrapolation from bands 3 and 4 may be necessary.

8	8	9	8	9
8	9	9	9	9
9	9	9	8	9
9	10	9	13	10
8	5	8	7	8

Normalization of marsh for bands 3 and 4 is determined in the image overlap. By convention, we use values of zero to define missing data. A simple model is applied to all bands to restore zeroes to a value of one within the image bitmap. The non-image area surrounding the scene is eliminated from this and the index calculations by the application of an "image-only" mask, detailed in Chapter 2.3 and 2.4.1.

1.5 INDICES

Analysis includes the calculation of a vegetation index, wetness index, temperature, and water reflectance. Each index provides a means to compare a particular feature between different scenes.

1.5.1 Vegetation Index

Subsequent analysis uses band ratioing as a surrogate measure of biomass. The vegetation index is a ratio of TM bands 3 and 4. The normalized difference vegetation index (NDVI) is a quantification of green biomass. It is not meaningful for water.

EQUATION 3. NDVI (NORMALIZED DIFFERENCE VEGETATION INDEX)

NDVI = R(4) - R(3) R(4) + R(3)

where 4 and 3 are near-infrared and red bands, respectively, and R = reflectance after aerosol correction. A weight of 0.01 for the ratio denominator will scale NDVI by 100, such that an NDVI of 1.0 produces a count of 100.

1.5.2 Wetness Index

The wetness index is a measure of the wetness in the soil observable through the canopy and is particularly effective in distinguishing tidal influence in the coastal marsh zone and in areas with thin vegetation canopy. It is calculated based on the inversion of a procedure used to delineate open or standing water. We find that the gradient provided with the inverse, the "wetness index", shown here is effective in delineating the extent of tidal flooding in the coastal marshes.

Temperature is calculated from the thermal band, TM band 6. The contrast in water temperatures is particularly noteworthy in Florida gulf coastal waters during the winter season, when it is possible to observe the source and redistribution patterns of the relatively warmer waters of the Floridan aquifer. We convert radiance of TM band 6 to Celsius + 5 . Resulting values include freezing temperatures, which may occasionally occur in the region.

Water reflectance is calculated as the difference of bands 2 and 4.