 Coweeta
LTER created incremental land
cover classifications along five year intervals for the southern
Appalachian Mountains in order to assist researchers efforts to study change in land cover over time.
Recently the United States Geologic Survey (USGS) has published land
cover for this area, the National Land Cover Database (NLCD), dated 2001.
Coweeta adopted USGS methodologies to create land cover datasets for 1986
onward using available Landsat imagery. In the future as imagery becomes
available, Coweeta LTER plans to create another classification every 5 years,
beginning with 2006, 2011, etc.
|
|
View and Download Land Cover |
Developing Land Cover
Classification
The NLCD 2001 dataset was developed by the
Natural Resources Spatial
Analysis Laboratory (NARSAL) at the
Institute of Ecology,
University of Georgia in Athens, Georgia. Coweeta LTER has worked in
collaboration with NARSAL to develop the new classifications using
methods as close as possible to those used in developing NLCD 2001. This
paper is a detailed description of those methods, created as a reference
for persons creating future land cover classifications.
Table of Contents
Chapter 1: Description of NLCD Land Cover Classes
Chapter 2: Preparing Imagery
Chapter 3: Digitize Cloud Areas, Create Cloud Masks
Chapter 4: Preparing DEMs
Chapter 5: Preparing Point Data
Chapter 6: Classifying images using NLCD classifier
and See5
Chapter 7: Preparing Wetland Classification
Chapter 8: Classifying urban areas using NLCD
classifier and Cubist
Chapter 1:
Description of NLCD 2001 Land Cover
Classes
The 2001 National Land Cover Database is a thematic raster image with
each pixel assigned an integer value that represents a particular type of
land cover. Each discrete value and its associated land cover is
described on the USGS
land cover website.
USGS describes the classification as a modification of the Anderson
land-use and land-cover classification system. The Anderson system was
modified for applicability to Landsat image classification, as many level
III classes are best derived from aerial photography, and some level II
classes are best consolidated for accuracy.
NLCD Land Cover Class Definitions
Water - All areas of open water or permanent ice/snow cover.
11. Open Water - all areas of open water, generally with less than 25%
cover of vegetation/land cover.
12. Perennial Ice/Snow - all areas characterized by year-long surface
cover of ice and/or snow.
Developed Areas characterized by a high percentage (30 percent or
greater) of constructed materials (e.g. asphalt, concrete, buildings,
etc).
21. Low Intensity Residential - Includes areas with a mixture of
constructed materials and vegetation. Constructed materials account for
30-80 percent of the cover. Vegetation may account for 20 to 70 percent
of the cover. These areas most commonly include single-family housing
units. Population densities will be lower than in high intensity
residential areas.
22. High Intensity Residential - Includes highly developed areas where
people reside in high numbers. Examples include apartment complexes and
row houses. Vegetation accounts for less than 20 percent of the cover.
Constructed materials account for 80 to100 percent of the cover.
23. Commercial/Industrial/Transportation - Includes infrastructure (e.g.
roads, railroads, etc.) and all highly developed areas not classified as
High Intensity Residential.
Barren - Areas characterized by bare rock, gravel, sand, silt, clay,
or other earthen material, with little or no "green" vegetation present
regardless of its inherent ability to support life. Vegetation, if
present, is more widely spaced and scrubby than that in the "green"
vegetated categories; lichen cover may be extensive.
31. Bare Rock/Sand/Clay - Perennially barren areas of bedrock, desert
pavement, scarps, talus, slides, volcanic material, glacial debris,
beaches, and other accumulations of earthen material.
32. Quarries/Strip Mines/Gravel Pits - Areas of extractive mining
activities with significant surface expression.
33. Transitional - Areas of sparse vegetative cover (less than 25 percent
of cover) that are dynamically changing from one land cover to another,
often because of land use activities. Examples include forest clearcuts,
a transition phase between forest and agricultural land, the temporary
clearing of vegetation, and changes due to natural causes (e.g. fire,
flood, etc.).
Forested Upland - Areas characterized by tree cover (natural or
semi-natural woody vegetation, generally greater than 6 meters tall);
tree canopy accounts for 25-100 percent of the cover.
41. Deciduous Forest - Areas dominated by trees where 75 percent or more
of the tree species shed foliage simultaneously in response to seasonal
change.
42. Evergreen Forest - Areas dominated by trees where 75 percent or more
of the tree species `maintain their leaves all year. Canopy is never
without green foliage.
43. Mixed Forest - Areas dominated by trees where neither deciduous nor
evergreen species represent more than 75 percent of the cover present.
Shrubland - Areas characterized by natural or semi-natural woody
vegetation with aerial stems, generally less than 6 meters tall, with
individuals or clumps not touching to interlocking. Both evergreen and
deciduous species of true shrubs, young trees, and trees or shrubs that
are small or stunted because of environmental conditions are included.
51. Shrubland - Areas dominated by shrubs; shrub canopy accounts for
25-100 percent of the cover. Shrub cover is generally greater than 25
percent when tree cover is less than 25 percent. Shrub cover may be less
than 25 percent in cases when the cover of other life forms (e.g.
herbaceous or tree) is less than 25 percent and shrubs cover exceeds the
cover of the other life forms.
Non-Natural Woody - Areas dominated by non-natural woody vegetation;
non-natural woody vegetative canopy accounts for 25-100 percent of the
cover. The non-natural woody classification is subject to the
availability of sufficient ancillary data to differentiate non-natural
woody vegetation from natural woody vegetation.
61. Orchards/Vineyards/Other - Orchards, vineyards, and other areas
planted or maintained for the production of fruits, nuts, berries, or
ornamentals.
Herbaceous Upland - Upland areas characterized by natural or
semi-natural herbaceous vegetation; herbaceous vegetation accounts for
75-100 percent of the cover.
71. Grasslands/Herbaceous - Areas dominated by upland grasses and forbs.
In rare cases, herbaceous cover is less than 25 percent, but exceeds the
combined cover of the woody species present. These areas are not subject
to intensive management, but they are often utilized for grazing.
Planted/Cultivated - Areas characterized by herbaceous vegetation that
has been planted or is intensively managed for the production of food,
feed, or fiber; or is maintained in developed settings for specific
purposes. Herbaceous vegetation accounts for 75-100 percent of the cover.
81. Pasture/Hay - Areas of grasses, legumes, or grass-legume mixtures
planted for livestock grazing or the production of seed or hay crops.
82. Row Crops - Areas used for the production of crops, such as corn,
soybeans, vegetables, tobacco, and cotton.
83. Small Grains - Areas used for the production of graminoid crops such
as wheat, barley, oats, and rice.
84. Fallow - Areas used for the production of crops that do not exhibit
visable vegetation as a result of being tilled in a management practice
that incorporates prescribed alternation between cropping and tillage.
85. Urban/Recreational Grasses - Vegetation (primarily grasses) planted
in developed settings for recreation, erosion control, or aesthetic
purposes. Examples include parks, lawns, golf courses, airport grasses,
and industrial site grasses.
Wetlands - Areas where the soil or substrate is periodically saturated
with or covered with water as defined by Cowardin et al.
91. Woody Wetlands - Areas where forest or shrubland vegetation accounts
for 25-100 percent of the cover and the soil or substrate is periodically
saturated with or covered with water.
92. Emergent Herbaceous Wetlands - Areas where perennial herbaceous
vegetation accounts for 75-100 percent of the cover and the soil or
substrate is periodically saturated with or covered with water.
Overview of Methods
The process used for development of Nationall Land Cover Databases (NLCD)
differs from many conventional classification methods in its use of
training data. Training data are point data, with each point
corresponding to a specific pixel of the image to be classified. The
pixels associated with training points are a subset of total pixels in
the image, and each point is assigned one of the NLCD land cover classes
by hand, with the operator looking at the Landsat imagery and Digital
Ortho Quarter Quad (DOQQ)aerial photography to determine the best fit
class. The points should be clustered on DOQQs and should be numerous
enough to proportionally represent all land cover classes. The DOQQs
should be spread evenly over the image to represent all different types
of areas. Once each point is assigned a classification, the point file is
converted to a raster and is used as training data in the See5 or Cubist
programs, programs which create regression models for classification
based on the training data. Urban classes are created in a separate
process from non-urban classes, as the urban classification is processed
using continuous data representing percent impervious cover (0% - 100%)
and the non-urban classification is processed using thematic data
(discrete land cover classes). The See5 model creates regressions based
on thematic data and is used for the non-urban classification, while
Cubist creates continuous data regressions and is used for the urban
classification.
Classification is performed using three different Landsat images as
independent variables: Leaf-on (June-August), Leaf-off
(December-February), and Spring (March-April). Additionally, a Digital
Elevation Model (DEM) and aspect and slope derived from the DEM may also
be used as independent variables. These images are particularly helpful
in mountainous areas, such as the Southern Appalachian area in which
Coweeta is interested.
Description of Software Requirements
Primarily, ERDAS IMAGINE and the See5 and Cubist models are the software
programs used to perform the classification. Additionally, ArcGIS is
useful for classifying point training data and general data viewing
purposes, while Arc/INFO and the associated GRID programs are easier than
IMAGINE for a few specific coverage and raster processes.
Southern Appalachian Study Area
The Southern Appalachian study area that is of interest to Coweeta
research covers parts of Georgia, Tennessee, North Carolina, and
Virginia. The study area used for classification covered each of these
counties plus a 5 kilometer buffer around the entire area (adding a small
portion of Kentucky). The study area is shown below.
Chapter 2: Preparing Imagery
Before beginning the classification process, Landsat imagery must be
selected and purchased from USGS and prepared for classifying. This
section will cover pre-classification preparation.
Selection of Imagery
Imagery can be browsed and selected through
USGS Online Catalog.
Some types of imagery can be ordered online, but for this classification,
terrain-corrected imagery is required and must be ordered by phone. A US
Government-Affiliated User form must be completed for access to this
highly-accurate imagery. The appropriate satellite data set is either
Landsat 4-5 TM or Landsat 7 ETM+, depending on the required dates.
The Southern Appalachian study area is covered by 7 Landsat scenes:
path 17 rows 34 and 35, path 18 rows 34, 35, and 36, and path 19 rows 35
and 36. Three seasons of imagery are required, leaf-on, leaf-off, and
spring. Three images must be ordered for each path/row combination, for a
total of 21 scenes. Ideally, images would contain no cloud, snow, or haze
cover, but most images are somewhat cloudy. In order to find high-quality
imagery, a three-year window is allowed around the target date, one year
on each side. For instance, in the 1996 classification, images were
selected from 1995, 1996, or 1997. Images selected should have as little
cloud, snow, or haze cover as possible, as these areas cannot be
classified. It should be noted that the study area does not cover all of
any of the scenes, and clouds in locations not in the study area are not
a factor. Nearly none of the images is completely cloud-free, and if the
best image for one season contains some clouds, try to select images for
the other seasons that do not have clouds in the same area. Once the 21
scenes are selected, order terrain-corrected imagery using Coweeta’s
preferred format:
Data Format: BSQ
File Format: EOSAT Fast
Image Orientation: Map
Pixel Size: 30m
Resampling Tech.: NN
Projection: UTM
Zone: 17N
Datum: NAD27
In 2006, the price for terrain-corrected images was $625 per scene, with
a reduced price of $310 for second and third scenes adjacent to another
scene on the same path and on the same date.
Importing Imagery
After receiving imagery from USGS on CDs, copy each CD to the location on
your hard drive/network drive where you will be working. Single scenes
will be on one CD; multi-scenes will be on multiple CDs. Multi-scenes
from USGS are single images, with different bands on different CDs, as
opposed to different scenes on different CDs with all bands of the same
scene on one CD. IMAGINE importer can import directly from EOSAT Fast
format, but it has trouble with multi-scene images. Thus the single-scene
images will be imported using a different method than the multi-scenes.
Single-scene importing
Open the import dialogue in IMAGINE by clicking the ‘Import’ icon on the
IMAGINE toolbar. Select ‘import’ at the top, then select Landsat TM Fast
(EOSAT) from the dropdown menu.
Mosaicing Imagery
After importing each scene into IMAGINE, the scenes for each season must
be mosaiced together. For the classification, it is imperative that
pixels from each scene line up exactly with pixels from adjacent scenes,
as well as overlay exactly with scenes from different seasons. If all
imagery was ordered at the same time and in the same projection from
USGS, then pixels should already line up.
Aligning Images
To correct misaligned images, the misaligned image will be shifted
slightly so that the pixels line up. Begin by opening a properly aligned
and the misaligned image in two different IMAGINE viewers. Geolink the
two viewers by right clicking in one, selecting geolink, and then left
clicking in the second viewer. Zoom in to an area in which the two images
overlap, showing a few pixels in each viewer. Locate a single pixel in
the properly aligned image and its corresponding pixel in the misaligned
image. Unlink the viewers and zoom in as far as possible to one corner of
the chosen pixel in the aligned image. Open the inquire cursor in that
viewer and move the crosshairs as close as possible to that corner; the
inquire cursor will show the X and Y coordinates of the pixel corner. In
the second viewer, zoom in as far as possible to the same corner of the
corresponding chosen pixel; the coordinates will be different as the
images are not aligned. Once zoomed in on the pixel corner of the
misaligned image, be sure that the arrow cursor is selected, click on
‘Raster’ at the top of the viewer and select ‘Set Drop Point.’ Click as
closely as possible to the pixel corner. The image offset box will open;
type in new X and Y coordinates that are shown in the inquire cursor
window of the properly aligned image and click OK. The image will be
shifted to align with the other. Repeat this process for all misaligned
images.
Mosaic
Once all images are properly aligned, the 7 scenes for each season should
be mosaicked using the IMAGINE mosaic tool. Click on the ‘DataPrep’ icon
on the IMAGINE toolbar, select ‘Mosaic Images’ and then ‘Mosaic Tool.’
Add all 7 images for a single season to the mosaic tool, then adjust the
order of overlap so that better images (less cloud cover) are on top of
worse images in areas of overlap. In the ‘Set Overlap Function’ dialogue,
be sure to select ‘Overlay.’ All other default settings should be
acceptable. Run the mosaic. Repeat for the two remaining seasons.
Subset mosaic
The mosaiced images should be subset to the area of interest. You will
need a shapefile of the area of interest for this operation; for the
southern Appalachian study area, the area of interest is all counties in
the study area plus a 5 km buffer. Open a shapefile/coverage of the
counties in ArcMap, merge them all into one polygon, then buffer the
polygon by 5 kilometers. In this initial operation, subset only one of
the mosaicked images. The two remaining images will be subset in the
IMAGINE Modeler using the initial subset image to ensure that the exact
pixels are contained in each subset.
Open the image to be subset (leaf-on) and the shapefile of the study area
in the IMAGINE viewer. Select the study area polygon in the shapefile. In
the AOI (area of interest) tools menu, select ‘convert polygon to aoi.’
This will create an *.aoi of the study area. Click on the ‘DataPrep’ icon
on the IMAGINE toolbar and select ‘Subset Image.’ Add the image to be
subset (leaf-on) to the dialogue box, click ‘aoi,’ and select ‘from
viewer.’ The study area aoi must be selected in the viewer for this
operation. Name the output file and run the subset operation.
Create image of study area
After subsetting the first image, you will need to create a thematic
image of the study area containing 1 layer, with a value of 1 for pixels
in the study area and a value of 0 for pixels outside of the study area.
This image will be invaluable for many subsequent operations, and will
ensure that every subset image contains exactly the same pixels.
Open IMAGINE Modeler. Insert one input raster and one output raster, with
one process between. Use the initial subset image (leaf-on) as the input
raster and create a name for the output raster such as ‘studyarea.img.’
Be sure that the output raster is thematic unsigned 8-bit. Click on the
process circle. In the dropdown menu in the top right, select
‘conditional,’ then select the ‘EITHER’ statement below. The EITHER
statement will appear in the lower box; change this statement to:
EITHER 1 IF (<subsetmosaic_layer1> gt 0) OR 0 OTHERWISE
Only one layer of the input image is necessary, any layer will work. The
output image ‘studyarea.img’ will have a value of 1 if the pixel is in
the study area, 0 if it is outside.
Subset remaining 2 mosaic images
The remaining 2 mosaicked images (leaf-off, spring) will be subset in
IMAGINE Modeler. This time use two input rasters and one output raster
with one process between. Use the image to be subset as one input and
‘studyarea.img’ as the second input. Run process with the conditional
statement:
EITHER <inputmosaic> IF (<studyarea> gt 0) OR 0 OTHERWISE
Run model on both remaining mosaics. You will end up with subset images
for all 3 seasons, each having the exact same pixel content.
Chapter 3: Digitize Cloud Areas, Create
Cloud Masks
Areas in images covered by clouds, cloud shadows, snow, or haze cannot be
classified. It is necessary to digitize an AOI for these areas in all 3
images, so that they can be removed later. You will create 3 different
AOIs, one for each image.
Digitize Cloud Areas
Open the leaf-on image (subset to study area) in the IMAGINE viewer.
Click on the AOI menu on the viewer toolbar and open the AOI Tools
toolbar. Click on the icon representing ‘create polygon aoi.’ After
zooming in to an appropriate scale, draw polygons around all clouds,
cloud shadows, snow, or significant haze (probably no snow in leaf-on
image). Save the AOI with a name such as ‘leafon_cloudaoi.aoi.’ Repeat
for leaf-off and spring images, resulting in 3 *.aoi files.
Create Cloud Masks
In the IMAGINE viewer, open ‘studyarea.img’ and ‘leafon_cloudaoi.aoi.’
Select all polygons in the AOI. Using the subset function under the
‘DataPrep’ icon on the IMAGINE toolbar, subset ‘studyarea.img’ using the
AOI in the viewer. Change the output name to leafon_clouds.img. The
resulting thematic image will have a value of 1 for areas containing
clouds in the leaf-on image and 0 for all other areas. This image will be
used later to remove cloudy areas from the classification. Repeat the
process for leaf-off and spring cloud AOIs, resulting in 3 cloud masks.
Chapter 4: Preparing DEMs
Digital elevation Models (DEMs) can be very useful in classification when
used in combination with Landsat imagery, particularly in areas with
significant topographical variation such as the Appalachian Mountains,
and are available free from USGS. Additionally, aspect and slope images
can be created from DEMs and also used in classification. Because DEMs do
not change over time like land cover, the same DEM, aspect, and slope
files can be used for multiple years of image classification, unless more
accurate versions of the same data are released in the future.
Downloading Data
The DEM data that is best for purposes here is the USGS National
Elevation Dataset (NED), and is downloadable free from http://seamless.usgs.gov.
Select only NED in the download menu on the right, then click and drag a
download box over the desired area in the viewer. There is a limit to the
amount of data that can be downloaded at one time, so it may take
multiple sessions to get the entire desired area. The NED files will
download in multiple tiles, which must be mosaicked together.
Reprojecting Data
After NED data has been downloaded for the entire study area, each
individual tile should be reprojected to Coweeta’s preferred projection,
UTM zone 17N, NAD 27. Reproject in IMAGINE, using the ‘reproject’ tool
under the ‘DataPrep’ icon on the imagine toolbar. Because the data is in
multiple tiles, it may be useful to use the batch process function in the
reprojection window.
Mosaic Tiles
The reprojected tiles should be mosaicked together in IMAGINE to create
one image. Using the IMAGINE mosaic tool, add all tiles. Change the
output image type to ‘unsigned 16-bit,’ as the default 8-bit does not
have enough range of values. Run the mosaic.
Subset Mosaic
The resulting mosaic will not align properly with the Landsat imagery.
Shift the DEM to line up with the imagery, using the inquire cursor on
the Landsat image and set drop point on the DEM, as described previously
in aligning images. Subset the properly aligned DEM in the IMAGINE
modeler using the same method used for the second and third Landsat
images.
Generating Aspect and Slope
The IMAGINE Modeler has functions to calculate aspect and slope from an
elevation image. Open modeler and create a model with one input raster,
one output raster, with one process between. Select the DEM image as the
input raster. Slope and aspect functions are under the ‘Surface’ function
list. For slope, select PERCENT SLOPE (<raster> , <units>). Replace
<raster> with the input DEM and replace <units> with ‘meters.’ Change the
output raster type to ‘continuous’ and ‘Unsigned 16-bit,’ then select a
name and location for the output. Run the model.
The aspect model will look the same, but the function should be ASPECT
(<raster>), also from the ‘Surface’ list. Replace <raster> with input
DEM. Change output raster type to ‘continuous’ and ‘Unsigned 16-bit,’
then select a name and location for the output. Run the model.
Chapter 5: Preparing Point Data
Classifying Non-Urban, Non-Wetland
Points
The point data will be the training data for the classification of
non-urban areas of the Landsat imagery. The points are clustered on
Digital Ortho Quarter-Quad (DOQQ) aerial photos, and the DOQQs are spread
across the study area to form a representative sample of all types of
land cover. The points are classified by the user with the help of the
aerial photos and the Landsat imagery, and are then used as training data
in the classification model. Points were generated for the NLCD 2001
classification, and the same points will be used for subsequent Coweeta
classifications.
Classifying Points
The easiest way to classify points is to open an ArcMap document and add
the shapefile containing the points, all three landsat images, and the
DOQQs. Adding only a few DOQQs at a time will save time by keeping the
ArcMap document small. Create a new field in the point shapefile for land
cover class. Open the editor toolbar, click ‘Start Editing,’ and select
the point shapefile. This will allow you to add values to the new
landcover field. Working with one DOQQ at a time, zoom in to each point,
looking at the entire pixel of the landsat image that it overlays and the
DOQQ for that area and assign a classification to that point. The classes
are described in more detail above, but the only ones of use at this time
are: 11 – water, 31 – barren, 41 – deciduous forest, 42 – evergreen
forest, 43 – mixed forest, 52 – scrub/shrub, 71 – grassland, 81 –
pasture, and 82 – row crop. Some classes are not applicable to the
Southern Appalachian area, and wetland and urban classes will be
determined at a later time. After selecting a class for a particular
point, input the class numeric value into the new land cover class field
of the shapefile. This process takes a bit of practice. DOQQ’s are taken
in winter and most are color infrared, so photosynthetic activity will
appear red. Thus evergreen forests (42) will appear red, deciduous
forests (41) will not. However, much deciduous forest has an evergreen
understory, and classification becomes more difficult. Understory
evergreen will have a finer texture than evergreen canopy; individual
trees should be discernable on the DOQQ for evergreen forests. Row crop
(82) and pasture (81) are also difficult to distinguish, and may change
from year to year. In general, row crops will be brighter and not reddish
on the landsat image as more bare soil shows than crop cover. Also, plow
rows may be visible on DOQQs. Grassland (71) and scrub/shrub (52) are
generally considered as a continuum between clearcut and forest regrowth.
Most grassland will be recently clearcut areas, most scrub/shrub will be
clearcuts a few years old, and older clearcuts will fall under forest
classes.
Convert Point Shapefile to Point
Coverage
Convert point shapefile to point coverage in Arc/INFO. Open Arc/INFO
workstation and change workspace to the location of the point shapefile.
On the Arc command line, type:
shapearc <inputpointshapefile> <outputpointcoverage>
This command will convert the shapefile to a point coverage, maintaining
all field data. The new coverage must be built. On the Arc command line,
type:
build <pointcover> point
The coverage will now be ready for use.
Convert Point Coverage to GRID
To convert the point coverage to a raster image, it will first need to be
converted to GRID format. In order to make sure that the pixels of the
point image line up precisely with the pixels of the Landsat imagery and
the DEM created earlier, the study area image used to subset the imagery
and DEM will need to be converted to GRID format. The point coverage will
then be snapped to that GRID during conversion. Change the workspace to
the location of the study area by typing on the Arc command line:
ws drive:path/foldercontainingstudyarea
Convert the study area image to a grid, placing the grid in the workspace
containing the point coverage by typing:
imagegrid <studyareaimage.img> <drive:path/foldercontainingpointcover/studyareagrid>
Change the workspace to the location containing the new GRID and the
point coverage. Enter GRID by typing ‘grid’ on the Arc command line. On
the GRID: command line, enter:
setwindow <studyareagrid> <studyareagrid>
The first ‘studyareagrid’ sets the extent of your GRID window to the
extents of the GRID ‘studyareagrid,’ the second sets the snap grid to ‘studyareagrid.’
This command will ensure that the pixels of the GRID created from the
point coverage will align properly with the study area and the landsat
images
Still in GRID, convert the point coverage to a GRID by entering on the
command line:
<outputgrid> = pointgrid(<pointcoverage>, <classfield>, #, #, 30)
‘Outputgrid’ is the name of the new GRID, ‘pointcoverage’ is the name of
the coverage of point data, ‘classfield’ is the name of the field in the
coverage with the classification values, and 30 is the size of the output
pixels.
Convert GRID to Image
Using IMAGINE importer, import the point GRID to an IMAGINE image.
Enter Wetland Points into Point image
Because wetland areas cannot be accurately discerned by eye from DOQQ’s
or Landsat imagery, they will be put in after classifying points. This
will be accomplished by finding any forest points (41, 42, 43) that lie
over NWI wetlands and changing them to forested wetland class (90) and
any grassland or scrub-shrub points that lie over wetlands and changing
them to emergent wetland class (95). For 2001 NLCD classification,
USGS-contracted labs created wetland classifications using these wetland
points and a series of models setting parameters to limit areas allowed
to be classified as wetland. This method is beneficial in that wetland
classes are determined from the imagery. However, wetlands constitute a
relatively small area in the Southern Appalachians, and this sort of
modeling is not necessary for Coweeta’s purposes. Coweeta will take
forested and non-forested wetlands directly from NWI data at the end of
the classification rather than classify them from the imagery. With this
method, the wetland points can essentially be eliminated from the point
data. Nonetheless, the process of reclassifying them will be described
here, though USGS-method models will not.
The NWI wetland image is divided into open water, forested wetland, and
emergent wetland. We will ignore these divisions as our determination of
classes should supersede these divisions and only consider wetland = yes
or wetland = no. Open the IMAGINE modeler and enter 2 input images and
one output image, with one process between. The modeler must be run
twice, once for forested wetland and once for emergent; first will be
forested. For the input files, use the NWI wetland image and the points
image. For the process, enter the conditional statement:
EITHER 90 IF (<pointimage> gt 40 AND <pointimage> lt 44 AND <nwiimage> gt
0) OR <pointimage> OTHERWISE
The output of this model will be ‘pointsintermediate.’ Run the model a
second time, with the inputs NWI wetland image and pointsintermediate.
For the process, enter the conditional statement:
EITHER 95 IF (<pointsintermediate> gt 51 AND <pointsintermediate> lt 72
AND <nwiimage> gt 0) OR <pointsintermediate> OTHERWISE
Output ‘pointsfinal.’ The output image will have all forest points (41,
42, 43) occurring over wetlands converted to forested wetland (90), and
all grassland (71) and scrub/shrub (52) points occurring over wetlands
converted to emergent wetland (95).
Chapter 6: Classifying Non-Urban Areas
using NLCD classifier and See5
Non-urban, non-wetland areas will be classified with the See5 model,
using the point image as training data. This process involves three
steps: the CART Sampling Tool, the See5 Model, and the CART Classifier.
Initial Classification
CART Sampling Tool
Click on the CART icon on the IMAGINE toolbar and select ‘CART Sampling
Tool.’ For independent variables, select the leaf-on TM image, the
leaf-off TM image, the spring TM image, the DEM image, the aspect image,
and the slope image. For the dependent variable select the points image.
Type 0 for ‘Ignore values.’ Under ‘Sampling Number,’ select ‘percent’ and
type 80 under ‘Training(%)’ and 20 under ‘Validation (%).’ Select a
location and name for the output *.names file (other output files will
automatically go to same place). Select See5 at the bottom and click
‘OK.’ This process will create *.names, *.data, and *.test files.
See5 Model
After the CART Sampling Tool finishes, open See5 (outside of IMAGINE).
Click the ‘Locate Data’ icon on the toolbar and run the See5 model. This
model will create several other files with the same name as the *.names
file. Elaborate
CART Classifier
After the See5 model finishes, click the CART icon on the IMAGINE toolbar
and select ‘CART Classifier.’ Select the *.names file created earlier as
the input and select ‘See5.’ Select ‘Tree’ and select a path and name for
the output image. Do not use a mask file at this point; click ‘OK.’ The
output will be a thematic, 1-layer IMAGINE image containing the
classification.
Revise/Troubleshoot Classification
The initial classification will probably not be so good. Problematic,
particularly in mountainous areas, is over-estimation of water areas due
to shadows and over-estimation of wetlands. To rectify this problem,
reduce the number of classes to water, forest, and non-forest and
re-classify, then subdivide the forest and non-forest areas.
Create Base Classification
Reduce the number of classes in the point image to three: water (11),
forest (41, 42, 43, 90), and non-forest (31, 52, 71, 81, 82, and 95).
This will be done changing the values in the point image: 41, 42, 43, and
90 will be changed to 41; 31, 52, 71, 81, 82, and 95 will be changed to
81. Open IMAGINE modeler and insert one input raster and one output
raster, with one process between. Choose the points image for the input
raster and choose a name and path for the output. In the process, select
‘Conditional’ on the Functions list, then select ‘CONDITIONAL.’ Change
the function to:
CONDITIONAL { (<pointsfinal> eq 11) 11 , (<pointsfinal> eq 31) 81 , (<pointsfinal>
eq 41) 41 , (<pointsfinal> eq 42) 41 , (<pointsfinal> eq 43) 41, (<pointsfinal>
eq 52 ) 81 , (<pointsfinal> eq 71 ) 81 , (<pointsfinal> eq 81) 81 ,( <pointsfinal>
eq 82) 81 , (<pointsfinal> eq 90) 41 , (<pointsfinal> eq 95) 81 }
Run the model. When the model finishes, re-run CART Sampling Tool, See5,
and CART Classifier, using the new reclassified point image as the
dependent variable.
Masking Cloudy Areas
Check the base classification, particularly in areas that are cloudy on
one or more of the images. Clouds tend to return barren (31, 81 in this
case), and cloud shadows tend to return water (11). In problem areas, the
model will need to be re-run omitting the image with clouds as an
independent variable, then using the resulting classification as a
replacement in the problem area of the base classification.
For cloudy areas in the leaf-on image, run Cart Sampling Tool using only
leaf-off TM image, spring TM image, DEM image, aspect image, and slope
image as independent variables, and use the 3-class points image as the
dependent variable. Run See5 model and CART Classifier with the output
files. This process will create a classification without the problem of
clouds from the leaf-on image.
In areas of leaf-on clouds or shadows, substitute this classification
into the base classification. Open IMAGINE modeler and insert 3 input
rasters, 1 output raster, with one process between. The three input
rasters will be the original base classification, the base classification
omitting the leaf-on image, and the image of leaf-on clouds created
earlier under ‘Create Cloud Masks.’ Choose a name and location for the
output classification. Select ‘Conditional’ under the Function list and
choose the ‘EITHER’ statement. Change the statement to:
EITHER <baseclass_leaf-on_omitted> IF (<leaf-oncloudmask> gt 0 ) OR <baseclass_original>
OTHERWISE
Run the model. The output will have the leaf-on cloudy areas replaced.
Repeat the process for problem areas such as clouds and shadows on other
images.
Sub-dividing Forest classification
Once all problem areas have been improved in the base classification,
create a mask of the forested areas using IMAGINE modeler. In the
modeler, put in one input raster, one output raster, and one process
between. Use the final base classification as the input and name the
output ‘forest_mask.img.’ In the process box, use the EITHER statement
from the conditional function list as follows:
EITHER 1 IF (<baseclass> eq 41 ) OR 0 OTHERWISE
The output will have a value of 1 in forested areas and 0 elsewhere, and
will be used as a mask in the CART Classifier.
Also needed in this step is a points image containing only points with
values of 41, 42, or 43. Create another model with one input raster, one
output raster, and one process. Use the points image containing all
values (11, 31, 41, 42, 43, 52, 71, 81, 82, 90, and 95) as the input
raster and name the output raster ‘forest_points.img.’ In the process
box, use the EITHER statement from the conditional function list as
follows:
EITHER <points_image> IF (<points_image> gt 40 AND <points_image> lt 44)
OR 0 OTHERWISE
The output image will contain only pixels classified as 41, 42, or 43
from the point data. These values are the only ones needed because you
are only classifying forested areas.
Run the CART Sampling Tool exactly as done in creating the base
classification, with the 3 landsat images, DEM, aspect, and slope as
independent variables, forest_points as dependent variable, 80%
training/20% verification, and selecting See5. When the sampling tool
finishes, run the See5 model as before. When See5 finishes, run CART
Classifier as before, but use ‘forest_mask.img’ as a mask layer.
The resulting classification will subdivide the original forested area
into its constituent classes. Because the forest classes are much closer
together, in terms of spectral signature, than the 3 classes of the base
classification, be sure to check this classification closely for
accuracy. Re-run CART Sampling Tool, See5, and CART Classifier 3 more
times, each time omitting one of the Landsat images as an independent
variable. This process was done earlier during base classification to
mask cloudy areas in one or more of the images, and that may still be
helpful, but due to inherent inconsistencies between images and the
spectral closeness of these classes, one of these classifications may be
better than the one using all 3 Landsat images. Check closely against
imagery to find the best. It may be helpful to run CART Sampling Tool,
See5, and CART Classifier 3 additional times, using only one of the
Landsat images as independent variable each time. Find the best
classification by checking against the imagery. A final note on this
process: the leaf-off image can be very good at the deciduous/evergreen
division, as evergreen areas will still be photosynthetically active at
that time of year.
After selecting the best classification from the group, check for cloudy
areas in the images used in that classification. If there are cloudy
areas, replace them as done in the base classification, using a
classification that omitted the image in question and the appropriate
cloud mask.
Sub-dividing Non-forest Classification
Re-classifying non-forest, non-water areas is done exactly the same as
dividing forest classification, only in several iterations. First
scrub/shrub (52) will be separated from other non-forest areas, then
grassland (71) will be separated from the remaining non-forest, then
barren land, pasture, and cropland will be split at the same time.
To subdivide scrub/shrub (52), you will need a non-forest mask and a
points image containing only points with classes 31, 52, 71, 81, and 82,
but with 31, 71, 81, and 82 all coded as 81. Create the mask from the
base classification in IMAGINE modeler as done earlier for ‘forest_mask.img,’
and call it ‘non-forest_mask.img.’ Recode the points image (containing
all classified points) in modeler using the conditional statement:
CONDITIONAL { (<pointsfinal> eq 11) 0 , (<pointsfinal> eq 31) 81 , (<pointsfinal>
eq 41) 0 , (<pointsfinal> eq 42) 0 , (<pointsfinal> eq 43) 0, (<pointsfinal>
eq 52 ) 52 , (<pointsfinal> eq 71 ) 81 , (<pointsfinal> eq 81) 81 ,( <pointsfinal>
eq 82) 81 , (<pointsfinal> eq 90) 0 , (<pointsfinal> eq 95) 0 }
Name this output file ‘class52_points.img.’ Run CART Sampling Tool, See5,
and CART Classifier as before, using ‘class52_points.img’ as the
dependent variable in CART Sampling Tool and ‘non-forest_mask.img’ as a
mask layer in CART Classifier. Again, run the classification tools and
models again up to 6 times, omitting one or two of the images each time.
Check against imagery to select the best classification. Most of the
scrub/shrub should appear in areas that were somewhat recently clearcut,
and should be apparent on the imagery. The southwestern portion of the
Southern Appalachian study area has the most intense clearcuts; check
this area carefully.
Again check for clouds, but they should be less of a problem due to the
relatively small number of pixels that will be class 52.
After creating the best scrub/shrub classification possible, begin on the
grassland (71) classification. Create a non-forest, non-scrub/shrub mask
in modeler using the final scrub/shrub classification as input and naming
the output ‘forest52_mask.img.’ In the process box, enter the conditional
statement:
EITHER 1 IF (<class52image> eq 81) OR 0 OTHERWISE
Create a new points image containing only classes 31, 71, 81, and 82,
with 31, 81, and 82 coded as 81. Use the original points image as input
and run the conditional statement:
CONDITIONAL { (<pointsfinal> eq 11) 0 , (<pointsfinal> eq 31) 81 , (<pointsfinal>
eq 41) 0 , (<pointsfinal> eq 42) 0 , (<pointsfinal> eq 43) 0, (<pointsfinal>
eq 52 ) 0 , (<pointsfinal> eq 71 ) 71 , (<pointsfinal> eq 81) 81 ,( <pointsfinal>
eq 82) 81 , (<pointsfinal> eq 90) 0 , (<pointsfinal> eq 95) 0 }
Name the output file ‘class71_points.img.’ Run CART Sampling Tool, See5,
and CART Classifier just as before, as many times as necessary to find
the best classification. Check closely against imagery; class 71
grassland should also be concentrated at clearcuts, but in areas more
recently clearcut than class 52. Check the southwestern area well. Check
for clouds, and replace areas as necessary.
Finally, classes 31, 81, and 82 will be divided. Repeat the processes as
before, creating a points image containing only classes 31, 81, and 82
and creating a mask of non-forest, non-scrub/shrub, and non-grassland
areas. Run the models with as many variations on input imagery as
necessary and check the classifications carefully against imagery.
Pasture and row crop are difficult to distinguish, and the same area may
be different on different images. However, it should be apparent that
some areas have generally more row crop, and some have more pasture.
(check agriculture statistics). Barren areas should be apparent on
imagery as exceptionally bright.
Overlay Non-urban classification
All of the non-urban classifications should now be overlaid using IMAGINE
modeler. Begin using 71 classification and 31/81/81 classification as
inputs. Name the output ‘class31718182.img’ and run the EITHER statement
from the conditional function list as follows:
EITHER <classification_318182> IF (<classification_71> eq 81) OR
<classification_71> OTHERWISE
Run again using ‘class31718182.img’ and 52 classification as inputs. Name
the ouput ‘nonforestclassification.img’ and run the EITHER statement from
the conditional function list as follows:
EITHER <class31718182.img> IF (<classification_52> eq 81) OR
<classification_52> OTHERWISE
Finally, run IMAGINE modeler again, this time with 3 input rasters: base
classification, forest classification, and non-forest classification.
Name the output ‘landcover_nourban.img’ and choose the CONDITIONAL
statement from the conditional function list in the process box. Enter in
the process box:
CONDITIONAL { (<baseclass> eq 11) 11 , (<baseclass> eq 41) <forestclassification>
, (<baseclass> eq 81) <nonforestclassification> }
The result will be the final classification of non-urban, non-wetland
areas.
Chapter 7: Preparing Wetland
Classification
As mentioned earlier, USGS used fairly complex models to classify
wetlands from imagery, but for Coweeta’s purposes, National Wetland
Inventory data will be used instead of models.
(in-progress)
Chapter 8: Classifying urban areas
using NLCD classifier and Cubist
Description of urban classification (in-progress) |
|