A method was devised to use Feature Analyst Extension of ArcGIS to extract the Lockheed Martin Corporation building from a high resolution aerial image of the South Valley Regional Airport.
Feature Analyst Extraction of Lockheed Martin building using ArcGIS
1. 2012
MULTI-CLASS EXTRACTION OF
URBAN ENVIRONMENT
NAME OF COURSE: GGR1911H
ENVIRONMENTAL REMOTE
SENSING
NAME OF STUDENT:
ARIEZ REYES TAMBOOWALLA
STUDENT NUMBER: 999294473
NAME OF INSTRUCTOR:
JANE LIU
2. ABSTRACT
A method was devised to use Feature Analyst Extension of ArcGIS to extract
urban environmental features from a high resolution ortho corrected aerial image,
especially large concrete roof structures. A cropped image of a downtown urban area near
the Salt Lake City, Utah was used to specify my training sets. It is understood that
running a supervised classification is effective to extract trees, major driveways, large
cement roof and wooden roof structures. A clutter removal algorithm was run to refine
the data in order to avoid overlapping of extraction features along smaller cement and
wooden roof houses to maintain continuous road pathways. Batch processing was then
used in order to implement this feature modeller to the complete original image. The
classification of the complete urban area was successful. These results were then paired
with an elevation image counterpart to project the dominant structures and vegetation
cover areas in ArcScene. Finally in order to check the universal applicability of my
cement roof training set, the same learning scheme was used on the South Valley
Regional Airport to extract the Lockheed Martin Corporation building. The extraction
was a success.
4. 1.INTRODUCTION, IMPORTANCE & PURPOSE
The importance of this project topic was to determine if Feature Analyst could be used to
classify and extract major urban features from multi-spectral and visible range images of
numerous locations across Salt Lake City (area of interest), by using the Feature Analyst
5.0 Extension of ArcGIS 2010 software. Image classification uses the spectral
information association of digital numbers in one or more spectral bands to classify each
individual pixel and extract target features from the image. Supervised learning
implemented here uses hand- digitized examples of cartographic features in an image to
extract features that closely resemble the examples provided, and returns them in a form
of a vector file. The conceptual model of the learner was reinforced by selecting correct
and incorrect examples from the return set of features provided by the agent. This project
report will give some insight into the techniques used, problems encountered, and
conclusions obtained to achieve optimum image classification.
2. DATA SOURCES
For this project, images of Salt Lake City and South Valley Regional Airport were used.
Figure 1 Salt Lake City & South Valley Regional Airport1
5. 3.METHODOLOGY
This project demonstrates another type of multi-class extraction. It will involve a general
workflow often encountered when approaching extraction problems. Beginning first with
image and target feature assessment, a decision on whether a single or multi-class
extraction is most appropriate, draw a good training set, and then set initial extraction
settings. After looking at the results, options to improve them through postprocessing,
and then apply this workflow to another image of interest using Feature Modeller will be
considered. In this project Feature Analyst is used to extract vegetation, structures and
pathway features. Creating training sets for vegetation; structures and pathway target
features; setting custom learning parameters then combine the supervised learning classes
into a single multi-class training layer, and setting custom parameters again will follow.
A new Feature Class can be created by clicking the Create New Feature class button on
the Feature Analyst toolbar (figure 2). This then created a new layer that I saved within
the output folder created earlier.
Figure 2. Feature Class creation
6. Now with a new Feature Class, drawing my Training Sites is the next step. This is done
by using the ArcMap Sketch tool within and Editing Session, similar to how created
Training Sites had been created in the previous exercises. A few screenshots of the
Training Site creations are shown below. (figure 3)
Figure 3. Feature Class training sets
After creating the Training Sites, Supervised Learning is set up by clicking the Learning
button on the Feature Analyst toolbar. Once the button is clicked the Learning dialogue
box opens up. (figure 4)
The Bands Selected list box must be verified so that the urban_crop.tiff (training image),
urban.img (reflectance image) and the urban_elev.img (elevation image) appeared in the
Bands Selected list. The Input Representation field (figure 14) is then set up by clicking
on the Input Representation tab and setting the Pattern width to the appropriate choice
depending on the feature class selected.
7. Figure 4. Supervised Learning Input Bands
By selecting Learning Options and setting the Aggregate areas, Min. Area field according
to the current feature class, the customization was complete.(figure 15)
After clicking RUN in the Set Up Learning dialogue box, ArcMap prompted me to name
and save the feature classes. Then the Feature Analyst Process Box opened up and the
resulting layers were added to ArcMap.(figure 16)
The Square Up Features tool (figure 17) allows you to straighten up the edges of
polygons by squaring up the corners. This is helpful for improving the appearance of
extracting buildings or other square-type objects.(figure 18) Also the Smooth Feature
(figure19) allows you to clean up the edges of polygons to smoothen the edges of the
roads and trees to more closely resemble the actual road pathway and vegetation cover
areas.(figure20)
8. A Feature Modeller was used to carry out the batch processing operation.(figure5)
Figure 5. Feature Modeller
9. Figure 6. Multi-class Input feature classification
Next a multi class extraction layer (figure 6) was created with wall to wall classification
(figure7) to ensure that all the image pixels fell into one of the above categories
Figure 7. Wall to Wall Classification
10. It was challenging to get the Feature Analyst’s 3D layer to display correctly within
ArcMap or ArcScene, ESRI’s 3D viewer; the look of 3D buildings was visualized by
extruding the building polygons within ArcScene.(figure 25,26,27) This is done by going
into the Properties of a layer, clicking the Extrusion Tab and typing in the appropriate
extrusion value.
Figure 8. After 3D Extrusion without image layer
Finally the cement structure training set was applied to the South Valley Regional Airport
to check for universal application of the training set.(figure 28,29)
Figure 9. South Valley Regional Airport with 3D extrusion
11. 4.ISSUES ENCOUNTERED
The first issue encountered was the problem with limited usage of the trial version of the
Feature Analyst 5.0 Extension of ArcGIS 2010 which allowed only working on the
limited number of images available in the tutorial section. I was hoping to work on urban
images more relevant locations like a satellite image of University of Toronto and the
nearby Porter Island Airport. Hence an urban landscape of Salt Lake City and the nearby
South Valley Regional Airport was used.
The next issue was the extraction of pathways i.e. roads, footpaths, parking lots and
minor intersections. Many of the cement roof structures i.e. that are greyish white are in
some cases too dark to be distinguished from the light and dark grey pixels that make up
most of the pathways. Hence the training sets had to be redone to take into account this
gradual shift of pixel color from white to grey in order to distinguish between pathways
and cement roofed structures to prevent overlapping of classification.
The next issue was the problem of shadows. Since both structures and trees have shadows
that cover areas of pathways and other structures and trees; classifying them into the
correct class proved to be challenging. The ‘input representation’ part of the supervised
learning feature was modified to classify that a combination of black and medium grey
pixels would be a road whereas a combination of red and black pixels would constitute
vegetation. However along narrow pathways there was still some misclassification
leading to broken pathways. This problem was solved by separating the 2 classes into 2
different layers with the pathway being given higher preference. This improved logic was
gave superior results.
12. The final issue encountered was using the elevation image of the area of interest to
project the ground structures from the image. Difficulty in getting the Feature Analyst’s
3D layer to display correctly within ArcMap or ArcScene made me create the look of 3D
buildings by extruding the building polygons within ArcScene.
5.RESULTS
Wall to wall classification was made possible allowing for all the land area to be
classified. Use of ‘convert to line’ feature was successful in creating a map like
representation of the area. (figure 24)Extrusion of land features was made possible
through ArcScene Reapplication of the concrete roof training set to another image
enabled classification of the Lockheed Martin Corporation building inside the South
Valley regional airport area.
6.ANALYSIS
Figure 10. Graph of South Valley Urban Landscape classification
13. As you can see the distribution of land area amongst the 4 different classes 49% was
covered with vegetation, 38% was covered with pathways, 8% with cement roof
structures and the remaining 5% with wooden roof structures.
7.CONCLUSIONS
It was possible to classify the entire image into the 4 chosen classes of classification.
Extracting the various pathways intersections of this image proved to be the most
challenging. Use of a 3D elevation image required a higher resolution image to
implement building extrusion. Overall this project has taught me how to construct good
training sets and implement correct search algorithms to classify the given image over a
wide range of input criteria. This opportunity was immensely motivational in cultivating
my interest in future GIS work.
8.REFERENCES
1. Feature Analysis 5.0 Tutorial
2. Google Earth
14. 9.FIGURES (IN ORDER OF OPERATIONS PERFORMED)
Figure 11. Feature Class creation
Figure 12. Feature Class training sets