The document describes using GRASS GIS to detect land cover change over 13 years at a mining site in British Columbia. Atmospherically corrected Landsat images from 2001-2014 were analyzed using image differencing of NDVI, TCT, and PCA outputs. Thresholding identified significant change areas. NDVI detected over 2300 ha of change, while TCT and PCA detected over 2000 ha. The open source and automated nature of GRASS GIS makes it suitable for replicable change detection.
2. Presentation Contents
Objective
Why GRASS GIS
Target Site
Data
Atmospheric Correction
Composites and Subsets
Image Differencing
Normalized Difference Vegetation Index
Tasseled Cap Transformation
Principle Component Analysis
The Change
Recommendations
Acknowledgments
Questions
3. Objective
To create an automated, easily replicated, accurate and free method to track landcover change
by scripting in GRASS GIS.
Compare the products of Open Source software to those created in commercial software
packages
4. GRASS GIS
Geographic Resources Analysis Support System (GRASS)
Created by the US Army Corp of Engineers
GRASS is a free GIS software suite used extensively for examining and managing geospatial data
6. Fording River Mine
South Eastern British Columbia
One of the largest coking coal reserves in Canada
263.8 million tonnes in reserves
8.34 million tonnes annually
Mine
7. Data
Continuing the ‘free’ theme -- http://earthexplorer.usgs.gov/
The scenes (Path 42, Row 25) cover 13 years of mining activity from 2001 to 2014 and collected
by three different Landsat sensors; 5, 7, and 8.
Each image was collected in August of that year to maximize the data continuity
8. Atmospheric Correction
In order to maintain data integrity and produce accurate, usable, products, the images must be
corrected for atmospheric distortion.
GRASS GIS includes an atmospheric Correction module; i.atcorr
The 6S Algorithm – Second Simulation of a Satellite Signal in the Solar Spectrum
9. i.atcorr
Can be applied to all Landsat satellites as well as several other common platforms including IKONOS,
RapidEye, and Modis, as well as Aerial Photography
The script command:
Flags indicate options within the software,
for ETM+ images taken before 1999 the flag
‘a’ must be used, and ‘b’ for ETM+ images
after 1999. None of the other sensors require
this flag.
The algorithm requires several parameters to be written to a parameter file, ‘pfile’
grass.run_command('i.atcorr',
flags='a',
input=line,
elevation='elev_int',
parameters=pfile,
output=oname,
overwrite=True)
grass.run_command('i.atcorr',
input=line,
elevation='elev_int',
parameters=pfile,
output=oname,
overwrite=True)
10. Parameter File
17 - Geometrical conditions associated with Landsat 8
8 11 18.5030759012 -114.43339 50.29982 - The month, day (decimal hours) and Long/Lat of image centre
2 - Atmospheric mode
1 - Aerosols model
0 - Visibility [km]
0.112 - Aerosol model concentration
-1.5278 - Mean target elevation above sea level
-1000 - Sensor height, -1000 is a preset for satellites
117 - The band number associated with Landsat 8 Band 3
The atmospheric mode was set to ‘Midlatitude Summer’
The aerosol model uses the pre defined ‘continental model’
Aerosol model concentration was used instead of a visibility map
11. Composites and Subsets
Atcorr must be run on the entire scene because corrections must
acknowledge the entire histogram. To reduce processing time and
memory requirements, a subset was created for each scene from
a near infrared composite.
13. Change Detection – Image Differencing
Image Differencing is one of the simplest forms of change detection, and as such, is one of the
easiest to understand. By subtracting the pixel values of the more recent image from the older
image, areas of change can be easily highlighted.
Three classification methods were chosen for this analysis:
Normalized Difference Vegetation Index
Tasseled Cap Transformation
Principle Component Analysis
14. Normalized Difference Vegetation Index (NDVI)
Focuses on the presence and health of vegetation it is calculated using the red and near infrared
bands by dividing the difference by the sum of the two bands. The output has a range of -1 to 1.
Creates a stark contrast between a healthy
forest canopy (represented by high values)
and the exposed soil (characterized by low
values)
15. NDVI
This map shows the cumulative
change over the whole dataset.
Subtracting the NDVI scene from
2014 from the 2001 scene.
16. Tasseled Cap Transformation (TCT)
Transforms the original bands into a set of four bands designed to target aspects used to assess
vegetation health:
Brightness
Greenness
Wetness
and Haze
The Greenness band was the focus of this
study since it will produce the most contrast
between bare soil and vegetation
17. TCT
This represents the change
detected through the Tasseled
Cap Greenness band
differencing between 2001 and
2014.
18. Principle Component Analysis (PCA)
This method transforms the dataset into a series of bands, Principle Components (PC), which
account for the most variance possible. Each subsequent PC accounts for the maximum
remaining variance. A composite of these PCs creates an image defined solely by the amount of
change present in the image.
Performed on near infrared composites
which already enhance the difference
between soil and vegetation. The composites
chosen for image differencing are created
with the fifth, fourth, and third Principle
Components.
Left: 5,4,3
Above: 3,2,1
19. PCA
Using the third, fourth, and fifth
Principle Components, this map
is created by differencing the
PCA composites.
20. The Change
Creating a change mask by setting a threshold for each image differencing method allows the
user to focus on areas which underwent significant change. These thresholds were set based on
where change was known to have occurred.
The amount of change was calculated based on the number of pixels found in the image once
the thresholds have been applied
21. The Change
By using the amount of pixels in each change class, the amount of change was calculated two
ways. The absolute change was calculated from everything above the threshold. The change was
also calculated separately for classes which cover at least two hectares.
NDVI TCT PCA
Total 2388.608 2141.049 2353.261
More than 2 ha 2351.425 2117.373 2121.599
22. To do..
Object Based Image Analysis: Quite possible in GRASS, time restrictions meant this portion of
the project had to be abandoned, but significant progress was made.
Add i.landsat.acca, Automated Cloud-Cover Assessment, not used because it does not work for
Landsat 8 scenes.
Add interactive g.region command to prompt users for subset
23. GRASS GIS community
The GRASS Mailing list was important to the success of this project and I would like to thank
Micha Silver, Vaclav Petras, and Nikos Alexandris for their help.
Definitely bit off more than I could chew with this project,
Definitely bit off more than I could chew with this project,
It was created by the US army corps of engineers for terrain modeling and analysis, and as such it is perfect for this study.
Since GRASS is one of the Open Source GIS systems, it is available to anyone which makes it appealing to a lot of start-ups and non-profit organizations. Pair this with its ability to process raster data and the myriad of task oriented modules and you have a very appealing software suite. But just how do the applications compare to commercial software?
As a proof of concept I chose to analyze the Fording River mine for several reasons, first off it is one of Canada’s largest open pit mines and so detecting a yearly change with Landsat data seemed quite feasible, also, it is in my home Province of British Columbia.
As I mentioned, the data for this project was limited to seven Landsat scenes and one mosaicked ASTER DEM, which I have to credit Nasa for I think, all of this data was downloaded from the USGS Earth Explorer website. Using Landsat data was a conscious choice as it kept the whole project free, and would allow me to demonstrate how medium resolution satellite data can still be relevant. Especially for environmentally minded non-profits
Atmospheric correction is an essential part of any remote sensing application, unfortunately, it is also an expensive one. Creating atmospherically corrected images in GRASS GIS was one of the original reasons behind choosing this software.
Using the 6S algorithm to correct the imagery, i.atcorr has presets for most types of data as well as an option to build your own extension if the platform you are using is not included. Luckily, the entire fleet of Landsat satellites are included.
The i.atcorr module uses the 6S algorithm to correct the imagery. This algorithm requires several variables which are found in the parameter file. Currently, the script creates the parameter file for each individual scene by reading the metadata file and extracting the values needed. There is, however, much room for improvement notably in using a visibility map or calculating a more accurate Aerosol model.
Once atcorr has run successfully, the script creates both true colour composites as well as Near Infrared Composites. The True colour composite is created simply to see how well the i.atcorr functions while the Near Infrared was used to study the mine itself. The script subsets the NIR composite to reduce processing time and memory requirements.