Wetlands Assessment
SAIT - Geos 451 - Data Capture II
The first part of this project was the classification of wetlands, as defined by open water, on a multi-spectral satellite image.
The area in question was just southeast of Edmonton, Alberta, Canada.
It was comprised of two large bodies of water, forest and a newly developed area in the southwest of the study area.
As the map extent is zoomed to the AOI, the current water bodies become quite visible in the basemaap imagery.
The goal of this project is to classify the image into more useable analytical data, and to specifically quantify the open water areas within the study area.
The multi-spectral imagery provided for the assignment was from the PlanetScope constellation, however here we see the Landsat layer from the Living Atlas.
The image consisted of four bands of information and was symbolized as a standard false colour image.
Swipe to compare the true color to the false colour imagery.
To begin classification a schema was required.
- For this project the schema was
- Water
- Grass
- Trees
- Developed
While this project only required the water to be classified, it was felt a slightly deeper schema would be more useful over time, and could prove useful within the LiDAR portions of the project.
For this project the Support Vector Machine(SVM) was chosen, a supervised classification algorithm.
SVM plots image bands vs each other to determine classifications.
For a supervised classification, training data is required to teach the algorithm about the informational classes.
The training areas were selected from representative areas within the four desired informational classes.
It is important to pick clean areas for the training areas, mixed, or dirty areas, can cause problems with the classification.
The classification was completed with the SVM algorithm and the default of 500 points.
Once completed a stratified random set of assesment points were created and truthed with the PlanetScope imagery (viewed as true colour).
Swipe to compare the classification to the base map imagery*.
*Classifications were loaded with simplified geometry for this presentation.
From these points, a confusion matrix was created showing an accuracy of 93% and a kappa of 0.90.
With these results, this classification was determined to be acceptable.
Next the water was extracted as a seperate polygon, that was used later during the LiDAR assesment...
And then clipped to the final AOI to determine the areal coverage of the water.
The second aspect of this project was to construct a Digitial Elevation Model over the area with a set of supplied LiDAR data.
The LiDAR dataset presented the first real issues with this project. It was provided without any supporting information, and while it loaded into, what was believed to be, the correct position using the same coordinate system as the provided study area feature class (NAD 1983 CSRS UTM Zone 12N). Its alignment with the base imagery lent further evidence to this being correct, but the provided was contacted to confirm this.
This swipe provides a comparison with the LiDAR and basemap imagery.
It was decided to move forward with the project with the only real unknown being the vertical datum. Other LiDAR datasets in the area use the Canadian Geodetic Vertical Datum of 1928 (CGVD 1928), and so it was chosen with the information noted in the final report.
The project required a DEM to be created from the LiDAR data.
All non-ground classified points were turned off and the point cloud was inspected, symbolized as a surface elevation.
It was noted there were a few anomalies in the developed area to the southeast, where it should have been flat.
The LiDAR points were classified to 2m elevation intervals. This allowed points with sudden changes to be seen and potentially reclassificed as ground points.
Swipe to compare the surface to the 2m point classifications.
Not all of these flagged areas seem to be causing issues. Some may be true ground changes. The larger the areal extent these features have, the more likely that is.
Swipe to compare the flagged points to the surface anomalies.
Within this LiDAR pointset, these two areas were re-classified to create a more accurate DEM of the area.
Once these had be re-classified, the resulting symbolization as a surface is much more consistent.
Swipe to compare the original surface to the updated surface.
It should be noted that within this LiDAR data set, many buildings have been classified as "High Vegetation." To create any surfaces beyond the DEM, these will need to be addressed.
Houses classified as vegetation are clearly visible.
It was also noted that all of the ground points on the lakes were classified as ground and not water.
While this shouldn't have an effect on the creation of the DEM, it was decided to reclassify them as water.
Manually reclassifying the water points was an impractical way of dealing with this.
To deal with this the water polygons created earlier with the SVM classification were employed.
Using the water polygon, all points that landed within the polygons were reclassified as water.
It should be noted that noise points were not reclassified, but ground and vegetation points were.
This provided a fast and efficient way to assign a proper classification to the water points.
With the LiDAR point cloud examined and adjusted where necessary, the next stage was to generate the DEM and examine it.
When the DEM was generated it was important to select both the ground points and water points, and to make sure noise was turned off. From within tool, the defaults were used for the most part. One change was select "lowest" value when determining points to use for the interpolation of the grid.
DEM compared to base imagery.
Q/C for the DEM was completed with the following steps.
First a basic visual inspection looking for any obvious bullseyes appearing in the grid.
The next stage was also a visual inspection, but this time a shaded relief was applied to the DEM to make sure it looked like plausible terrain.
With this applied it was easy to see lagoons, roads and ditches in the developed areas.
The final examination was to look at the elevation histogram. As this is east of Edmonton, a great deal of variation was not expected.
While the right side of the chart wasexpected, the left side was not. It was suggested this was due to the lakes taking up a good portion of the study area and having a water level lower than the surrounding terrain.
To check this the CON tool from ArcGIS Pro was execute and classified elevations above and below 757m to a 0 or a 1.
The resulting grid confirmed the theory that the lakes were causing the histogram distrubution.
Using a four-band PlanetScope image, a supervised classification schema was created that allowed the classification of developed, tress, grass, and water. With this information the study was able to assess that the study area was composed of 90 hectares of wetlands. In addition to this information a high-resolution digital elevation model was created from a LiDAR survey that was flown over the area.
These two products can be used individually to assess the area, or they could be combined to create visualizations of the area, or slope areas to be used as an input into water flow and flooding analysis in the area.
