Module 9 - Special Topics in GIS - Accuracy of DEMs

This week in Special Topics in GIS, we learned about the sources of error involved in creating DEMs, and conducted two analyses on elevation data.  In the first analysis, we compared elevation and land cover class data collected by hand in the field.  This elevation data is considered to be the more accurate, or reference data.  The sample points were divided among five classes of land cover, shown in the table below.  We overlaid a LIDAR image on the field sample point map, and compared the LIDAR-obtained elevations at points corresponding with the field-measured points.  The table below shows the accuracy results for the five types of ground cover, separately and together.

Bare earth and low grass, predictably, have the lowest land-cover, because there are few obstructions in the LIDAR view.  Fully-forested land has the highest error, which is logical because the canopy of trees will interfere with the LIDAR view.  The range of error values was also greater for forest.  This might be from the variability of the forest cover.

Another interesting outcome of this analysis was that while the composite error appeared to be fairly unbiased, most of the component land cover data types did show various strong biases.  Land cover types a, b, and c (bare earth and low grass, high grass, weeds and crops, and brushland and low trees) all produce LIDAR elevations that are more often higher than the field measured sample elevations. 

Although fully-forested land cover has the highest error, it is also shows the least bias in elevation error. Urban land cover has a very strong negative bias in error: LIDAR elevations are nearly always lower than the field-sampled reference elevations. It just so happens that the combination of these 5 types of data produce a more-or-less unbiased composite.  This would be very misleading if the results were not closely analyzed by ground-cover type. 

Module 8 - Remote Sensing - Thermal and Multispectral Analysis

Thermal and Multispectral Analysis to Identify a Geographic Feature

This week, we added Thermal Infrared, or TM band 6, to our multispectral analysis toolbox.
Unlike the near-infrared and mid-infrared wavelengths, which are reflected from the surface of the Earth, back to the sensors, thermal infrared energy is actually emitted from the body of the Earth itself.  When shorter-wavelength energy hits the earth, some of it is converted to heat energy, and that is what is emitted back to the thermal infrared sensors on platforms such as LANDSAT.

Sources of heat, such as fires, can be identified by the thermal infrared band, but cool features can be as well, from their anomalously low emission of thermal infrared energy: they are cool.  Water bodies are examples of features such as this.  In the map above, a band combination of Near Infrared (band 4: red), Visible Green (band 2: green) and Thermal Infrared (band 6: blue) were used, along with several other band combinations, to identify the dark green grid-like feature above as a collection of ponds or flooded fields.  In the image pictured here, it can be seen that the ponds reflect varying degrees of visible green light, depending on their depth.  The fact that they are a fairly dark green compared to the river to the west show that they are filled with relatively clear water.  Whether they are dark or light green also depends upon their depth.  Almost all Near Infrared wavelengths are absorbed by the water, and they emit very little thermal infrared energy.  This is shown by the very low amounts of blue and red displayed by the ponds in the image.  The area to the east of the ponds emits considerably more thermal infrared energy, as shown by its bright blue color, while the vegetated areas shown by pink and red around the ponds reflect much more Near Infrared energy.

Module 8 - Special Topics in GIS - Surface Interpolation

Spline Interpolation with input data points

This week we worked with several different types of interpolation, that process by which new values can be estimated between known values. I could have  devoted several weeks to this topic, to get a good grasp of which types of interpolation are best for various applications and how best to use them.  I felt that this week, I barely scratched the surface, as it were, in understanding this topic.

The map at left is an example of Spline interpolation of elevation (white and purple are highest, turquoise is lowest). Locations of input data are shown as black dots. This is a method which provides a smooth result, and with which it's possible to extrapolate beyond the available data values and study area.  Newly interpolated values can be more liberally influenced by input data in their general neighborhoods.

Legend (relative Elevation)

IDW Interpolation

This map is an example of IDW, or Inverse Distance Weighted interpolation of the same data points.   It is not as smooth as the Spline method, because interpolated values get less and less influence from the input data points with increased distance; that is, less weight is given.  The IDW has a much rougher and spottier aspect because the influence of a single data point rapidly diminishes with distance away from it.  All of the small dots in this map are centered around input data points (not shown in this view) that are isolated enough from other points that their weight is felt only in their immediate vicinities.  The more extensive areas of color are the result of the combined influence of input data that are more densely arrange,  and of similar values.

Both IDW and Spline interpolation honor the input data, that is, the data points retain their original values.

Module 7 - Remote Sensing - Multispectral Analysis

Three examples of ground features emphasized by Spectral Band Combinations.

This week in Remote Sensing, we learned about multispectral analysis, specifically with Landsat TM 5 imagery in Washington State.  The image we looked at has 6 spectral bands that can be examined one at a time with gray scale imagery, or in combinations, displayed in Red, Green and Blue (RGB).
The bands in this Landsat imagery are 1, 2, and 3, which are the visible wavelengths of red, Band 3,or Near Infrared (NIR), and Bands 5 and 7, which are Middle Infrared.  Any combination of 3 bands can be used, represented by the visible red, green and blue in the display.  Each of the example features on the left is shown in a different combination, which best emphasizes deep or shallow water and snow, for example.

Module 7 - Special Topics in GIS - TIN's and DEM's

Symbology in a TIN elevation image of a landscape, with 5x exaggeration of relief.

This week in Special Topics, we began a 3-module unit on surfaces. In the first lab, we had an introduction to DEM's or Digital Elevation Models, and TIN's or Triangulated Irregular Network Surfaces.

DEM's are raster-based, and made of regular grids of pixels, each of which holds a value.

TIN's however, are based on vector data.  Data points values can be turned into a TIN by means of triangulation: the points are connected to their neighbors by edges, and the edges form triangles, which make up facets in the surface in question.  Vector lines and polygon boundary lines can be incorporated into a TIN, to show the positions of roads, rivers, lake shores, and other significant features that may affect the shape of a landscape.  These linear features are included in the network of triangles, with series of small triangle edges defining the linear features.

In an elevation TIN, each triangle or facet has a uniform slope and aspect, while elevation changes across the triangle. In the figure above, different degrees of slope are shown by different tones of gray.  The surface can also have contours overlaid, as well as the original mass points (red).  Aspect (azimuthal direction) can also be symbolized.

The figure above shows a close-up example of a TIN.  Many features of its can be shown in the symbology, including the nodes (points), edges of the triangles, slope, aspect, and surface.

ArcGIS uses Delaunay triangulation.  This process avoids long skinny triangles by requiring that a circle that includes the three vertices of any triangle does not enclose vertices of  another triangle.

Module 6 - Remote Sensing - Spatial Enhancement

Enhancement of Landsat 7 Imagery with Fourier Transform and
3x3 sharpening, ERDAS Imagine v.2014

In this week's Remote Sensing Lab, we tried out various digital filters in order to enhance imagery.  High-pass filters can be used to bring out more detail and contrast in an image, while low-pass filters smooth the imagery and emphasize larger-scale features.

The larger image to the left is an example of how imagery that has been corrupted by missing data can be improved.  We used Fourier Transform in ERDAS Imagine to partially blend the values of pixels in the black stripes that show up in the original image (left) into the pixels surrounding them.  This processing was followed by convolution filtering in ERDAS.  Here, we make a "kernel" of 8 pixels that will surround each pixel in the image in turn.  In filtering with a kernel, the values of the 8 pixels surrounding each pixel are averaged, and that value is then applied to the center pixel.  This serves to de-emphasize the black stripes.  Finally, radiometry was adjusted in ERDAS Imagine, by way of the LUT histogram breakpoints.

In the final enhanced image, the stripes are fainter, but still visible.  The main disadvantage to this type of processing is that the image can become very blurry and lose a lot of detail if the analyst is not careful.

Module 6 - Special Topics in GIS - Location-Allocation Analysis

In this final week of our Network Analysis unit, we practiced Location-Allocation analysis, in which supply points are allocated to demand points.

A hypothetical retail chain has 22 distribution centers throughout the United States, to supply to more than 4,300 customers.  It has divided the customers into more than 200 market areas, shown by the smaller polygons in the map at left.  However, the company wants to determine if there is a more efficient scheme for supplying the market areas from the distribution centers.  With the help of a Network Dataset of roads, each customer is allocated to the ideal distribution center.
In this study, we did not want to go to the trouble to reassign the individual customers to the distribution centers, so we looked at a way to reassign the market areas, based on if the majority of customers within each market area had been reassigned to a new distribution center.  If the majority are better served by a new distribution center, all of the customers within a market area will be reassigned.  We had only a few market areas (less than a dozen) of the 227 total reassigned.  This scheme does not insure that every single customer is being served by the nearest distribution center, but it does make for less reorganization for the company in terms of reassigning customers to new distribution centers.