GIS4035 Module 10: Supervised Classification

ERDAS Imagine was used to develop two different land use maps during this module, with the method of supervised classification being the focus of the exercises. In areas of interest (AOIs), polygons were drawn around representative pixels of different land uses or land covers, or a polygon was "grown" from a seed pixel to encompass a sample of a land use or land cover area. These samples ("signatures") were then used to classify the remainder of the area covered by the image. The classes were critiqued using histograms and Mean (statistically speaking) Plots, and additional signatures were gathered as needed. (That was the intent anyway. Initially I thought the classification was looking fairly decent after some adjustments, but then I noticed something fishy about the roads, including the acreage attributed to roads. Obviously there is some other feature lumped in with roads which needs to be evaluated further.) After the additional signatures were incorporated, the histograms and Mean Plots were used to critique the classification again. Bands 3, 4, and 5 had the least amount of overlap, and so various combinations of them were used to edit the signature colors before finally settling on R3, G4, B5 for the Germantown map shown below. The image was classified with this band combination, and a distance file was created. The purpose of the distance file was to highlight specific areas which might have been misclassified. Similar classes were merged by the process of recoding, and areas were calculated.

Supervised Land Use Classification
Germantown, Maryland

GIS4035 Module 9: Unsupervised Classification

Unsupervised classification of land cover allows the software to group pixels according to similarity (in this case creating 50 different classes). While this simplifies the classification process, the software is unable to distinguish different surfaces which reflect similarly. Because of this the results of unsupervised classification need to be reviewed and adjusted by an analyst. For instance, some grassy athletic fields (identified by soccer nets and light fixtures) reflected similarly to paved surfaces and were classified accordingly during the process of unsupervised classification. After review of the classes created with unsupervised classification, the pixels grouped like this were manually assigned to a more appropriate class. The reviewing process was made easier with the use of the Swipe, Flicker, and Blend tools along with the use of color-highlighting of pixels. 100% accuracy is not expected with classification. Some pixel groups did not clearly belong to one particular class; these pixel groups were assigned to a Mixed class. A few dormitories appear to have sod roofs, but this is due to those pixels being classified as grass like the overwhelming majority of pixels in a particular class.

After the image was reviewed and classes manually adjusted, the classes were merged to reduce the number to five using Recode. In the Recode Table, an area column was created from which percentages of permeable and impermeable surface areas were calculated. In this case, the information was transferred to Excel for computations. The resultant map shows the recoded image along with acreages for each of the classes of land cover and the percentages of permeable and impermeable surface area.

Unsupervised Classification of Land Cover and Surface Area
University of West Florida Campus

GIS4035 Module 8: Thermal & Multispectral Analysis

Building on what was practiced in Module 7 Lab, additional experimentation was done with different band combinations and adjustments of breakpoints after composites were created from layers in both ERDAS Imagine and ArcMap. Thermal imagery allows the comparison of different features by the energy emitted. This is dependent not only on the material but also on season,  time of day, and moisture content. The following map was created after much experimentation with different band combinations, including some adjustments of breakpoints. The goal was to obtain an image and band combination in which a selected feature really stood out.

Thermal and Multispectral Analysis
Feature: Lake

GIS4035 Module 7: Multispectral Analysis

Map 2: Snow identified with band

combination Red: 5, Green: 4, Blue: 3

Map 1: Water feature identified with band

combination Red: 4, Green: 3, Blue: 2

Map 3: Lake identified with band
combination Red: 7, Green: 5, Blue: 3

This lab explored different ways to identify features using ERDAS. After reviewing histograms to identify patterns and shapes in the data, the imagery was viewed in grayscale to look for dark and light shapes as well as patterns. EMR bands were manipulated to emphasize different features in the imagery. Trying to find the right band combination to help make a feature noticeable was a challenge, but the band combination references posted on the discussion board were helpful. Exact pixel values of specific areas were obtained by using the Inquire Cursor. Map 1 shows the use of TM False Color IR to highlight bodies of water. Map 2 utilizes a Short Wave Infrared Color Composite of 5-4-3 to illustrate snow. The final map, Map 3, uses a band combination of 7-5-3 to make certain water features, such as this lake, stand out.

GIS4035 Module 4 - Ground Truthing and Accuracy Assessment

Land Use / Land Cover Classification Accuracy
Pascagoula, Mississippi

This module included a lot of completely new material for me. Although the lab exercise was straightforward and rather easy, the preparatory information was rather overwhelming. The lab exercise itself involved using last module's classification map and adding sample points. The actual land use or land cover for these sample points was determined using Google Maps Street View where possible or zooming in on Google Maps aerial view. This was compared to the classification selected last week for the same points. The results were tabulated for accuracy. Best accuracy for this map was determined for non-forested wetlands, probably because they were so easy to distinguish. Barren lands had lowest accuracy because of having classified open areas as barren lands when a better classification would have been herbaceous rangeland. The lab itself was a fun exercise. The resultant map with sample points is shown here.

GIS4035 Module 3 - Land Use / Land Cover Classification Mapping

Land Use / Land Cover Classification
(Pascagoula, Mississippi)

Creating a classified land use and land cover (LULC) map from an aerial image was the goal for this lab. After a polygon feature shapefile was created and additional fields were added to the attribute table for LULC code and description, Editor was used to create features by digitizing polygons for different categories of land use and land cover. Vertex and edge snapping were used to digitize adjacent polygons, but eventually some issues involving the Snapping Toolbar were encountered that could not be resolved promptly. The ability to clip out polygons from within other polygons is a skill that will be handy for future projects. The identify tool was very helpful while clipping out polygons.

Although time-consuming, especially because of the scale and zoom level I selected (it seemed like a good choice initially), this was a very enjoyable lab to complete.

GIS4035 Module 2 - Aerial Photography Basics & Visual Interpretation of Aerial Photography

Map 1 - Degrees of Texture and Tone in an Aerial Image

Module 2 focused on the basics of aerial photography (vantage points, cameras, filtration, and films) along with how to use recognition elements to interpret aerial imagery. Recognition elements include location, tone and color, size, shape, texture, pattern, shadow, height/depth, and site, situation, and association. The first map depicted here focuses on the identification of different degrees of tone (from very light to very dark) and texture (from very fine to very coarse). Degree of tone indicates the degree of energy reflected by surfaces. Degree of texture is relative for any particular image. For this specific image, the texture ranges from very fine for water to very coarse for residential neighborhoods. For another image, these descriptions may vary significantly dependent on the nature of the subject of that particular image.

Map 2 - Features Identified by Visual Attributes

The second map focuses on utilizing other recognition elements to identify specific objects in the image. Shape and size were utilized to identify roads (long and linear), a pool (roman shape with rectangular diving board), and cars (rectangular shapes associated with large, open parking area). Shadows were used to identify a water tower (long, thin shadow with flared bottom and topped with bulbous shape), trees (slender trunks topped by clusters of fronds/branches), and a warning siren (horn-like shadow at top of tall, slender shadow which was distinctly different from the shadows of other poles in the image). Objects identified by patterns included parking spaces (herringbone pattern), utility poles (regularly spaced tall, thin objects with cross bars at the top), and houses (buildings of similar size, height, and spacing with a driveway to each structure). Some features were identified by their association with other features. This included the pier (a long, thin structure projecting from the shoreline out into the water with a shadow indicating height above the water) and a driveway (a short linear object which did not cast a shadow and leads to a house).

From personal and professional (surveying) use of aerial imagery, identification of objects is a rather comfortable practice. An area of growth for me was demonstrated in the third part of the lab which focused on comparing objects between a true color image and a false color infrared image. Along with learning more about the use of recognition elements in identifying objects in imagery, learning to interpret information from natural color and false color IR images is a skill I look forward to developing.