GIS 4035 Module 10: Supervised Classification

Supervised classification and spectral distance file of Germantown, Maryland.

The task for module 10 was to utilize ERDAS Imagine to create, analyze, and edit supervised classifications of multispectral images. We began by reviewing the fundamentals of the Signature Editor tool (to create and edit the training samples used for a supervised classification) and creating an AOI (area of interest) layer. We then used the Inquire tool to locate specific points on the image on which we created polygons to capture to areas to be used as the spectral signature of a class. We also reviewed the use of the Seed tool; this tool automatically creates a polygon capturing pixels of a set spectral distance from the initial point based on input parameters. I preferred using the Seed tool to manually inserting polygons as it gives more control over keeping each class distinct. We also spent some time analyzing the histograms and mean plots of spectral signatures. These tools give use a more precise way to ensure our classifications capture unique features and are not spectrally confused.

The assignment required us to use all the tools reviewed in the module to create our own supervised classification of an image of Germantown, Maryland. We were given coordinates of three urban features, two fallow field features, four agriculture features, and one each for grasses, deciduous forests, and mixed forests. We also were to create signatures for water and roads. With this completed, we then needed to recode the number of classes down to eight, consolidating the categories with multiple classes into one each. We also needed to include the spectral distance file in our final map; this image shows us at a glance the pixels in the image that are furthest away from any of our classes. The final layout was completed in ArcMap.

GIS 4035 Module 9: Unsupervised Classification

Five category unsupervised classification of the UWF campus based on true color imagery.

The ninth module of the course guided us through performing unsupervised classifications of imagery in ArcMap and ERDAS Imagine. An unsupervised classification takes basic guidelines from the user (such as the number of desired categories) and creates categories based on the appearance of each pixel. A perfect classification would, for example, classify all water in a category, all trees in another category, and so on. There has likely never been a perfect classification, however, so the resulting classified image must be edited to better capture the desired categories.

In order to create the above classified image of the UWF campus, the original true color image was run through Imagine's Unsupervised Classification tool to create fifty categories. The resulting image (not shown) looked very similar to the original image. We then reclassified each of the fifty categories into the five classes of trees, grass, buildings/roads, shadows, and mixed (grass/urban). The main source of error was the overlap of bright grass and ground areas to some urban areas; this created the need for the "mixed" class. 

GIS 4035 Module 8: Thermal and Multispectral Analysis

Imagery combining the thermal infrared band (6) with the shortwave infrared bands (5 and 7) to highlight fires.

The eighth module of the course focused on using thermal infrared imagery to extract data unavailable from other EMR wavelengths. The general concepts and tools were similar to previous exercises, yet thermal infrared imagery presents unique challenges. First, thermal infrared EMR is emitted, not reflected; the amount of thermal infrared EMR emitted by a feature is a combination of the amount of energy absorbed, the feature's composition and surface characteristics, and the sensitivity and exposure length of the image (among other factors). In simple terms, thermal infrared EMR reflects temperature, but we cannot assume a direct correlation without calibration.

Our deliverable for the week was more open-ended than usual; we were to choose an area or feature in one of the two composite images created for the module and create an image that highlights the chosen area or feature. The thermal infrared band needed to at least be used to identify the feature even if it wasn't used in the final layout. While we had already identified select fires in the imagery, I was struck by how defined the fires' core extents appeared when the thermal infrared band was combined with the shortwave infrared bands. After applying a Gaussian stretch in Imagine and a minimum-maximum stretch in ArcMap, the fires popped out from the background significantly. This did allow me to identify a third, small fire southeast of the largest fire that I had not noticed previously.

I enjoyed the experimental nature of this module and learning of the unique data one may extract from thermal infrared imagery. I am still not confident in my ability to manually manipulate histogram breakpoints to achieve fruitful results, but this module did help me significantly understand the areas I need to work on.

GIS 4035 Module 7: Multispectral Analysis

Feature 1 multispectral analysis: Deep water.

Feature 2 multispectral analysis: Snow.

Feature 3 multispectral analysis: Shallow water.

Our task this week was to practice using different methods of multispectral data analysis to find three features with particular EMR signatures. The methods practiced in the assignment included histogram analysis and manipulation, displaying different combinations of bands as composites, creating indices based on different bands (e.g., Normalized Differential Vegetation Index), and using the Inquire Cursor to get detailed data on individual pixels. In order the find the three features, I primarily used the Inquire Cursor method based on histogram analysis.

The assignment helped me greatly in understanding histograms, although I am still not confident in manipulating them for a better image. By looking at the histograms of each band, one gets a good sense of how the data will display. I also appreciated the practice in combining bands into custom composites to highlight particular data. This was something I did not quite grasp in ArcMap, but the methods in Imagine made the concepts clearer.

GIS 4035 Module 6: Spatial Enhancement

Landsat 7 image run through Fourier transformations, sharpening, and statistical filter in ERDAS Imagine.

Past modules in this and other courses have mentioned that GIS operators often must correct acquired imagery for various errors prior to using it for their analyses. This module introduced us to some of the ways such corrections can be done. We explored the tools available in both ERDAS Imagine and ArcMap, including Fourier transformations, high pass filters and low pass filters. Low pass filters generate output that appear smoother and less detailed than the original image; noise in the image is also removed, according to the size of the kernel chosen. High pass filters create high contrast and noisy output that highlights the edges of features.

For the assignment, we were to experiment with Imagine and ArcMap to create an enhanced version of a Landsat 7 image that minimizes as much as possible the scan line corrector failure striping without removing too much detail. The assignment walked us through using the Fourier Transform Editor tool as the initial step in the process. While I experimented with different wedge placements in the editor, I was unable to create an image that was significantly better than my first attempt. After experimenting with various filters and settings in Imagine and ArcMap, the final image used above was the result of sharpening and a statistical filter in Imagine. Other filter and setting combinations created images similar to this, and a few were significantly worse. I did not experiment much with the histogram, however; I hope to learn more about histogram manipulation in future modules.

GIS 4035 Module 5a: Introduction to ERDAS Imagine and Digital Data

Subset of a classified image selected from ERDAS Imagine and exported to ArcMap for layout finalization.

Our module this week was divided into two topics. First, we dove deeper into the details of the electromagnetic spectrum, including the relationship between wavelength, frequency and energy along with how to calculate each. While we need not be physicists to use remotely sensed data, we do need to have an understanding of the structure of the electromagnetic spectrum and the type of data we can extract from different wavelengths.

Our second topic was an introduction to the ERDAS Imagine program. I am glad to finally be learning this program, in part due to the positive things I've heard of it from others but also as a change of "scenery" from ArcMap. Our deliverable (above) was simply a selected section from a classified image exported and finalized in ArcMap. Most of the assignment was designed to get us familiar with the program for use throughout the semester and, hopefully, our careers.

An issue that often arose as I worked with ERDAS Imagine was the difference between Imagine 2011 (the basis for the assignment) and Imagine 2014. The majority of changes I encountered were simple to recognize and overcome. The only major problem I had was in exercise 3; the subset image I created using the inquire box never updated the area attribute. I made several attempts without success. Most likely I was in error but I could never locate the source. In the end, I calculated the updated area totals in ArcMap using the Count field of the attribute table.

I enjoyed working with Imagine as the program seems to be much more efficient than ArcMap in simple manipulations of raster data (such as panning and zooming). I am looking forward to learning more about what the program can do.

GIS 4035 Module 4: Accuracy and Ground Truthing

Land use land cover classification of Pascagoula, MS with accuracy of thirty sample points symbolized.

Our task for week 4 of Aerial Photo Interpretation and Remote Sensing was to check the accuracy of last week's efforts to classify an aerial photograph. The best methods for ground truthing such a classification involve direct testing sample sites in the field, but the online nature of the course prevents such tests. Instead, we utilized Google Maps (and especially Street View) to test the classification accuracy of thirty sample sites. I roughly followed a stratified random sampling pattern based on classification category. However, small or homogeneous categories were given fewer samples (e.g., the large body of water in the west, the cemetery) while large, heterogeneous categories were allotted more samples (e.g., residential, commercial and services). The accuracy of my classification turned out to be about 73%. The main error sources were misclassifications of bodies of water and of forest cover. While my misclassification of Krebs Lake as a bay could, perhaps, be forgiven, my misclassifications of forest cover were due to not correctly distinguishing deciduous and evergreen trees. As a result of this assignment, however, I believe I would have a higher accuracy percentage on a similar aerial photo.

GIS 4035 Module 3: Land Use Land Cover Classification

Aerial photo of Pascagoula, MS with land use/land cover classification overlay.

This week for our Photo Interpretation and Remote Sensing lab we were given the challenge of classifying an aerial photograph based on land use and land cover.  While we were not required to get too detailed with our classification, we were required to classify everything.  We were to create a new shapefile and then create polygons over each classification.  I extensively used the "trace" and "clip" tools from the editor toolbar as I classified the image to ensure everything was properly classified with no gaps remaining between polygons.

This was a rather challenging lab assignment, due both to its intrinsic difficulty as well as to the catastrophic failure of my computer.  It was poor timing, but I believe I was successful with the above map in the end.

GIS 4035 Module 2: Visual Interpretation

The first two modules of our Photo Interpretation and Remote Sensing course focused on the background and fundamentals we need to build on throughout the semester.  In particular, the two maps above show our experimentation in module two with the different criteria by which we interpret remotely sensed data (specifically aerial photography).  In the first map, we were required to identify five categories of tone (from very light to very dark) and five categories of texture (from very fine to very coarse).  Our next task, reflected in the second map, was to use four additional identification criteria (shape and size, shadow, pattern, and association) to interpret the photograph.  The most difficult aspect of this task was attempting to consider each criterion in isolation; in normal interpretive practice we obviously use all available criteria.  However, considering them singly was helpful; I had not considered how useful shadows might be to correctly interpreting aerial photography.

These introductory modules provided a good foundation in the science of remote sensing and the basics of interpreting the resulting data.  I look forward to learning much more throughout the semester.