Photo Interpretation and Remote Sensing - Supervised and Unsupervised Classification

 

The classified map of Germantown, showing land cover/a degree of land use based on ERDAS extrapolation of manually selected signatures. The Distance map shows areas where the classification may be uncertain.

This week's lab focused on conducting feature classification on aerial imagery using a computer-driven tool (unsupervised) and classification based on analyst inputs (supervised). Both methods require manual retooling before or after, and adjustment and manual assessment of the band representations to check that all features are displayed as clearly as possible. In particular, I used histograms and spectral plots to check for spectral confusion in the pixels I had classified - this is where one pixel may have a value that has been classified in two groups at the same time. At the end, I ran the entire process on an image from Germantown, Maryland, to practice supervised classification and the resulting recoding and display of land cover in Germantown.  

Photo Interpretation and Remote Sensing - Spatial Enhancement and Band Indices

To identify this feature we were given information about it's appearance in different spectral layers. From there we had to compare layers and select appropriate band combinations to highlight the feature and make it distinct from it's surroundings. 

 In this week's lab we examined techniques for manipulating the display of aerial imagery based on which areas of the visible spectrum we wished to emphasize. Using ERDAS and ArcPro, I explored the distribution of reflectance in satellite imagery from USGS, and adjusted the band assignments in the visible spectrum to identify different features. Much like the True Color and False Color Infrared from previous labs, the amount of different wavelengths reflected depends on the type of surface encountered by electromagnetic radiation. By changing how the bands of visible light are visualized, analysts can discern between different types of surfaces, such as water, compact ice, new ice, healthy vegetation, and dry vegetation.    

Photo Interpretation and Remote Sensing - Working with EMR and Digital Image Processing

 This week we explored and learned to work with ERDAS software, which is used for importing, displaying and analyzing images from aerial sensors. Though this software has been around for some time, learning the quirks and best practices was essential for beginning to analyze aerial images outside of ArcPro. In addition to working with ERDAS, we reviewed the types of electromagnetic radiation, and how wavelength, frequency, and energy are related as well as how this may influence a sensor's image of the Earth. By manipulating images in ERDAS and ArcPro I was able to discern and utilize different color models to focus on different points of image analysis. 

This map was created by selecting a subset of an aerial image from ERDAS and successfully importing it into ArcPro for further analysis and cartography. Note that "Cloud" is a type of cover, as it interferes with the sensor ability to see the ground.  

Photo Interpretation and Remote Sensing - Manual Land Use/Land Cover Assessment

Digitized polygons classified by land use or land cover.

For this module we each classified land use and land cover from an aerial photograph for a portion of Pascagoula, Mississippi. We were encouraged to classify land cover and land use down to a Level II classification, which included delineations such as Residential, Commercial, Evergreen Forest, Bays and Estuaries, but did not specify further to type of home, species of tree, use of waterway etc. 

In this exercise I used the photo interpretation techniques from the previous lab to make determinations on structures, vegetation, and feature boundaries. I determined an appropriate minimum mapping unit (MMU), digitized polygons based on my interpretation, and then conducted an accuracy analysis by comparing 30 randomly sampled points to the features at those locations in Google Maps/Streetview. Of 30 sampling points, only 2 were incorrectly classified, for an estimated accuracy of 93%.