GIS 4035 - Module 10: Supervised Classification for Thematic Mapping

This week focused on creating thematic maps via supervised classification. This involves indicating portions of a raster as training areas to feed to an algorithm. The training areas may be based on previously conducted ground survey, on interpretation of the rasters to be classified or by signature files created by experts and/or expert systems. In any case, the spectral information of the training areas is then used as a guide to classifying the rest of a raster image set, relying primarily on a comparison of multiple spectral bands to interpret their euclidean spectral distance. This later phrase refers to plotting out values of a raster pixel location on two or more bands and assessing how various pixels are grouped.

The map below was created by creating polygons based on spectral 'seeds' grown from a single pixel chosen from various locations on a raster set representing Germantown, MD. The seeds were expanded to polygons with a limited spectral variation from the initial pixel. Multiple locations were chosen for most of the classifications shown on the map, which were then merged after the classification was created using a Maximum Likelihood algorithm (using Bayesian methods to assess each pixels most likely classification). The spectral distance inset map shows the locations that were furthest, spectrally, from one of the classification classes, represented by the bright areas. The brightest areas tended to be fallow fields, which were deemed to have been correctly classified despite the indication of the distance map.

Thematic Map of Germantown, MD

GIS 4035 - Module 09: Unsupervised Image Classification

This week we dealt with unsupervised image classification. This is the process of using an algorithm to differentiate raster image pixels (often across and comparing multiple bands of an image set) in order to create classifications that produce (It is hoped) classifications of pixels that reflect actual ground features. The identification of groupings (and adjustment of classifications) are done after the classification is processed. This stands in contrast with supervised classification where an algorithmically based system is trained before classification by receiving input on areas across the image space that are grouped together into classifications.

The map below was produced by running an unsupervised classification on an orthophotograph of UWF's main campus. The initial classification was set to produce 50 classes. Once those were created, they were each identified as belonging to one of five very simple classes: Trees, grass, building/road, shadow or mixed. The mixed class was reserved for classes that were ambiguous as to their proper identification - encompassing a mix of grass and road. Once all the classes were identified, the classes were collapsed to encompass just those five and the map was created.

UWF Campus Thematic Raster Map