Remote Sensing: Module 10 - Supervised Classification

     This week is a continuation of techniques in automated digital image classification. Specifically, supervised classification is covered. Unlike unsupervised methods, this method requires the use of training sites (a priori knowledge) to guide classification. It also uses statistical methods as opposed to Euclidean distance data to assign pixels to information classes. To practice this technique, the lab assignment walks through a supervised classification the process for which is then reproduced to compile the weekly deliverable. The classification was performed in ERDAS Imagine and the map deliverable was made in ArcMap.
     The general method is to import an image, create AOI features surrounding training sites (using the region growing tools and drawing tools), record the signatures from these AOIs in the Signature Editor, perform the supervised classification (via the Raster tab), and then recode this information to represent final class designations. As a result of running the supervised classification, a distance image is also created. This image along with examining comparative class histograms and mean plots (options in the Signature Editor window) examine the accuracy of the classification.
     My map below shows Germantown, Maryland classified via a supervised classification. The Euclidean distance image is included as a reference to the degree of error. (The brighter spots indicate areas that are likely classified incorrectly.) Eight classes are ultimately used to represent the different land uses/land classes. Areas are also included and noted in the legend. The original band combination used is 5, 4, 6 as I felt these bands had the best separability between spectral signatures. As an aside, one of the more difficult things was creating training sites for roads. The borders between roads and everything else for in this image were quite pixelated. As a result, there is a fair amount of error in that class since my training sites may have included the occasional pixel belonging to a different class. When I was growing areas of interest for some of the other classes I also set higher distances between pixels to capture some of the variation in that feature and as a result introduced error into certain classes (particularly the fallow and agriculture classes) and that is validated by the bright spots in the distance image inset.

Map 1 An exercise in automated classification using a supervised classification technique

Remote Sensing: Module 9 - Unsupervised Image Classification

     This week we delved into digital image classification. Up first is unsupervised classification, the following module (10) covers supervised classification. Unsupervised classification is an iterative (that is, until a threshold is met) procedure involving the clustering of pixels (with similar spectral reflectance) into classes (a parameter determined by the user). Supervised image classification uses a priori knowledge to guide classification (by way of 'training sites').
     In this lab we took a high resolution aerial photograph of the University of West Florida campus and performed an unsupervised classification (using the ISODATA algorithm) in ERDAS Imagine. The result of which is a thematic raster with a lower spectral resolution than the original photograph. This image was classified into fifty categories. It was our job to reclassify and reduce these categories into one of  five choices: trees, grass, buildings/roads, shadows, and mixed (for pixels that fall into more than one category). This was done by selecting pixels and reassigning them into one of the categories and giving them a designated color. Once all fifty classes were reclassified, the Recode tool was used to reduce the number of classes into the final five. Descriptive data (in the form of areas and percentages) was calculated for each category and for permeable/impermeable surfaces. The map deliverable below displays the classified image and accompanying data.
     ArcGIS was used to make the final map product but it can also run an unsupervised classification by using the Iso Cluster tool and the Maximum Likelihood Classification tool. It is  necessary, however, to possess the Spatial Analyst extension.

Map 1 Unsupervised Digital Image Classification of the University of West Florida Campus