Supervised Classification - Lecture Material - Completely Remote Sensing tutorial, GPS, and GIS - facegis.com
Supervised Classification

Supervised classification is much more accurate for mapping classes, but depends heavily on the cognition and skills of the image specialist. The strategy is simple: the specialist must recognize conventional classes (real and familiar) or meaningful (but somewhat artificial) classes in a scene from prior knowledge, such as personal experience with what's present in the scene, or more generally, the region it's located in, by experience with thematic maps, or by on-site visits. This familiarity allows the individual(s) making the classification to choose and set up discrete classes (thus supervising the selection) and then, assign them category names. As a rule, the classifying person also locates specific training sites on the image - either a print or a monitor display - to identify the classes. The resulting Training sites are areas representing each known land cover category that appear fairly homogeneous on the image (as determined by similarity in tone or color within shapes delineating the category). In the computer display one must locate these sites and circumscribe them with polygonal boundaries drawn using the computer mouse. For each class thus outlined, mean values and variances of the DNs for each band used to classify them are calculated from all the pixels enclosed in each site. More than one polygon is usually drawn for any class. The classification program then acts to cluster the data representing each class. When the DNs for a class are plotted as a function of the band sequence (increasing with wavelength), the result is a spectral signature or spectral response curve for that class. The multiple spectral signatures so obtained are for all of the materials within the site that interact with the incoming radiation. Classification now proceeds by statistical processing in which every pixel is compared with the various signatures and assigned to the class whose signature (usually as a data set within the computer rather than a plot) comes closest. A few pixels in a scene do not match and remain unclassified, because these may belong to a class not recognized or defined.

Many of the classes for the Morro Bay scene are almost self-evident: ocean water, waves, beach, marsh, shadows. In practice, we could further sequester several such classes. For example, we might distinguish between ocean and bay waters, but their gross similarities in spectral properties would probably make separation difficult. Other classes that are likely variants of one another, such as, slopes that faced the morning sun as Landsat flew over versus slopes that face away, might be warranted. Some classes are broad-based, representing two or more related surface materials that might be separable at high resolution but are inexactly expressed in the TM image. In this category we can include trees, forests, and heavily vegetated areas (the golf course or cultivated farm fields). For high spatial resolution and hyperspectral data, the trees can often be subdivided to the species level.

For the first attempt at a Supervised Classification, the writer set up 17 discretional classes. These were picked largely from my on site experience (as indicated on page 1-1, I went to Morro Bay on vacation in California even before I attempted the classification). These sites are shown as color polygons traced on the true color (Bands 1,2,3) composite, as shown next. (Note that their site colors are assigned here for display convenience and do not correspond to their class equivalent colors in the maps shown on the next page).

The location(s) of training sites and color key for each class set up for the Morro Bay scene is superimposed on a true color image of that scene.

Note that IDRISI does not name them during the stage when the signatures are made. Instead, IDRISI numbers them and names are assigned later. Several classes gain their data from more than one training site. IDRISI has a module, SIGCOMP, that plots the signature of each class. Here we show Signature plots for 8 general classes (not all the same as the training sites shown above; this is done to simplify the first product). These are Ocean = Deep Blue; Waves = Green; Beach = Lighter Blue; Town = Maroon; Marsh = Purple; Vegetation = Olive; Hillslope = Light Gray; Shadow = Darker Gray. All TM bands (listed on the abscissa as Morobay1..., Morobay2, etc.) except 6 are involved. The ordinate DN values range from 0 to 255.

Spectral signatures for 8 general classes found in the Morro Bay scene; abscissa shows TM Bands 1 through 7; ordinate gives DN values from 0 to 255.

The signature positions at each band are somewhat artificial in that the raw data were not adjusted for relative gains; this affects Bands 1 and 5 in particular. However, it is clear from these plots that most of the signatures are different from one another over the 6 bands used, even though some are close-spaced (nearly coincident) over intervals consisting of several adjacent bands, e.g, all but Waves and Beach between TM1 and TM2. The greatest separability among all classes occurs in Band 5.

IDRISI also has a program that presents pixel information for each signature, recording the number of pixels contributing to the data, and the mean, maximum, minimum, and standard deviation of DN values for each signature. To help you get a deeper feel for the numerical inputs involved in these calculations, we have reproduced a simplified version of these data in the following table:

Table of Band Means and Sample Size for Each Class Training Set

BAND: 1 2 3 4 5 6 (TH) 7 No. of

Pixels

Class
1. Seawater 57.4 16.0 12.0 5.6 3.4 112.0 1.5 2433
2. Sediments1 62.2 19.6 13.5 5.6 3.5 112.2 1.6 681
3. Sediments2 69.8 25.3 18.8 6.3 3.5 112.2 1.5 405
4. Bay Sediment 59.6 20.2 16.9 6.0 3.4 111.9 1.6 598
5. Marsh 61.6 22.8 27.2 42.0 37.3 117.9 14.9 861
6. Waves Surf 189.5 88.0 100.9 56.3 22.3 111.9 6.4 1001
7. Sand 90.6 41.8 54.2 43.9 86.3 121.3 52.8 812
8. Urban1 77.9 32.3 39.3 37.5 53.9 123.5 29.6 747
9. Urban2 68.0 27.0 32.7 36.3 52.9 125.7 27.7 2256
10. Sun Slope 75.9 31.7 40.8 43.5 107.2 126.5 51.4 5476
11. Shade Slope 51.8 15.6 13.8 15.6 14.0 109.8 5.6 976
12. Scrublands 66.0 24.8 29.0 27.5 58.4 114.3 29.4 1085
13. Grass 67.9 27.6 32.0 49.9 89.2 117.4 39.3 590
14. Fields 59.9 22.7 22.6 54.5 46.6 115.8 18.3 259
15. Trees 55.8 19.6 20.2 35.7 42.0 108.8 16.6 2048
16. Cleared 73.7 30.5 39.2 37.1 88.4 127.9 45.2 309

1-22: Examine the signature plots and the table. What can you say about the plots in terms of similarities and differences? Based on numbers in the table, would you predict any notable differences in the signatures for towns; marsh; sunlit hillslopes and shadows? ANSWER

We can deduce from this table that most of the signatures have combinations of DN values that allow us to distinguish one from another, depending on the actual standard deviations (not shown). Two classes, Urban 1 and Cleared (Ground), are quite similar in the first four bands but apparently are different enough in Bands 5 and 7 to suppose that they are separable. The range of variations in the thermal Band 6 is much smaller than in other bands, suggesting its limitation as an efficient separator. However, as we will see next, its addition to the Maximum Likelihood Classification increases the spatial homogeneity of some classifications.

As another aside: In some older computer-based classification systems, one (sometimes the only) hard copy output was a printout that showed the different classes in alphanumeric symbols. Here is an example for a scene in central Pennsylvania:

Alphanumeric printout of a classification produced on Penn State University's early ORSER system; the area is near Clearfield, PA.

Since the mid-1980sm most final products are color images that can be printed as photographs.

There are scores of different classifiers.We will limit our examination of specific Classification Methods to just three, the first two being widely used in determining classes in space imagery. Among those omitted we cite: Parallelepiped; Piecewise Linear; Binary Decision Tree; Density Based Clustering; others.

Source: http://rst.gsfc.nasa.gov