Creating training samples

Resource Center

To create training samples, use the training sample drawing tools on the Image
Classification toolbar.
Steps:
The steps below show how to create training samples with the Image
Classification toolbar:
1. On the Image Classification toolbar, choose an appropriate image layer in
the Layer list.

2. Click one of the drawing tools on the Image Classification toolbar. Notice
that there are three drawing tools available, for drawing polygons, circles,
and rectangles.

3. In ArcMap, identify an area that belongs to a known class. Draw a training

sample to enclose it. The following image shows a polygon training sample
in ArcMap:

4. Once you finish drawing the training sample, a new class is created
in Training Sample Manager with a default name, value, and color.

Open Training Sample Manager and change the class name, value, and color
if desired.
5. Repeat steps 2 to 4 to create a few more training samples to represent the
rest of the classes on the image. The following is what Training Sample
Manager looks like after five classes are created:

Managing training samples

Resource Center

During the process of creating and evaluating training samples, you will need to use the
tools on the Training Sample Manager to maintain the list of classes. For example, you
may need to change the class name, value, and color. Also, classes may need to be
merged, deleted, reorganized, and so on. See the following on how to perform some
common training sample management tasks.

Managing training samples

The first three tools on the toolbar are used for managing the training samples.

Clearing all training samples

1. In the Training Sample Manager, click the Clear Training Samples button

Loading training samples

1. In the Training Sample Manager, click the Load button
dialog box appears.

. A file browser

2. In the file browser dialog box, select a training sample feature class and
click Add.
Note:
When you load training samples from a feature class, the loaded training
samples are appended to the existing list of training samples. If you do
not need the existing training samples, you can click the Clear Training
Samples button to remove them before you load the feature class.

Saving training samples

1. In the Training Sample Manager, click the Save button
dialog box appears.

. A file browser

2. In the file browser dialog box, navigate to the desired location and specify a
name for the feature class or shapefile, then click Save.

Managing classes
Other tools on the toolbar are used for managing the classes.

Merging classes
1. In the Training Sample Manager, select two or more classes that you want
to merge.
2. Click the Merge button

Note:
The Split button
does the opposite of the Merge button. It splits a class
into multiple classes so that each class has one training sample.

Deleting classes
1. In the manager, select the classes you want to delete.
2. Click the Delete button

Moving a class up or down

1. In the manager, select the class you want to move.
2. Click the appropriate Up button

or Down button

Renumber all classes in ascending order

1. In the manager, click the Reset class values button

Renumber an individual class

1. In the manager, click in the Value cell of a row. The cell becomes editable.
2. Type the new integer value and press ENTER.

Rename a class
1. In the manager, click in the Class Name cell of a row. The cell becomes
editable.
2. Type a new name and press ENTER.

Changing the display color

1. In the manager, click in the Color cell of a class. A color selector is shown.
2. Pick a color from the color selector. The color of the training sample is
updated in the ArcMap display area and any opened evaluation window.

Training sample evaluation tools

Resource Center

Training samples are created to represent classes in a supervised classification.

Different classes should be separated in the multidimensional attribute space. If any of
them are overlapping, you might consider merging them into one class. To check the
separability and distribution of your training samples, Training Sample
Manager provides three evaluation tools: a Histograms window, a Scatterplots window,
and a Statistics window. These tools are accessible as buttons on the manager.

The Histograms window

The Histograms window allows you to compare the distribution of multiple training
samples. If the training samples represent different classes, their histograms should
not overlap each other.

To compare the distributions of two or more training samples, select their classes
in Training Sample Manager and click the Histograms button
.
The color of the histogram matches the color of the class in Training Sample
Manager.

The Histograms window contains the same number of graphs as the number of
bands in the image layer.

This window displays four graphs in one screen. If you have more than four
bands in the image layer (which means there are more than four graphs), a
vertical scrollbar will be available.

This tool works for integer images only. If the image data is stored as floating
point, the Histograms button
will be unavailable on Training Sample Manager.

The Change Order button (located at the bottom of the window) allows you to
change the overlapping order of the histogram series. If a class is hidden in the
background in one or more histograms, you can bring it to the foreground by
clicking on this button.

The following image shows an example of the Histograms window:

The Histograms evaluation window

The Scatterplots window

The Scatterplots window is another way to compare multiple training samples. If the
training samples represent different classes, their scatterplots should not overlap.

To compare the distributions of two or more training samples, select the classes
represented by the training samples in Training Sample Manager and click
the Scatterplots button
.

The color of the scatterplot matches the color of the class in Training Sample
Manager.

There is one scatterplot for each unique band combination of the image layer. For
an n-band image, there are 1 + 2 + + (n 1) scatterplots. For example, there
are six scatterplots for a four-band image.

This window displays four graphs in one screen. If you have more than four
bands in the image layer (which means there are more than four scatterplot
graphs), a vertical scrollbar will be available.

The tool works for both integer and floating-point images.

The Change Order button (located at the bottom of the window) allows you to
change the overlapping order of the scatterplot series. If a class is hidden in the
background in one or more scatterplots, you can bring it to the foreground by
clicking on this button.

The following image shows an example of the Scatterplots evaluation window:

The Scatterplots evaluation window

The Statistics window

The Statistics window displays the statistics for the selected classes.

To view the statistics of one or more training samples, select the classes
in Training Sample Manager and click the Statistics button
.

The available statistics include minimum values, maximum values, means,

standard deviations, and a covariance matrix.

The text in the Statistics window can be copied to the Windows clipboard by
pressing CTRL+C.

The following image shows an example of the Statistics window:

The Statistics window

An overview of the Multivariate toolset

Resource Center

Multivariate statistical analysis allows the exploration of relationships among many

different types of attributes. There are two types of multivariate analysis available in
ArcGIS Spatial Analyst, Classification (both Supervised and Unsupervised) and Principal
Component Analysis (PCA).
The goal of classification is to assign each cell in a study area to a class or category.
With Supervised classification, you have a specific knowledge about the study area and
can identify representative areas, or samples, of each class. Unsupervised classification
uses naturally occurring statistical groupings in the data to determine the clusters into
which the data will be classified.
Learn about multivariate classificationLearn how to produce signature files and class
and cluster analysisLearn how to evaluate classes and clustersLearn how to perform the
classification
The general procedure for both Supervised and Unsupervised classification follows:
1. Identify the input bands.
2. Create the classes or clusters.
The following tools can be used: Create Signatures, Iso Cluster,
or Sample from the Extraction toolset.
3. Evaluate and edit the classes or clusters.
Use the Dendrogram or Edit Signatures tools.
4. Perform the classification.

Use the Maximum Likelihood Classification or Class Probability tools.

The Image Classification toolbar provides an integrated environment to help with the
multistep workflow that is necessary for performing a classification.
Learn more about image classification
To eliminate redundancy in the data and make it more interpretable, you can transform
your multivariate data through PCA.
Learn more about transforming multivariate data with Principal Component Analysis
(PCA)
The following table lists the available multivariate tools and provides a brief description
of each.
Tool

Description

Band Collection Statistics

Calculates the statistics for a set of raster bands.

Class Probability

Creates a multiband raster of probability bands, with one band being created for each class

Create Signatures

Creates an ASCII signature file of classes defined by input sample data and a set of raster b

Dendrogram

Constructs a tree diagram (dendrogram) showing attribute distances between sequentially m

Edit Signatures

Edits and updates a signature file by merging, renumbering, and deleting class signatures.

Iso Cluster

Uses an isodata clustering algorithm to determine the characteristics of the natural grouping
signature file.

Iso Cluster Unsupervised

Classification

Performs unsupervised classification on a series of input raster bands using the Iso Cluster

Maximum Likelihood Classification

Performs a maximum likelihood classification on a set of raster bands and creates a classifi

Principal Components

Performs Principal Component Analysis (PCA) on a set of raster bands and generates a sing

Tools of the Multivariate toolset