ArcGIS Tutorial: Spatial Analyst expansion module for image classification

In the ArcGIS Spatial analyst Extension Module, the multivariate toolset provides tools for monitoring classification and unsupervised classification. The Image Classification toolbar provides a user-friendly environment for creating training samples and feature files that are used in supervised classifications. The maximum likelihood classification tool is the main classification method. The signature file that identifies the category and its statistics is a required input for this tool. For supervised classification, use the training sample to create the feature file through the Image Classification toolbar. For unsupervised classification, the feature file is created by running the clustering tool. Spatial Analyst also provides tools for post-categorization processing, such as filtering and boundary cleanup. The detailed steps for the image classification workflow are described in the table.

　　

1. Data exploration and preprocessing

　　Data browsing

Classification analysis is based on the assumption that band data and training sample data conform to the normal distribution. To check the distribution of data in a band, use the interactive Histogram tool on the Spatial Analyst toolbar. To check the distribution of each training sample, you can use the Histogram tool on the training sample manager.

　　Stretching of band data

The classification process is sensitive to the range of values in each band. To make the number of properties of each band approximately the same, the range of values for each band should be similar. If the value range of a band is too small (or too large) relative to other bands, you can use the math tools in the Spatial Analyst Toolbox to extrude it. For example, you can use the multiply arithmetic tool to multiply a band by a constant to stretch its range of values.

　　Principal component Analysis

Principal component analysis transforms a multiband image to remove the correlation between bands. The information in the output image is mainly concentrated in the first few bands. By enhancing the first few bands, the imagery can see more details in the image when it is displayed in ArcMap. This may help to collect training samples. Principal component analysis is performed through the PCA tool on the multivariate analysis toolset.

　　Creating multi-band imagery

The Image Classification toolbar is available for multiband imagery layers. To load a single band into a new multiband image, you can use the band compositing tool.

　　Create a subset of bands for classification

To use all the bands in an image dataset in a taxonomy, you can add the image dataset to ArcMap and select the imagery layer in the Image Classification toolbar.

To use only some of the bands in an existing dataset in a taxonomy, you can use the Create Raster Layer tool to create a new raster layer for it. The new raster layer will contain only the specified subset of bands and can be used in the Image Classification toolbar.

2. Collect Training Samples

In supervised classification, training samples are used to identify classes and to calculate their characteristics. You can interactively create training samples using the Training sample drawing tool on the Image Classification toolbar. Creating a training sample is similar to drawing a graphic in ArcMap, except that the training sample shape is managed by the training sample manager instead of the ArcMap graphics layer.

To create a training sample, select a Training sample drawing tool (for example, the Polygon tool) on the Image Classification toolbar, and then draw on the input imagery layer. The number of pixels in each training sample should not be too small or too large. If the training sample is too small, you may not be able to provide enough information to fully create the class characteristics. If the training sample is too large, it may include pixels that are not part of the class. If the number of bands in the image is N, the best pixels for each training sample should be between 10n and 100n.

　　

3. Evaluate training samples

When you draw a training sample in the display interface, a new class is automatically created in the Training sample manager. The manager provides three tools to evaluate the training sample-histogram tool, scatter plot tool, and statistical data tool. You can use these tools to explore the spectral characteristics of different regions. You can also use these tools to evaluate training samples to see if there are enough gaps between classes.

　　

4. Editing classes

Depending on the evaluation of the training samples, it may be necessary to merge classes that overlap each other into a single class. You can do this by using the Merge tool in the Manager window. In addition, you can rename or renumber a class, change the display color, separate classes, delete classes, save and load training samples, and so on. Shows how to merge two classes:

　　

5. Create a signature file

Once you have determined that the training sample can represent the required classes and can be differentiated from one another, you can create a feature file using the Create feature file tool in the Manager window.

　　

6. Clustering (unsupervised classification)

In supervised classification, feature files are created from known and defined classes (for example, land-use types) that are identified by pixels enclosed in polygons. In unsupervised classification, clusters (not classes) are created by the statistical attributes of the pixels. Groups of pixels similar to statistical attributes in a multivariate space are grouped to form clusters. Unlike classes in supervised classifications, clusters have no category meaning (for example, land-use types).

For unsupervised classification using the Image Classification toolbar, you can create a signature file by running the ISO cluster unsupervised classification tool. You can also use the ISO Clustering tool set in the multivariate analysis toolset.

The Iso clustering tool only creates feature files that can then be used in the taxonomy (step 9 in the above workflow diagram). Created a new tool ISO cluster unsupervised classification (accessible from the Image Classification toolbar or the multivariate toolset) allows you to create feature files and output categorical imagery (steps 6 and 9) with just one tool.

7. Checking the signature file

You can use the tree Chart tool to examine attribute distances between successive merged classes in a feature file. The output is an ASCII file, and the tree shows the class interval. With a tree view, you can determine whether two or more classes or clusters are distinguishable, and if not, you might decide to merge them in the next step.

Access the tree Chart tool from Spatial analyst's multivariate toolset.

8. Editing the signature file

You should not edit the signature file directly in the text editor. Instead, use the Edit features tool in the multi-analysis toolset. This tool allows you to combine, renumber, and delete class features.

9. Application classification

To classify images, use the maximum likelihood classification tool. This tool is based on the maximum likelihood probability theory. It assigns pixels to a different class based on the average and variance of the class characteristics (stored in the signature file). The tool can also be accessed through the Image Classification toolbar.

You can also categorize images by using the interactive supervised classification tool. This tool can speed up the maximum likelihood classification process. Use this tool to quickly preview classification results without running the maximum likelihood classification tool.

10. Post-categorization processing

Classified imagery created with the maximum likelihood classification tool may incorrectly classify some cells (random noise) and form small, invalid areas. To improve classification, it is best to reclassify the cells of these error classifications and classify them into classes or clusters that can surround them directly. The most common techniques for cleaning up classified imagery include filtering, smoothing class boundaries, and removing small isolated areas. The map looks better when you apply the data cleansing tool.

　　Filter the sorted output

This process removes a single orphaned pixel from the classified image. You can do this with the following tools: The main filter tool, or the focus statistics tool with the majority as the statistic type. The difference between the two tools is that the main filter tool assumes a square field of 3 x 3 during processing, while the focus statistics tool supports more neighborhood types (for example, rings or circles).

　　Smoothing class Boundaries

The Boundary Cleanup Tool aggregates classes and smooths out the irregular edges of classes. This tool works by extending the class first, and then shrinking the class. It will increase the spatial consistency of the classified imagery. Adjacent areas may be connected together.

Generalize output by removing small isolated areas

After filtering and smoothing the process, the classified image should be much cleaner than before. However, there may still be some small isolated areas on the classified image. The generalization process further cleans up the imagery by removing such small areas from the image. This is a multi-step process involving multiple Spatial Analyst tools.

Run the region Merge tool on the classified image to assign unique values to each area on the image.

Opens the property sheet for the new raster layer created by the region Merge tool. Use the pixel count to identify thresholds for small areas to be removed.

Create a mask raster for the area you want to remove. This can be done by running the SET NULL function tool, which sets the area with fewer pixels to be a null value.

Run the Nibble tool on the classified image. Use the mask raster created in the previous step by using the Set as NULL function tool as the input raster mask. This clears the small area on the output image.

ArcGIS Tutorial: Spatial Analyst expansion module for image classification