Lab 10: Object-Based Classification

Introduction:

This lab exercise was set up to teach the class how to properly perform object-based classification using eCognition, an object-based image processing tool.  Object-based classification integrates both spectral and spatial information to aid in extracting land use/land cover features from remotely sensed imagery.  Some of the objectives in this lab included:  segmenting an image into spatial and spectral clusters, selecting which of these clusters (objects) to use as training samples using a nearest neighbor classifier, and executing and refining the object-based classification output.


Methods:

As this was the first time the class had used eCognition, the first portion of the lab involved getting to know the software and then importing an image of the same study area of Eau Claire and Chippewa Counties that was classified in previous lab exercises.  A project was then created with the imagery being set using layer mixing to the 4, 3, 2 band combination that the class has become so used to.

At this point, it was time to create the different image objects (Figure 1).  This is a simple process that involved opening up the process tree and creating a new process.  Multiresolution segmentation was used along with a scale parameter of 10.  From there, the process was executed and the various objects were created.  It was possible to view the objects in several different ways including pixel view, object mean view, with or without the outline, and transparent or not.

The different objects that were created can be seen here as outlined in blue.  (Figure 1)

From here classes were created, these included:  agriculture, forest, green vegetation, urban, and water.  A nearest neighbor classifier using mean was then assigned to the classes.  Sample objects were selected  based on the classes knowledge of the spectral reflectance of the classes and the appearance of the land cover in a 4, 3, 2 band image (Figure 2).  After the samples were selected the classification was run and it was possible to review the results.

Here several green vegetation training samples and one urban training sample can be seen.  (Figure 2)

One interesting aspect of object-based classification is that it's possible to easily, manually edit objects that are known to be incorrectly classified.  After all manual editing was performed it was possible to export the result to a raster and create a map from the classified image (Figure 3).

This is the final map of the image that was classified using object-based classification.  (Figure 3)

Conclusion:

Image based classification is an interesting, newer way to perform image classification.  It is more useful in some application than other and seems to be especially customizable if the original result isn't up to a high enough quality.  eCognition is also an extremely useful and rather user-friendly software that could help make object-based classification more relevant in many future applications.