Image classification versus manual mapping

Classification is a common tool for deriving thematic information, such as maps,  from satellite imagery. A range of algorithms have been developed over the years; these are typically broken down into supervised and unsupervised methods. Supervised classification requires the user to define training areas from which class statistics can be derived. These statistics are then used to identify pixels with given statistical properties which are then assigned to the given class. Unsupervised methods identify common statistical properties for a given number of classes and assign pixels to those classes based on the class statistics. It is common to over classify when using unsupervised classification and then combine classes (this allows better refinement of the result).

Classification methods can be very powerful and useful in reducing the complexity of an image to more manageable levels.  However, there are limitations. Cloud, haze and shadow can all complicate accurate classification. Unsupervised approaches require careful analysis and may result in classes that do not represent targets of interest. Targets need to be spectrally separable for the approaches to work: texture and shape are rarely considered.

Some approaches are more sophisticated. Object-oriented methods such as Definiens eCognition allow shape and context based decision making in classification. Spectral angle mapping and spectral unmixing use spectral libraries to identify known targets. Nevertheless, the answer to mapping may sometimes be old fashioned manual mapping.

Manual mapping approaches are sometimes dismissed as qualitative and irreproducible. Sannel and Brown (2010) amongst others have shown that mapping decisions are surprisingly similar: a +/-2 pixel deviation was common. Other experiments such a Raup et al. (2007) and experiments at Stockholm University, have shown that manual mapping can be both accurate and reproducible, even in the presence of shadow and other complications. Hence for some applications, particularly when a small number of classes or objects need separating, manual mapping may provide an alternative to classification. At the end of the day the human eye, coupled to the brain, is a very powerful tool for object identification, and the brain even sub-consciously, is incredibly good and separating objects of interest from complex backgrounds.

 

Sannel, A.B.K. and Brown, I.A., 2010: High resolution remote sensing identification of thermokarst lake dynamics in a subarctic peat plateau complex. Canadian Journal of Remote Sensing 36, S26-S40.

Raup, Bruce; Andreas Kääb; Jeffrey S. Kargel; Michael P. Bishop; Gordon Hamilton; Ella Lee; Frank Paul; Frank Rau; Deborah Soltesz; Siri Jodha Singh Khalsa; Matthew Beedle; Christopher Helm (2007). “Remote Sensing and GIS Technology in the Global Land Ice Measurements from Space (GLIMS) Project”. Computers and Geosciences 33:104–125.

 

Advertisements
This entry was posted in Articles, GIS, Remote Sensing, Uncategorized and tagged , , . Bookmark the permalink.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s