Image classification is based on the assumption that variation that can be detected in imagery is indicative of real variation on the ground. Furthermore; it is assumed that when there is true variation on the ground, there is a corresponding change in the spectral information contained within the imagery that denotes this change. These statements might seem so obvious that you might believe they are completely trivial. Yet, by now, you should be able to think of many situations where one or both of these statements might not be true. There may be variation in the imagery caused by the sensor (an extreme case would be the SLC-off gaps in Landsat 7 imagery), by the atmosphere (clouds, for example), or by lighting conditions which cause shadows across an otherwise uniform surface. There may be important variations on the ground surface that are not easily detected in the imagery due simply to insufficient spatial resolution. Therefore, the analyst should always revisit these two simple assumptions before beginning a classification project, making sure that the imagery selected is appropriate for the desired task and that all possible preprocessing steps have been taken to reduce variation in the image caused by sources other than the ground surface.
According to Green and Congalton (2012), the success of a classification project depends on:
- defining a classification scheme that corresponds to real variation on the ground;
- understanding variation in the imagery and ancillary data used to support classification;
- linking variation in the imagery to real variation on the ground;
- exploiting linkages between variation in the imagery and variation on the ground to create a useful land cover map.
Green and Congalton (2012) go on to state that an effective classification scheme should:
- meet the user's needs;
- consist of classes that are
- mutually exclusive (an object/target should belong to one and only one class),
- totally exhaustive (all objects/targets in the study area can be assigned to a class),
- hierarchical (high level general classes further subdivided into more specific subclasses);
- include a minimum mapping unit which defines the smallest size area to be classified (which should never be as small as a single image pixel)
Finally, a classification scheme must include both:
- labels (class names)
- rules (definitions)
As you have been proceeding through the lab exercises in this course, you have probably realized that simple labels, such as "forest" or "urban," are highly subjective. In order to be applied consistently, labels must be associated with a set of specific criteria, such as "at least 30% of the ground as seen from above is covered by tree canopy," or "at least 50% of the ground as seen from above is built infrastructure." Only with a clear and unambiguous set of rules can a classification scheme retain its meaning from one scene to the next or from the individual analyst to the end user.