Geoscience Reference
In-Depth Information
Figure 10.5
Land-cover classification of a portion of scene 17/29. Three classes are shown: water (dark), non-
forest land (medium grey), and forest (white).
10.4.2
Aspects of Scene-Based Consistency Overlays
Once a final classification was obtained for a given scene, a “confidence” overlay was produced
wherein the confidence value of each pixel corresponded to the consistency level of its parent
cluster. Example results are shown in Figure 10.5 for the 1980s path/row scene 17/29. Figure 10.5
illustrates the three primary classes (i.e., water [dark], nonforested land [medium grey], and forested
[white]). Figure 10.6 shows an enlargement of the confidence layer of the central portion of Figure
10.5. The confidence range 0.0 to 10.0 was presented as a grey-level scale from black to white.
The following points are worthy of note: (1) Water was easily recognized, and hence the central
portions of most water bodies exhibited a high, uniform confidence level. (2) Pixels along interclass
boundaries, such as the edges of lakes or forest patches, tended to be of low confidence (Figure
10.6). They are members of clusters containing primarily “mixed” pixels and therefore have a low
accuracy. (3) Forested areas exhibited a slightly higher average confidence than nonforested areas.
This is related to the fact that this scene has more forest than nonforest cover. Consequently, the
population of pixels classed as forest will contain a relatively lower proportion of commission
errors, resulting in a corresponding higher level of interscene classification consistency.
10.4.3
Aspects of the Accumulated Confidence Layer
Figure 10.7 and Figure 10.8 illustrate a portion of a three-class LC product and accompanying
confidence overlay respectively. The interscene overlap regions are readily distinguishable in Figure
10.8 by their higher levels of accumulated confidence. In these regions significant confidence
variations still arise, either from conflicting classifications or information loss in one of the con-
stituent scenes because of cloud contamination (e.g., in central Michigan). Finally, in Figure 10.7
there are data gaps, appearing as nearly horizontal black lines, that arise because of along-track
data loss during the preprocessing steps of resolution reduction and haze removal (Guindon and
Zhang, 2002).
