GISCI 241 Chap.8 Supervised and Unsupervised Classification
Supervised and Unsupervised Classification
Image classification assigns thematic meaning to spectral measurements. Supervised classification begins with analyst-defined classes and representative training samples. Unsupervised classification begins with spectral clusters and adds semantic labels after grouping.
This chapter compares those workflows, shows why training distribution and seasonal transfer matter, and uses a confusion matrix to distinguish overall agreement from class-specific omission and commission. The guiding principle is that labels are hypotheses that independent reference data must test.
What this chapter covers
- 01
Thematic classes from multispectral pixel vectors
- 02
Representative training areas and spectral signatures
- 03
Supervised assignment using labelled examples
- 04
Unsupervised clustering and post hoc interpretation
- 05
Class count, mixed pixels and seasonal transfer
- 06
Confusion matrices, omission and commission
Worked example · free
Audit a two-class map
- +1Total reference observations equal 42 plus 8 plus 6 plus 44, giving 100.
- +1Correct observations equal 42 plus 44, giving overall accuracy of 86 percent.
- +1Reference forest accuracy equals 42 divided by 50, giving 84 percent.
- +1Mapped forest reliability equals 42 divided by 48, giving 87.5 percent.
- +1Forest omission count of 8 exceeds forest commission count of 6.
- +1Investigate missed forest for shadow, mixed edges or unrepresentative training.
Key terms
- Supervised classification
- A workflow that learns class characteristics from analyst-labelled training samples before assigning other pixels.
- Unsupervised classification
- A workflow that clusters pixels by spectral similarity before an analyst interprets and labels the groups.
- Training area
- A spatial sample selected to represent the spectral variation of a known thematic class.
- Spectral cluster
- A group of pixels with similar multiband values that may or may not equal one semantic class.
- Confusion matrix
- A table comparing mapped classes with independent reference classes to show agreement and error direction.
- Omission error
- Reference members of a class that the map fails to include in that class.
- Commission error
- Mapped members of a class that reference evidence places in another class.
Supervised and Unsupervised Classification FAQ
When is supervised classification preferable?
Use it when meaningful target classes are known and representative training areas can be collected across their variability. It gives control over the legend but can fail when training omits locations, seasons or mixed conditions. Keep independent reference data out of training so transfer can be tested honestly.
How many unsupervised clusters should I choose?
There is no universal number. Too few clusters merge distinct covers, while too many fragment one cover into shadows and material variants. Explore a defensible range, inspect spectral and spatial coherence, and merge clusters only with reference support. Explain how the final count serves the intended legend.
Why is overall accuracy not enough?
A common class can dominate the total and hide poor performance on the class that matters. Inspect class-specific omission and commission, reference sampling and spatial error pattern. If a rare wetland or hazard class drives the question, its recall and mapped reliability may matter more than aggregate agreement.
Why can high overall accuracy still be misleading?
A common class can dominate the diagonal and produce a strong total while a rare class important to the research question is frequently missed. Inspect reference-based and mapped-class accuracy for every class, map the spatial pattern of mistakes and confirm that validation samples are independent of training. Agreement is always conditional on the reference and sampling design.
Exam move
Draw parallel supervised and unsupervised workflows, marking exactly when semantic labels enter. For supervised practice, design training areas that cover class variation across space and season while avoiding mixed boundaries. For unsupervised practice, run a mental cluster-count experiment and explain why very low and very high counts fail.
Work small confusion matrices by hand and always name the numerator, denominator and error direction. Review lab-style outputs by asking whether colours match the legend, captions explain classes and comparisons discuss strengths and weaknesses. End by stating what independent reference evidence would validate the map and whether the trained model can transfer to another date.
Write operational definitions for each class before selecting samples. A class definition should identify the surface meaning, season and spatial scale rather than relying on a display colour. Map training locations and ask whether they cover illumination, moisture, management and geographic variation. Reserve independent reference observations so validation does not recycle the training evidence.
For unsupervised work, compare several cluster counts and record which materials are merged or fragmented. Attach semantic labels only after inspecting signatures and spatial context. Practise confusion matrices with an imbalanced common class and a rare class that matters to the research question.
Calculate overall agreement, reference-based accuracy and mapped-class reliability, then describe omission and commission in words. Map the errors to find shadows, mixed edges or seasonal transfer failures. Finally, test transfer by changing date or region. Explain why representative training in one scene may not cover another scene's spectral distribution, and propose new training, robust features or separate classification.
Conclude with the measured agreement and the class-specific limitation, not a claim that one percentage proves the map correct.