Download Supervised Classification - Laboratory | WILD 6750 and more Lab Reports Forestry in PDF only on Docsity!
Supervised Classification
Part 1
Unsupervised Classification
--Supervised Classification-- WILD 6750
Image Æ Clustering & Cluster Analysis Æ Cluster Grouping & Labeling Æ Accuracy Assessment
Supervised Classification
Image Æ Supervised Training Æ Pixel Labeling Æ Accuracy Assessment Development of a classification scheme by selecting representative areas using reference sources such as higher resolution imagery (DOQ, NAIP), or field notes/ data. The software then characterizes the statistical patterns of the representative areas and classifies the image.
The software is used to explore the image and compute clusters that represent groups of pixels with similar spectral properties. Those clusters are evaluated to determine what number of clusters (or classes) is desirable, and then the pixels are assigned to the best-fit cluster.
Classification: Unsupervised vs. Supervised
WILD 6750 Defining the Classes
Using the unsupervised classification method, the software was instructed to create x number of clusters in x iterations while attempting to meet a predetermined threshold value. By calculating and plotting the cluster statistics we were able to determine what number of clusters (or classes) to use.
SLC Area Sum of Variance and Standard Deviation Plot
0
200
400
600
800
1000
1200
1400
0 5 10 15 20 25 30 35 40 Clusters (Classes)
Avg. Sum of Var.
0
10
20
30
40
50
60
70
80
Avg. Sum of SD
Sum of Variance Standard Deviation
--Supervised Classification-- WILD 6750 Defining the Classes
When completing a supervised classification it is critical that a accurate and complete set of class definitions are created. These classes are typically created with a specific goal or target in mind.
The more well defined the classes the less the ambiguity and inconsistency within the classification. However, not all data can be fit into a “class” because of fuzzy or mixed areas within the image. There is often no clear boundary between certain land cover types, which merge into each other with changing bio-physical conditions.
Much like unsupervised classification, the point of diminishing returns will eventually be met and thus adding additional classes becomes unproductive.
WILD 6750 Defining the Classes
For this lab we will begin by using the region grow AOI (seed) utility in the AOI tools to select areas within the image that constitute the land cover categories which we want to classify. It may be necessary to adjust the region grow AOI parameters to increase or decrease the search criteria.
Image displayed, note the inquire cursor.
AOI tools active: AOI Æ Tools Region Grow AOI tool
Region Grow Properties
--Supervised Classification-- WILD 6750 Defining the Classes
Our goal is to identify and develop 5 different land cover types (or classes) with 4 training sites per cover type defining their characteristics.
Note that by editing the Region Growing Properties the AOI signature that is created will change.
Remember that the larger the Spectral Euclidian Distance the more variance, or the more inclusive, the signature. Adjust with caution! You may also want to experiment with the Neighborhood function and the Geographic Constraints to see the different types of growth as a result of various combinations.
WILD 6750 Defining the Classes
--Supervised Classification-- WILD 6750 Defining the Classes
Once a signature has been defined we must begin building our signature set by adding that signature to the Signature Editor.
Open the Signature Editor. Classifier Menu.
Note the toolbar, to add a signature to the Editor, click the Create New Signature from AOI tool.
The signature is added to the editor, note that the signature represents the “average” of those pixels selected in the seeded AOI (now known as a training site).
WILD 6750 Evaluating Signatures: Urban vs. Ag
The classes displayed here represent Agricultural areas and Urban areas. The mean plot would suggest that there is a clear separation of the classes (good!). We can also check this relationship by using the Evaluate Æ Separability utility:
When using the mean signature plot, use the Scale Chart to Fit Current signatures button to bring all plots into the graph.
--Supervised Classification-- WILD 6750 Evaluating Signatures: Urban vs. Ag
Evaluate the CellArray and determine if there is sufficient separability between training sites. If areas said to be different types appear close to each other, consider removing that site and selecting a new location.
Note these values are low, thus more closely related to each other. Urban vs. Urban.
Note these values are distant, thus less related to each other. Urban vs. Agriculture.
WILD 6750 Evaluating Signatures
--Supervised Classification-- WILD 6750 Combining Signatures
After the signatures have been evaluated and are determined to be satisfactory, the signature set SHOULD BE SAVED AT THIS POINT. Additionally, it is critical that you SAVE THE AOI TRAINING SITES for future processes.
Because we want our signatures of each cover type to be inclusive of each training site the signatures must be combined using the Merge Selected Signatures utility in the tool bar.
In this case the Ag classes were all selected and, using the Merge Selected Signatures button, a new weighted signature was created. Renamed here to Agriculture.
This “new” signature will be used in our supervised classification.
This is done for each class.
WILD 6750 Assignment – Using an image of your choice
Using the region grow utility in the AOI tools, select areas within the image that constitute the land cover category which you want to map. You may need to adjust the grow (seed) parameters to increase or decrease the search area. You should have, at least, 5 different land cover types with a minimum of 4 training sites per cover type. Once you have identified your training sites, develop the statistics and evaluate your selections. Use the evaluate and mean plot options signature editor to evaluate the spectral distance of your training sites. Add or remove training sites as necessary to create the most robust signature set possible.
Combine your training sites to create the weighted training site for each land cover type.
--Supervised Classification-- WILD 6750 Assignment
- List the land cover types you will attempt to classify. Briefly describe the characteristics of each cover type. Be careful to select land cover not land use.
- Create individual mean signature plots for at least five cover types (as seen on slide 12). Explain the signatures and what the mean values are representing.
- Provide a snapshot of your signature editor file.
- Using the Evaluate Æ Separability tool, compare your signatures. Select three cover types and compare (Slide 13 -14). Provide snapshots of these comparisons and explain what you see.
- Collapse your signatures and provide a Signature Mean Plot of your final classes. Again, using the separability tool and evaluate your final classes. Provide a snapshot and explanation of the Cell Array.
Make sure you have saved both signature files and your AOI training sites for next lab.
