Lab 4: Supervised Classification

Goals and Background:

This lab had students extract informational classes from Landsat 7 (ETM+) images. The skills practiced included collection of training samples for the classifier to use in classification and evaluation of the quality of the training samples collected. In collecting samples for each of the classes the user must pick enough spectral samples that appropriately show the variance within the class, but must make sure that the samples do not overlap into areas that would be otherwise classified to ensure the class has a unique mean spectral signature. A minimum of 5 to 10 samples must be created for each class and each sample should show a very similar spectral profile. Evaluation of samples in this lab included the use of histograms, an image alarm tool, and a spectral separability report. After these steps were completed, classification was executed using tools in ERDAS Imagine, resulting in a raster classification image that could be made into a neat figure in ArcMap. 

The classes used mimicked those used in unsupervised classification (Lab 3) and can be seen beside their color assignment below in Figure 1 below.

Figure 1: Classes

Methods:

Part 1: Collecting training samples for supervised classification

Part 1 involved collecting the initial spectral samples to be used in classification. 12 samples were collected for water, 11 were collected for forest, 9 were collected for agriculture, 11 were collected for urban areas, and 7 were collected for bare soil areas. In ERDAS Imagine, the image used to classify was brought in and the drawing tool was used to create polygons for sample areas of interest (AOIs). These samples were then input one by one into the Signature Editor tool (under the raster and then supervised menus), and named and organized by color. Forested land was identified by a red or dark red false infrared image, while agricultural land appeared light red to pink, urban land appeared light gray, cyan, or white, and bare soil appeared turquoise or slightly lighter. These training samples were then saved as their raw and original version before moving on to evaluation and editing. Figure 2 shows the collection of a signature, while Figure 3 shows the raw and original training samples.

Figure 2: Signature Collection

Figure 3: Training Samples

Part 2: Quality evaluation of training samples

In part 2 quality of training samples was assessed. This was first done through the use of spectral signature plots of each class of signatures. Outlier signatures were deleted and replaced so that there was not too much variation between signatures (but still enough so that the range of signatures would be represented). One such plot of a specific class is shown in Figure 4. Histogram plots of these faulty signatures were also made on all bands, revealing in some cases multimodal distributions signifying a bad signature.

Figure 4: Plotted Water Signature Samples

Next, the image alarm was used in order to find the areas that the training samples would identify as water on the image. If the image alarm failed to find all areas of a class, samples were identified in the areas no identified as the class, and were added.

Figure 5: Signature Mean Plot for All Samples

Finally, after displaying a signature mean plot and finding it was hard to discern the bands with the best separability (Figure 5), a separability report was utilized with 4 layers per combination and transformed divergence for the distance measurement (Figure 6). This report found that bands 1, 2, 4, and 6 had the best separability, and my best average separability to be 1971, an adequate score.

Figure 6: Separability Evaluation and Report

Next, the samples for each class were merged into single mean samples after the final sample choices were saved. These were labeled and colored accordingly, then the mean signatures were saved as a .sig file. 

Part 3: Performing supervised classification

After opening the Supervised Classification (under the raster tab and the supervised menu), the image was brought into the tool, the best mean signature file was brought in, and the settings seen below in Figure 7 were chosen. The tool was then run.

Figure 7: Performing Supervised Classification

Results:

The final result of the classification is seen below in Figure 8. The result was unsatisfactory in that even in initial qualitative assessment, it was a less accurate land use/land cover map than the unsupervised classification of Lab 3 produced. This was unexpected. Quantitative assessment of the two classification methods, comparison, and discussion of what could have caused the unsupervised classifier to perform better classification than the supervised classifier can be found in the Lab 5. Specific commentary on the performance of this classifier with each class are seen below.

Water: The classifier did a fairly good job at finding water bodies and rivers. Some sections of rivers were not classified as water however. They were classified as urban instead. This doesn’t make sense as there is little to no overlap between the signatures for the water and urban bands. The only similarity seen is the overall shape of the signature.

Forest: Forest was often classified as urban, bare soil, or agriculture. The large amount of mixing between these classes was not expected, though it is understandable due to the similarities between especially forest, agriculture, and bare soil in their mean signature plots. This could be due to error in selecting signatures. Signatures should perhaps have been collected so that informational classes had less similar signatures.

Agriculture: Large portions of agricultural areas were filled with bare soil classification. Areas were also mixed up with urban classifications. This again goes back to the same problem of these three classifications having very similar mean spectral signatures.

Urban/Built-up: The classifier was very good at picking up urban areas. This was due to the careful choice to only include small samples of road for signature samples. The classifier did too good of a job classifying urban areas, however, and overreached into agricultural, bare soil, and forested lands. The urban class covers almost all roadways in my image and also all of the cities. The area in between the urban classified areas in cities is often classified as bare soil, but this is expected because of the similarity between the two mean spectral signatures.

Bare Soil: Bare soil was well picked up in the images, but it sometimes stretched into areas of agriculture and of forest.

Figure 8: Final Result

Sources:

Lab instruction from Dr. Cyril Wilson. Landsat imagery is from the Earth Resources Observation and Science Center of the US Geological Survey.