Lab 3

Lab 3: Rule Based Classification

Jackson Radenz

Introduction:

Through Object-Based Image Analysis, the user is not only able to apply nearest-neighbor rule based classification features to the classification samples acquired, they are also able to implement rule based classifications that are manually defined within the class hierarchy.  This allows the user to identify features that stand out between classes in order to classify them much more accurately.  Because of the rule-based classification defined by manually assigned thresholds to each class, the user is able to classify areas to level 3, level 4, level 5 etc. with much more ease and accuracy.  In lab 3 we explore the abilities of rule-based classification through classification of imagery from North Wales, a subset of Landsat 7 imagery.     

Rule-Based Classification Thresholds

Within rule-based classification thresholds found in each class, there are two types of thresholds, absolute thresholds and fuzzy thresholds.  Absolute thresholds are the easiest to define.  An example of an absolute threshold would be:

Class: Improved Grassland
Mean NIR > 125
Mean NDVI >= 0.5

Figure 1. Absolute Thresholds

This implies that if the mean layer value of the NIR band in a segmented object is greater than 125 and if the mean NDVI is greater than or equal to 0.5 it is able to be classified as 'Improved Grassland' if all other parameters are not met.  This allows the user to refine the image into much more specific, accurate classes if spectral, textural, and geometrical differences between classes can be derived.

However, our due to our prior knowledge in Remote Sensing we understand that not all pixels are homogeneous.  In reality, a pixel and/or object classified as 'Forest' may be 70% forest but may encompass 15% bare soil and 15% grassland, as well.  This is especially a problem in imagery with  with very low spatial resolution.  Therefore, fuzzy classification is able to be utilized due to this problem.  So, in this case objects could be classified as

Forest = 0.7
Bare Soil = 0.15
Grassland = 0.15

Figure 2. Fuzzy Thresholds

This does not provide the user with a rigid classification, however it does provide a more accurate classification as it accounts for all land types that encompass the object.

Study Area & Imagery:

• Llyn Brenig in the Denbigh Moors, North Wales, United Kingdom
• Subsetted image collected from Landsat 7 
• Minimum X: 7800 Maximum X: 8500
• Minimum Y: 7400 Maximum Y: 8100
• PAN: 15m
• BLUE, GREEN, RED, NIR, SWIR1, SWIR2: 30m

Method:

• Open eCognition Developer → 'Create New Project' → Open Panchromatic imagery first, followed by the VIS & Infrared imagery
• Adjust project and subset parameters to the same as figure 3 & figure 4

Figure 3. Project Parameters      

Figure 4. Subsetted Image

• Set up a Customized 'Mean NDVI' feature → 'feature view' → 'create new arithmetic feature'→ enter figure 5 details

Figure 5. Arithmetic Feature Window

• Next create classes under 'Class Hierarchy' window.  Class hierarchy window should match figure 6

Figure 6. Class Hierarchy window

• Create rule-based classification thresholds for each class by double clicking each class → contained → insert expression → insert threshold
• Create  rule-based classifications for each class as follows:
• Acid Semi Improved Grassland: N/A
• Bog/Heath: Mean GREEN > 30 & Mean GREEN < 42
• Forest: Mean NIR < 100 & Mean SWIR1 < 40 & Mean NDVI > 0.3 & Mean NDVI < 0.6
• Improved Grassland: Mean NIR > 100 & Mean NDVI >= 0.5
• Not Vegetation: Mean NDVI <= 0.275
• Water: Mean NDVI <= 0.05
• Create 'Rulebased Classification Example' in Process Tree → 'insert child' Multi resolution segmentation → refer to figure 7

Figure 7. Multi resolution segmentation parameters

Figure 8. Process Tree

• Create a Process Tree that matches Figure 8 
• Export image into Arc Map
• 'Define Projection' and add cartographic elements
• After creating map, adjust segmentation parameters and rule-based classification thresholds in class hierarchy in an attempt to more accurately classify imagery
•  Take screenshots, upload, and discuss

Discussion:

Figure 9. Classified Land Cover map of North Wales, UK

Classified Map:

Visual comparison of the RGB Landsat 7 imagery and classified map:

The water is the southwest corner of the map in inaccurately classified.  Areas of  the tip were classified as 'Not Vegetation', which we can conclude may be turbid water.  However, the classified map accurately mapped forest.  The patches of forest were homogeneous throughout, which made it easy for the algorithm to segment and map it.  Bog/Heath and Acid Semi Improved grassland are often mixed with each other in very geometrically irregular ways.  Therefore, I believe more rule-based classification thresholds must be produced to more accurately distinguish between the two.         

Figure 10.  Scale: 5

Figure 11. Scale 30

Figure 12. SWIR1 & NIR Weighted Layer: 3
Rest of layers: 1

Scale Factor:

The scale factor of 30 was too large for this imagery.  Just through visual comparison of the Landsat image and the classified image of 30 there are large inaccuracies.  However, the scale of 30 was effective in smoothing out the map where there were large areas of homogeneous land cover, especially west of the large water body. 

The scale factor of 5 presented a similar issue that pixel based image analysis encounters as well, salt and pepper effect.  There are small areas of heterogeneously classified land within large areas of homogeneously classified land.  There are small areas of Acid Semi Improved Grassland within large areas of Improved Grassland and Bog/Heath.  This most likely calls for manual correction.  However, the small scale factor did accurately classify small changes within the landscape.

Finally, weighting SWIR1 and NIR layers with 3 populated the map with much more Acid Semi Improved Grassland classified land.  Acid Semi Improved Grassland most likely has higher SWIR1 values and lower NIR values because more soil cover and less vegetation cover would be present.  Therefore, this allowed the algorithm to distinguish between the two types of Improved Grasslands more easily. 

Figure 13. Forest: 0.4 < Mean NDVI < .55

Water:  Mean NDVI <= 0.09

Bog/Heath: 70 <= Mean NDVI <= 100

Figure 14. Thresholds Listed Above + Not Vegetation: Mean NDVI <= .35

Forest: Mean SWIR1 <= 41

Figure 15. Thresholds Listed Above + Improved Grassland: Mean NIR > 125

Change of Existing Rule-Based Classification Thresholds:

Changing the existing rule-based classification thresholds under the class hierarchy gave the ability to more accurately classify the existing imagery.  After segmentation, I examined various feature values of different land cover types under the 'Object Feature View' window.  This allowed me to gain a better understanding of what type of object feature values better differentiated different types of land cover classes from another.  I found that mean NDVI was a great feature value to manipulate in classifying the image.  NDVI was able to not only accurately classify between water and vegetation, but it was also key in classifying between different types of vegetation.

Also, I found creating rule-based classification thresholds on mean NIR and SWIR1 values to be sufficient in classifying the data set because of both band's significance in differentiating between vegetation.

Figure 16. Thresholds Listed Above + Not Vegetation: GLCM Homogeneity <= 0.5

Water: 21 < Mean Brightness Temp. < 28

Figure 17. Thresholds Listed Above + Improved Grassland: Mean GLCM => 65

Figure 18. Thresholds Listed Above + Acid Semi Improved Grassland: Area => 300px 

Added Features:

The two feature values that were most efficient in distinguishing between different types of classes were mean brightness temperature and texture feature values of various objects.  I found that water had a higher brightness temperature than land.  Therefore, the brightness temperature was able to very accurately classify between land and water.  I found it interesting when I changed the mean GLCM of Improved Grassland to => 65 almost all of the Improved Grassland in the prior classification turned into Acid-Semi Improved Grassland.  That makes me conclude that mean GLCM may have been an inaccurate feature to use to classify the image because of such drastic change from its prior classification.  In the final classification I used area as a feature value.  Upon classification, it left large amounts of area in the image unclassified.  This makes me conclude that it is possible to develop too many rule-based classification thresholds in an algorithm that prohibits the program's ability to classify the image. 

Conclusion:

In conclusion, developing rule-based classification thresholds within the class hierarchy gave the user the ability to more efficiently and accurately classify the dataset.  This method was much more efficient than selecting samples before each classification, and it allowed the user to further refine their classification thresholds until an accurate classification is reached.  In my opinion, I achieved the most accurate classification in figure 13.  The ability to create 'mean NDVI' and other customized object feature value rule-based classification methods bring in the advantages of ENVI in a much easier user-friendly and efficient way.