Lab 8: Expert System/Decision Tree and Artificial Neural Network Classifiers
Goals and Background:
The goal of this lab was to gain experience in correcting classification with the expert system/decision tree method leveraging ancillary data, and performing a neural network classification. The expert system/decision tree method for classification allows the user to create shapes that designate an area where the user would like to change specific classes to other classes in an existing classified image by using whatever ancillary data is available to the user to consult. This is performed in ERDAS Imagine. The neural network classifier simulates the structure of the human brain in performing its black box classification based on a users input parameters and data, and can be performed in ENVI. The
Methods:
Part 1: Classification using the expert system/decision tree method
This section was performed using the Knowledge Engineer tool in ERDAS Imagine.
In examining the provided classified image of the cities of Eau Claire and Chippewa Falls it was evident that some areas classified as residential were classified as urban, some areas that were classified as agriculture should have been classified as green vegetation and vice versa, and some other areas that were classified as urban/residential should be classified as other urban/other. Rules were assigned per ancillary data to each area that needed to be changed. For example, in Figure 1 a rule was written to assign the residential class to the areas where the ancillary data (a raster file where the value was 1 in areas of urban/other LULC) did not equal 1, and where the classified image was classified as urban/residential already.
Similar rules were written to denote the areas of all of the other classes desired, leveraging the existing classified image and the ancillary raster data. Of particular interest to this process was the creation of two green vegetation classes, and two agriculture classes, which were later recoded as one class each.
The full resulting logic is shown in Figure 2.
The classification was run, the recode mentioned above was performed, and a map figure was then created using ESRI ArcMap.
Part 2: Classification using the artificial neural network method in ENVI
After opening the image to be classified, the Available Bands List window appeared (Figure 3). A 4,3,2 false infrared image was opened by clicking on RGB Color, then selecting the bands in the 4,3,2 order, then clicking Load RGB. The region of interest tool was then opened, and ROIs that were provided were restored to the tool and displayed over the image. The supervised Neural Net classifier was then opened and the image to be classified was opened in the resulting window. In the next dialog window parameters were set (Figure 4). 1000 training iterations were used in addition to logistic activation, but all other parameters were kept as default.
Training rate and training momentum can be changed to adjust the classification. If the rate is made too high, the risk of oscillations and non-convergence will increase. The training momentum is a value from 0 to 1.0 and rates above zero will allow a greater training rate without oscillations. This influences the weight of change between iterations of training.
Results:
Part 1:
Below in Figure 5 are the results of the expert system/decision tree reclassification. The areas that were classified prior as residential are now classified as other urban. This was possible via ancillary rasterized zoning data. Other areas were changed as well using ancillary data as is indicated in the logic presented prior.
Part 2:
Figure 5 shows the results of the artificial neural network classification. The output image shows the three classes that resulted from the input regions: vegetated high lying regions, low lying areas of agriculture, and unidentified areas of high reflectance.
Sources:
A Quickbird high resolution image was sourced from the University of Northern Iowa Department of Geography. Landsat imagery is from the Earth Resources Observation and Science Center of the US Geological Survey. Instruction from Cyril Wilson.
The goal of this lab was to gain experience in correcting classification with the expert system/decision tree method leveraging ancillary data, and performing a neural network classification. The expert system/decision tree method for classification allows the user to create shapes that designate an area where the user would like to change specific classes to other classes in an existing classified image by using whatever ancillary data is available to the user to consult. This is performed in ERDAS Imagine. The neural network classifier simulates the structure of the human brain in performing its black box classification based on a users input parameters and data, and can be performed in ENVI. The
Methods:
Part 1: Classification using the expert system/decision tree method
This section was performed using the Knowledge Engineer tool in ERDAS Imagine.
In examining the provided classified image of the cities of Eau Claire and Chippewa Falls it was evident that some areas classified as residential were classified as urban, some areas that were classified as agriculture should have been classified as green vegetation and vice versa, and some other areas that were classified as urban/residential should be classified as other urban/other. Rules were assigned per ancillary data to each area that needed to be changed. For example, in Figure 1 a rule was written to assign the residential class to the areas where the ancillary data (a raster file where the value was 1 in areas of urban/other LULC) did not equal 1, and where the classified image was classified as urban/residential already.
 |
| Figure 1: A rule specifying areas of residential areas |
Similar rules were written to denote the areas of all of the other classes desired, leveraging the existing classified image and the ancillary raster data. Of particular interest to this process was the creation of two green vegetation classes, and two agriculture classes, which were later recoded as one class each.
The full resulting logic is shown in Figure 2.
 |
| Figure 2: All logic contributing to the expert system/decision tree reclassification |
Part 2: Classification using the artificial neural network method in ENVI
 |
| Figure 3: ENVI training sample interface |
After opening the image to be classified, the Available Bands List window appeared (Figure 3). A 4,3,2 false infrared image was opened by clicking on RGB Color, then selecting the bands in the 4,3,2 order, then clicking Load RGB. The region of interest tool was then opened, and ROIs that were provided were restored to the tool and displayed over the image. The supervised Neural Net classifier was then opened and the image to be classified was opened in the resulting window. In the next dialog window parameters were set (Figure 4). 1000 training iterations were used in addition to logistic activation, but all other parameters were kept as default.
 |
| Figure 4: Artificial neural network image classification parameters |
Training rate and training momentum can be changed to adjust the classification. If the rate is made too high, the risk of oscillations and non-convergence will increase. The training momentum is a value from 0 to 1.0 and rates above zero will allow a greater training rate without oscillations. This influences the weight of change between iterations of training.
Results:
Part 1:
Below in Figure 5 are the results of the expert system/decision tree reclassification. The areas that were classified prior as residential are now classified as other urban. This was possible via ancillary rasterized zoning data. Other areas were changed as well using ancillary data as is indicated in the logic presented prior.
 |
| Figure 5: Expert system/decision tree classifier results |
Figure 5 shows the results of the artificial neural network classification. The output image shows the three classes that resulted from the input regions: vegetated high lying regions, low lying areas of agriculture, and unidentified areas of high reflectance.
 |
| Figure 5: Neural network classifier results |
Sources:
A Quickbird high resolution image was sourced from the University of Northern Iowa Department of Geography. Landsat imagery is from the Earth Resources Observation and Science Center of the US Geological Survey. Instruction from Cyril Wilson.
Comments
Post a Comment