Lab 9: Advanced Classifiers 2

Goal and Background

                This lab was a continuation on how to use advanced classifiers. The advanced classifiers for this lab were a decision tree and an artificial neural network. The lab helped with an introduction into these types of powerful advanced classifiers. A decision tree helps to define land classification with hypotheses, rules, and variables. An artificial neural network tries to mimic human brain activity when deciding land classifications. It runs numerous iterations before the end product can be classified.

Methods

                The first part of the lab was to use an expert system classification to improve upon an existing image. Expert systems use ancillary data to help improve images and have very high accuracy assessments. The first stage in expert system classification is to create a knowledge base. Within the knowledge base are rules that help classification. The image used in this part of the lab was eau_cpw_al2011cl.img. This is an image of the Eau Claire-Chippewa Falls-Altoona area and there are some wrongly classified pixels. In ERDAS Imagine the Knowledge Engineer is used this is found under the Raster tab. Knowledge Engineer has three main components which are hypotheses, rules, and variables. Hypotheses are planned LULC classes, rules are functions that help prior imagery and ancillary data, variables are the previous images and ancillary data. Hypotheses are created in the Knowledge Engineer. They are selected by choosing the hypotheses icon. For this lab the hypotheses were water, urban/built-up, forest, green vegetation, and agriculture. To create rules the corresponding icon is used. This creates a new rule. Each rule will have a name corresponding to the class. For water “WTR” was used. After rules new variables are to be created. For this the variable was ec_cpw_al2011. Variable type is raster. The image for this lab is au_cpw_al2011cl.img. The following classifications are done a similar way. The variable does not need to be changed in rules props. Each class is given a value of either 1, 2, 3, 4, or 5. To help wrongly classified pixels ancillary data is used. Creating separate classes for already existing classes is done as well. The first new class is other urban and it is colored as yellow. The new rule is OTR. The new variable is other_urban and is given a value of 1 in rule props. The second row in rule props is changed to ec_cpw_al2011.img and is given a value of 2. This creates an argument that pixels classified as urban in ec_cpw_al2011 and that are within the ancillary data should be reclassified as other urban. The urban rule is opened to help create a new rule for the urban_other class. Urban_other is in the second row. The rule is Urban_other !=1. This means that urban_other is not classified as urban. The same is done for green vegetation and agriculture. Arguments are written to help classify wrongly classified land and rules are made to show that the new agriculture and green vegetation are not classified as their parent classes. A Knowledge Classification is now ran. In the Knowledge Classification window all available classes are selected. Cell size should be 30x30. Each class should be changed so that there is only 6 classes; water, residential, forest, green vegetation, agriculture, and other urban.

The reclassified image using Knowledge Engineer

                The second part of this lab is a neural network classification. This part is carried out in a different remote sensing software called ENVI. In ENVI the new image is can_tmr.img. RGB colors are set and the available bands are set to 4, 3, 2. Regions of Interest training sets were to be used. In the main image menu Region of Interest is selected. CLASSES.ROI restore the regions of interest. This put 3 regions of interest into the image. These images appeared to be bare soil, forest, and agriculture. These 3 regions help classify the entire area. To use neural network classification the process is Classification, Supervised, Neural Net. Can_tmr.img is the input file, in the parameters the 3 regions are selected to classify, the logistic radio button is the activation method, and 1000 iterations are entered. The final step of this part was to create our own land classifications. Three ROIs were selected for the campus of the University of Northern Iowa. These three classes for me were buildings, grass, and sidewalks. The ROIs are created using polygons placed around pixels.

Results 

Neural Network Iterations

UNI neural network image, personal classification