2013

Digital Image Processing Assignment # 1 Land Classification of Orillia, Ontario, Canada

Bennett White White Winter Environmental Consulting 1/29/2013

32 Robertson Rd., Niagara-on-the-Lake, L0S 1J0 905.932.1998

Index: Index: .......................................................................................................................................... 1 Purpose:........................................................................................................................................... 2 Introduction: .................................................................................................................................... 2 Background: .................................................................................................................................... 2 Study Area: ..................................................................................................................................... 3 Unsupervised Classification............................................................................................................ 4 Supervised Classification: ............................................................................................................... 9 Supervised Classification: ............................................................................................................. 10 Maximum likelihood................................................................................................................. 10 Mahalanobis distance ................................................................................................................ 11 Minimum Distance (Minimum Distance to Means) ................................................................. 12 Classification Comparison: ........................................................................................................... 12 Supervised ................................................................................................................................. 12 Advantage ............................................................................................................................. 12 Disadvantage ......................................................................................................................... 12 Unsupervised............................................................................................................................. 13 Advantage ............................................................................................................................. 13 Disadvantage ......................................................................................................................... 13 Differences between the two ..................................................................................................... 13 Work Cited:................................................................................................................................... 13

32 Robertson Rd., Niagara-on-the-Lake, L0S 1J0 905.932.1998

Purpose:
This report requires users/students to create a subset image and then analyze it through the use of ERDAS Imagine, conducting supervised and unsupervised classification in the process. The user will be required to perform sub-procedures during the classification process, supervised signature evaluation and recoded unsupervised classes taking in mind the parametric rules. The procedures that will be undertaken will be analyzed using ERDAS Imagine 2010 software and Landsat raster images. The legal documentation for the procedures of this assignment can be obtained from the Environmental Science department at Niagara College Canada, or with granted permission can be accessed from the Niagara College Canada server [x-drive/GIS Resources/GIS Second Semester/GISC9216].

Introduction:
It is the users duty to select a subset region of southern Ontario and analyses this area using classification procedures, unsupervised and supervised, to distinguish features present in the subset region. This assignment will involve the creation of training areas and the production of a supervised classification. Multispectral classification is the process of sorting pixels into a finite number of individual classes, or categories of data based on their data file values. In the context of remote sensing, this is the collection and analysis of reflected, emitted, or backed-scattered energy from an object or an area of interest in a multiple band electromagnetic spectrum of a region. Depending on the type of information the user will use, this data can be arranged into known features within the area of focus. It is the use of multispectral classifications that these pixels from satellite imagery can be given feature classification enabling them to be analyzed and recognized with use of remote sensing techniques. The global society is finding it more appropriate to use multi-band images of the Earths surface to access information pertaining to land cover and land use around the world. Landsat Thematic Mapper raster mages have become widely used in the field of remote sensing, which have made this industry a fast growing and lucrative sector for companies and countries to invest in. With respect to Landsat imagery, band 1-5 & 7, cover wavelength ranges between .45-1.75, 2.08 2.35nm with 30 meter resolution. These bands comprise of RGB, near and far infrared and panchromatic bands. Band 8 should also not be used in this case as it is a higher spatial resolution *** Band 6 is not used for this study, as its resolution is not needed and far too costly***

Background:
In the late 1950s United States and the USSR were in a tight race to be the first nation into space and experience what no other human had ever witnessed or experienced. Since the time of the cold war and the race to space, there has been a significant technological advance into aerospace engineering. There are hundreds, possibly thousands of satellites that have been sent into space for one purpose and that is to photograph the earth surface. 2

32 Robertson Rd., Niagara-on-the-Lake, L0S 1J0 905.932.1998

In modern time there is a more data pertaining to the earth surface then ever before, and with applicable use and help of remote sensing techniques and applications, users have the ability to map out the earth surface and provide some insightful evidence to that the earths surface looks like in at this point in history.

Study Area:
The user was provided an original raster image, shown in figure 1, that was taken from the Landsat thematic raster image with 30 meter resolution of an the South-Central Ontario. The Landsat raster image provides a false colour representation of the region of Ontario, providing a detailed account of the vegetative and bare surface on Earth.

Figure 1: Landsat Raster Imagery

The user was responsible for selecting a subset region (512 pixels by 512 pixels) of southern Ontario which they could perform a land classification on. In this scenario, the region of Orillia, Ontario was selected to conduct a land classification on, not visible with a true colour image.

32 Robertson Rd., Niagara-on-the-Lake, L0S 1J0 905.932.1998

Figure 2: Subset image of Orillia, Ontario, Canada

The region of Orillia was selected for its condensed urban space and numerous vegetative areas located in a condensed space, specifically agricultural land and forest area surrounding the city. For a user to gain knowledge of land classification it would not be appropriate to select a subset region that would comprise of a large majority of water bodies, as their reflective aspects would be low and in result, exemplify very little analysis.

Unsupervised Classification
The primary procedure which is required to be undertaken for this scenarios land classification is the primary, user interactive, unsupervised classification. Although the procedures name indicates unsupervised, the user has a considerable amount of interactive touches throughout this procedure. Unsupervised classification finds spectral classes in a multi band image without the need to analyses the entire process. The unsupervised classification process is aided by the image classification toolbar extension in ERDAS Imagine 2010, that provides access to certain tools, clusters, cluster analyze and classification tools. The use of unsupervised classifications to assign each pixel cell in within the subset image into a feature class (trees, water bodies, etc). In this process, users want to aggregate each location in the subset into one a specified number of clusters or groups. To determine where each cell is placed, each cluster will be assigned dependent on the cells multivariate statistics. Each cluster is independent in that it is statistically different from the other clusters based on the values for each band within the multivariate.

32 Robertson Rd., Niagara-on-the-Lake, L0S 1J0 905.932.1998

In this process, the use of unsupervised training is important in establishing a good image. Being computer automated, it enables the user to specify parameters that the computer uses to identify statistical patterns within the data. The process identifies clusters of pixels with similar spectral characteristics. This specific training process is dependent on the data and the definitions of the classes identified. In unsupervised, users are not familiar with the study areas, so this method is used at point to fill the gaps or when classes are appropriately interoperated.

Figure 3: Unsupervised Class Section.

Within the process which was conducted for this study, the number of classes were specified by the user at the immediate beginning. In this regard, the user selected 8 classes which were to be identified in the unsupervised process.

32 Robertson Rd., Niagara-on-the-Lake, L0S 1J0 905.932.1998

Figure 4: Unsupervised Classification Class Names

Following the computer computing the data, 8 classes were identified: 1. Border 2. Fresh Water 3. Forest 4. Forest 5. Forest 6. Agriculture/urban 7. Agriculture/ urban 8. Agriculture 9. Agriculture For more clarity of the study area to be understood, within the unsupervised classification process, recoding had to be undertaken. The recoded values are: 1. 2. 3. 4. Fresh Water Forest Agriculture/Urban Agriculture

32 Robertson Rd., Niagara-on-the-Lake, L0S 1J0 905.932.1998

Figure 5: Unsupervised Classification Recode

Figure 6: Unsupervised Classification Recode Pixel Identification

32 Robertson Rd., Niagara-on-the-Lake, L0S 1J0 905.932.1998

The problem with recoding was that it did not provide enough clarity for the user to distinguish the features on the ground. It grouped features together which were not of the same pixel characteristics.

Figure 7: Unsupervised Classification Recode Image

Having the user undertake an unsupervised land classification it is evident that the user did not find the unsupervised process to be the most accurate in identifying and distinguishing between different features. Question: What does the x-axis show and what does the y-axis show? The x-axis is the identification of pixel value and the y-axis illustrates the frequency.

32 Robertson Rd., Niagara-on-the-Lake, L0S 1J0 905.932.1998

Supervised Classification:
Supervised Classification uses the spectral signatures obtained from training sample sites to classify an image. With the assistance of the Image Classification toolbar, as was used in Unsupervised, the user has the ability to create training sites at their disposal and preference. Once the training sites have been created, the signature files can be obtained with ease. The signature file is then used by the Spatial Analyst multivariate classification to identify and classify the image.

32 Robertson Rd., Niagara-on-the-Lake, L0S 1J0 905.932.1998

Figure 8: Supervised Classification Signature Input

Supervised Classification:
There are three procedures that user has the advantage to select when undertaking a supervised classification. These three optional procedures; maximum likelihood, mahalanobis distance, and minimum distance, are selected in the dialogue box prior to the processing the supervised land classification. Maximum likelihood
Maximum Likelihood uses both variance and covariance to classify pixels. Assuming that the distribution of each category has normal spread, assumptions can be made about classifying the data. Undefined pixels will be defined by identifying the probability of that pixel falling into each category. The category with the highest likelihood (probability) will be selected for classification. This classifications limitation is its mathematical calculations for every pixel are complicated therefore the computational power required to classify a large image will be time consuming. Figure # 1 illustrates the issues with this maximum likelihood classification. For this particular classification maximum likelihood produced the best result among the three techniques listed. The image size is relatively small and took very little time to classify. In particular, this classification defined roads extremely well which is impressive due to the similarity between roads and urban spectral signatures. The maximum likelihood classification did not classify all bare fields correctly, some fields were classified as urban areas as seen in site # 1. See appendix A for the maximum likelihood classification.

10

32 Robertson Rd., Niagara-on-the-Lake, L0S 1J0 905.932.1998

Figure 9: Maximum Likelihood

Mahalanobis distance
Mahalanobisdistance is very similar classification technique to maximum likelihood in that is also uses both variance and covariance. The difference is that mahalanobis assumes that every categories covariance is equal. Because of this difference classification takes far less computational power. See appendix B for the mahalanobis distance classification. Figure # 2 illustrates the issues with this mahalanobis distance classification. For this particular image classification vegetation was classified correctly but the urban/ roads/ bare fields were done poorly as seen in sites # 1, 2 and 3. The spectral separation was too similar for this classification technique.

Figure 10: Mahalanobis Distance

11

32 Robertson Rd., Niagara-on-the-Lake, L0S 1J0 905.932.1998

Minimum Distance (Minimum Distance to Means)

Minimum Distance is a form of supervised classification in which pixels can be classified by the distance between that pixel and the various category averages. Because the mathematical calculations required to classify unknown pixels is relatively simple this classification is computationally efficient. The restrictions for this classification are apparent when classifying categories that have similar spectral signatures. Minimum distance is not sensitive to data that contains a large variance in spectral signatures within a single category. Figure # 3 illustrates the issues with this minimum distance classification. For the purposes of this particular classification minimum distance to means was effective at classifying vegetation, healthy fields, deep water and shallow water. Conversely, classification of urban areas, roads and bare fields was done poorly as seen at site # 1. This is because urban areas have a large variance in spectral signatures and pixels were only classified based on their distance to bare fields or roads category means. See appendix C for the minimum distance to means classification.

Figure 11: Minimum Distance (Minimum Distance to Means)

Classification Comparison:
Supervised
In a supervised classification the user creates numerical descriptors of the various categories by selecting training sites. The selected computer algorithm will then sort the pixels based on the user defined training sites. Please see appendix A for the maximum likelihood supervised classification.

Advantage
The advantage of the supervised classification is the users ability to define training sites that best represent a land surface. The ability to choose between various mathematical classifications techniques within a supervised classification also allows for further ability to classify an image. The choice of training sites in this classification created a spectral separability that enabled the classification to yield better results than that of the unsupervised classification.

Disadvantage
The apparent disadvantage of the supervised classification is user error. The image analyst doing the supervised classification has to have a previous knowledge of the area being supervised to yield the most accurate results when selecting training sites.

12

32 Robertson Rd., Niagara-on-the-Lake, L0S 1J0 905.932.1998

Unsupervised
Unsupervised classification involves no training site selection; instead classification is based on natural clusters that occur within the image. These clusters are derived from computer algorithms and sorted into a user defined number of categories. These spectral groupings are not known to the user and class identity is created by referencing the classified image to a reference image. Please see appendix D for the unsupervised classification.

Advantage
The advantage to an unsupervised classification is the absence of user error. The user is limited to only specifying number of classes and iterations. The user does not require any previous knowledge of the area. This particular unsupervised classification did the best job classifying water; all water was classified correctly with no issues, even the water found in small pockets inland were correctly classified. This classification has issues separating urban and residential land cover.

Disadvantage
The disadvantage to an unsupervised classification is the lack of user defined classes. There are only a user defined number of classes applied to the image. These classes may not include classes that the user wants to identify. This particular classification could have had more classes added but the unclassified would still lack user input as to which land cover types will be addressed.

Differences between the two

The main difference between both the supervised and unsupervised classification is identifiable by addressing how each method handled land cover classification. Classifying water was done well by both methods, unsupervised identified all water including small ponds inland correctly, however if the user wanted to identify shallow or turbid water this classification method is not the best option. The supervised classification of the same area identified both deep and shallow water; however there were some issues with some transitional zones being classified as urban area. Distinguishing urban and roads were done effectively by supervised classification, providing the best result among all classifications when it came to roads. The unsupervised classification did have issues separating urban from residential and bare fields. Vegetation was classified well by both methods. The unsupervised classification distinguished between vegetative land and crop fields well. The supervised classification identified these as well as bare fields however there was some issues in regards to the similarity between urban and bare fields.

Work Cited:
Lillesand, Thomas, R Kiefer, and J Chipman.Remote Sensing and Image Interpretation. 6. United States: John Wiley & Sons, 2008. 123-188. Print.

13