Sunteți pe pagina 1din 14

[135 Taylor Road Niagara-on-the-Lake Ontario L0S 1J0] [Nathan Page]

January 28, 2014 GISC9216-D1 Mrs. Janet Finlay GIS-GM Program Coordinator Niagara College 135 Taylor Road Niagara-on-the-Lake, ON L0S 1J0 Dear Mrs. Finlay, RE: GISC9216-Deliverable 1-Introduction to Supervised Classification Please accept this cover letter as my formal submission of Deliverable 1: Introduction to Supervised Classification for GISC9216-Digital Image Processing. The deliverable files are located in the X: drive under GISC9216\Assignment#1\PageNGISC9216D1. This assignment proved as a great introduction to processing digital images; having zero background in remote sensing and very little in ERDAS, I was frequently stuck trying to figure out which little step I had forgotten or carried out incorrectly, but it was very enjoyable to produce the final classifications. Should you have any questions regarding the enclosed documents, or if there are technical issues regarding the files please contact me at your convenience at nathanpage90@gmail.com. Thank you for your time and attention. I look forward to your comments and suggestions. Warmest Regards,

Nathan Page BA, GIS-GM Certificate Candidate N.P./n.p.

[135 Taylor Road Niagara-on-the-Lake Ontario L0S 1J0] [Nathan Page]

[135 Taylor Road Niagara-on-the-Lake Ontario L0S 1J0] [Nathan Page]

Contents
Discussion...................................................................................................................................................... 3 Conclusion ................................................................................................................................................... 12

Figure 1 Water Signature Histograms ........................................................................................................... 3 Figure 2 Vegetation Signature Histograms ................................................................................................... 4 Figure 3 Residential Signature Histograms ................................................................................................... 5 Figure 4 Commercial/Industrial Signature Histograms ................................................................................. 6 Figure 5 Agricultural Signature Histograms .................................................................................................. 7 Figure 6 Maximum Likelihood Supervised Classification .............................................................................. 9 Figure 7 Mahalanobis Supervised Classification ......................................................................................... 10 Figure 8 Minimum Distance Supervised Classification ............................................................................... 11 Figure 9 Unsupervised Classification .......................................................................................................... 12 Figure 10 Supervised Classification ............................................................................................................. 13

[135 Taylor Road Niagara-on-the-Lake Ontario L0S 1J0] [Nathan Page]

[135 Taylor Road Niagara-on-the-Lake Ontario L0S 1J0] [Nathan Page]

Discussion
When evaluating our signatures for this assignment, we viewed each signatures multiple histograms in order to get a feel for how each bands pixel values are distributed. The x-axis shows the value of an individual pixel, whereas the y-axis shows how many pixels of that value exist in the image in question. For this assignment I ended up creating 5 signatures; Water, Vegetation, Residential, Agricultural and Commercial/Industrial. When comparing the various bands histograms within this assignment there are many things to be noted.

The signature that was created for water appears below (Figure 1 Water Signature Histograms). One can see that these histograms are Leptokurtic and for the most part are Unimodal (The first three bands have one or two very small peaks but they arent large enough to have any significant impact. This tells us that the pixels within the Water signature layer do not vary too much and are all concentrated roughly around the same pixel value.

Figure 1 Water Signature Histograms

Moving on we see the histograms for the Vegetation signature (Figure 2 Vegetation Signature Histograms). These histograms are generally unimodal but the RGB bands are more mesokurtic; not as
[135 Taylor Road Niagara-on-the-Lake Ontario L0S 1J0] [Nathan Page]

[135 Taylor Road Niagara-on-the-Lake Ontario L0S 1J0] [Nathan Page]

tightly packed together. This tells us that the pixels vary somewhat, more than the water signature, but are still normally distributed.

Figure 2 Vegetation Signature Histograms

Moving on to the Residential signature, the histograms exhibit a more sporadic distribution; they are more so meso and platykurtic than the water and vegetation signatures. This sporadic distribution of pixels can possibley be attributed to the fact that residential building greatly vary in colour, material, size, and as a result have more varied pixel values.

[135 Taylor Road Niagara-on-the-Lake Ontario L0S 1J0] [Nathan Page]

[135 Taylor Road Niagara-on-the-Lake Ontario L0S 1J0] [Nathan Page]

Figure 3 Residential Signature Histograms

Moving on to the Commercial/Industrial signature, it can be pointed out that the histograms are bimodal and multi-modal but are also quite mesokurtic. This could be attributed to the same reasons as just previously stated; buildings vary greatly in colour, size and materials used and so the pixel values of the remotely sensed images of them will reflect that increased variance in value.

[135 Taylor Road Niagara-on-the-Lake Ontario L0S 1J0] [Nathan Page]

[135 Taylor Road Niagara-on-the-Lake Ontario L0S 1J0] [Nathan Page]

Figure 4 Commercial/Industrial Signature Histograms

Finally we come to the agriculture signature. These histograms are extremely leptokurtic, but are bi and tri-modal, suggesting that within my area there are two and possibly three distinct agricultural land types. Besides that though the high leptokurtic state of the histograms suggests that the pixel values have little variation.

[135 Taylor Road Niagara-on-the-Lake Ontario L0S 1J0] [Nathan Page]

[135 Taylor Road Niagara-on-the-Lake Ontario L0S 1J0] [Nathan Page]

Figure 5 Agricultural Signature Histograms

Moving on to the supervised classification method, we come to a fork in the road. There are three algorithms, three ways to classify, each three with specific parameters rendering slightly different results. The first, Minimum Distance uses the spectral distance between the candidate pixel and the average vector of each signature. Once that distance is calculated for all possible classes then the lowest spectral distance gets assigned the class of the candidate pixel. Mahalanobis is similar to the previous type of classification except that variance and covariance are worked into the equation in order to deal with clusters of pixels that are highly varied. The final type of classification (Maximum likelihood) uses probability statistics to decide the probability of whether a pixel belongs to a certain class or not. It assumes normal distributions for the input bands and because of this reason can sometimes be unsuitable when dealing with bands that do not have a normal distribution of the data. Each of these three classification types makes sure to keep a certain aspect of the data in mind, whether it be the variability of classes (Mahalanobis), the probability of pixels and the class they belong too (Maximum Likelihood) or the actual distance between the pixels and the mean vector for each signature (Minimum Distance). Each type has its advantages and disadvantages which are appropriate for specific situations.
[135 Taylor Road Niagara-on-the-Lake Ontario L0S 1J0] [Nathan Page]

[135 Taylor Road Niagara-on-the-Lake Ontario L0S 1J0] [Nathan Page]

For this assignment, I decided to go with the default option of the Maximum Likelihood Classification. I performed all three (Figures 6, 7 and 8) and examined and compared them too each other. Based upon my ground truthing via Google Earth and my personal knowledge of this area, I decided that Max. Likelihood displayed the most accurate representation of the mix between forests and the vegetation as well as the mixed areas of residential and industrial/commercial areas.

[135 Taylor Road Niagara-on-the-Lake Ontario L0S 1J0] [Nathan Page]

[135 Taylor Road Niagara-on-the-Lake Ontario L0S 1J0] [Nathan Page]

Figure 6 Maximum Likelihood Supervised Classification

[135 Taylor Road Niagara-on-the-Lake Ontario L0S 1J0] [Nathan Page]

[135 Taylor Road Niagara-on-the-Lake Ontario L0S 1J0] [Nathan Page]

Figure 7 Mahalanobis Supervised Classification

[135 Taylor Road Niagara-on-the-Lake Ontario L0S 1J0] [Nathan Page]

[135 Taylor Road Niagara-on-the-Lake Ontario L0S 1J0] [Nathan Page]

Figure 8 Minimum Distance Supervised Classification

[135 Taylor Road Niagara-on-the-Lake Ontario L0S 1J0] [Nathan Page]

[135 Taylor Road Niagara-on-the-Lake Ontario L0S 1J0] [Nathan Page]

Conclusion
For this assignment we performed two classification types on a specific area of interest of our choice within the GTA. We first performed an unsupervised classification on this area of interest then we performed a supervised classification using layer signatures that we created from training sites within our Area of Interest. For my unsupervised classification I left the default number of classes, which were 20. This created a very busy looking output. As you can see below (Figure 2 Unsupervised Classification), the areas of the cities are all one similar shade, making it almost impossible to distinguish between residential, industrial and commercial areas. Furthermore the farm fields and forests are not easily discernible within this unsupervised classification; they are two shades of red that are very close in color making it very difficult to discern one from the other, especially in highly diversified areas such as what I have in my Area of Interest where theres agriculture and natural vegetation mixed in together throughout the whole area.

Figure 9 Unsupervised Classification

This Unsupervised Classification clearly needed to be given some revision and some direction so as to more clearly and easily discern each of the major categories of land use that I chose which consisted of: Water, Vegetation, Agriculture, Residential, and Commercial/Industrial.
[135 Taylor Road Niagara-on-the-Lake Ontario L0S 1J0] [Nathan Page]

[135 Taylor Road Niagara-on-the-Lake Ontario L0S 1J0] [Nathan Page]

Moving on to the Supervised Classification, I chose to limit each category to just one class and one color so as to more clearly convey the contrast of used land between the major land use types rather than have multiple classes within each major land use type and distinguish differences within each major land use type. As can be seen below (Figure 10 Supervised Classification), its not difficult to discern where the farm lands are (Brown), where the natural plots of vegetation/forest cover are (Green), where the lakes, canals and rivers are (Blue), where the cities are (Grey) and also where the commercial/industrial areas are within the cities (White).

Figure 10 Supervised Classification

The advantages and disadvantages between these two types of classifications are numerous. If you perform an unsupervised classification you dont have to have prior knowledge of the cite that youre investigating it automatically generates unbiased training sites and gives you results. The disadvantage of this is that the results do not take certain factors into consideration such as variance and covariance, and also it can leave many pixels unidentified making it difficult to read the final output. To summarize the advantage is that unsupervised is quick and easy, good for preliminary and early stage work, but it
[135 Taylor Road Niagara-on-the-Lake Ontario L0S 1J0] [Nathan Page]

[135 Taylor Road Niagara-on-the-Lake Ontario L0S 1J0] [Nathan Page]

can end up being time consuming and can contain errors and unidentified pixels. The advantage of the supervised classification is that the researcher/analyst has more control over the parameters. This is most appropriate when the analyst is highly knowledgeable of the area of interest and furthermore, it allows for the processing to be tightly and accurately tied to the areas of known identity, resulting in a much more accurate and reasonable output. This advantage, however, can serve as a disadvantage if what the analyst thinks that what he knows is in fact inaccurate. If this is the case the output classification can end up being very skewed and inaccurate. Each type of classification is sensitive to these varying factors, making them both highly advantageous in certain situations; the analyst must layout their area and figure out how much they know of the area in order to decide which method will benefit their purposes the most.

[135 Taylor Road Niagara-on-the-Lake Ontario L0S 1J0] [Nathan Page]

S-ar putea să vă placă și