Presentation is loading. Please wait.

Presentation is loading. Please wait.

Statistical Classification on a Multispectral Image

Similar presentations


Presentation on theme: "Statistical Classification on a Multispectral Image"— Presentation transcript:

1 Statistical Classification on a Multispectral Image
Susan Kolakowski Kristin Strackerjan Donald Taylor Seth Weith-Glushko Multivariate Statistics for Imaging Science May 12, 2005

2 Outline Introduction Background Procedure Results/Analysis Conclusion
Questions I will give you a brief introduction and then Kristin will give some background. Seth will talk about the procedure. Don will talk about the results/analysis. I will close with the conclusion and then we will allow some time for questions.

3 Introduction MISI Image Supervised Unsupervised Analysis Methods
Gaussian Maximum Likelihood (GML) Unsupervised Minimum Distance to the Mean (MDM) Analysis Methods Confusion matrices, Jeffries-Matusita distances, transformed divergences

4 Background Why classify an image?
Separate pixels into different classes based on their spectral information Human visual system does not distinguish between different classes of material outside of the visible spectrum (RGB vs VIS/NIR) Automate this process to remove as much bias as possible

5 Background Methods of Classification (1) Unsupervised Classification
Little interference from the user Input parameters: Number of classes Threshold Number of iterations K-means vs. ISODATA

6 Background Methods of Classification (2) Supervised Classification
Gaussian Maximum Likelihood (GML) classifier Use Regions of Interest (ROIs) defined in ENVI to train the GML classifier GML is calculated based on the probability of pixel X belonging to class i:

7 Background Separability of Classes
In order to quantify the degree of overlap between potential classes, we used Transform Divergence and the Jeffries-Matusita (JM) distances Both of these methods allow us to see how separable the data is by comparing groups of three bands at a time

8 Background Transformed Divergence

9 Background Jeffries-Matusita (JM) Distance

10 Background Quality of the Classifier Confusion Matrices
Dependent data set Independent data set Randomly selected data

11 Procedure Selection of test data
Code generation for each algorithm in the IDL programming environment Run the classification algorithm and generate a classification map Calculate the best three-band combo for display using Jeffries-Matusita distance and transformed divergence Execute the unsupervised algorithm Select training data based off of unsupervised classification map

12 Original Image with Regions of Interest overlaid

13 Procedure Run the supervised classification algorithm
Calculate quality metrics Visual comparison of the classification maps Confusion matrices Visual inspection of scatterplots generated with Jeffries-Matusita distance and transformed divergence

14 Unsupervised Color Class Contents Water Sand/Dirt Grass Trees
Roads/Pavement Urban

15 Unsupervised – Classification Matrices

16 Unsupervised - Scatter Plots

17 Supervised

18 Supervised – Classification Matrices

19 Supervised - Scatter Plots

20 Conclusion Supervised vs. Unsupervised Advantages and Disadvantages
MDM vs. GML

21 Questions??


Download ppt "Statistical Classification on a Multispectral Image"

Similar presentations


Ads by Google