classification - lcluc programlcluc.umd.edu/sites/default/files/lcluc_documents/... ·...
TRANSCRIPT
![Page 1: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/1.jpg)
Crop Classification
Classification
![Page 2: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/2.jpg)
Classification
• Key remote sensing processing technique
• Placing pixels into thematic categories that correspond to land cover types – e.g. forest, crops, water, urban, etc.
• Basis for classification are the spectral signatures of landcover types
![Page 3: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/3.jpg)
![Page 4: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/4.jpg)
Landsat TM Bands
![Page 5: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/5.jpg)
Spectral Resolution of Landsat TM
TM Band: 1 2 3 4 5
These bands provide a coarse summary of spectral signatures.
(NIR) (MIR)
![Page 6: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/6.jpg)
Feature Space (2 bands)
Red BV’s
NIR BV’s
0
0 255
255
v
![Page 7: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/7.jpg)
Feature Space
Pixels in the same classes cluster
water
soil
agri
forest
Red BV’s
Infrared BV’s
0
0 255
255
![Page 8: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/8.jpg)
Classification
• Two essential types of classification – Unsupervised
• Unsupervised by humans
• Computer algorithm-based
– Supervised • Supervised by human(s)
• Human-trained
• Algorithms also used
![Page 9: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/9.jpg)
Unsupervised Classification • Classes are created based on the
locations in feature space of the pixel data
Red BV’s
Infrared BV’s
0
0 255
255
v
![Page 10: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/10.jpg)
Computer Algorithm Finds Clusters
Red BV’s
Infrared BV’s
0
0 255
255
v
Unsupervised Classification
![Page 11: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/11.jpg)
Unsupervised Classification
• Attribution phase – performed by human
water
Soil
agri
forest
Red BV’s
Infrared BV’s
0
0 255
255
![Page 12: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/12.jpg)
Problems with Unsupervised Classification
Red BV’s
Infrared BV’s
0
0 255
255
v
The computer may consider these 2 clusters (forest and agri) as one cluster The computer may consider
this cluster to be 2 clusters
![Page 13: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/13.jpg)
Supervised Classfication • We “train” the computer program using
ground truth data
Coniferous trees Deciduous trees
![Page 14: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/14.jpg)
Supervised Classification
Red BV’s
Infrared BV’s
0
0 255
255
v
Sample pixels
Other pixels
![Page 15: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/15.jpg)
Problems with Supervised Classification
Red BV’s
Infrared BV’s
0
0 255
255
v forest
agri
water
Soil
What’s this?
v
![Page 16: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/16.jpg)
The Result: A Classified Image
![Page 17: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/17.jpg)
Classification scheme (important for both supervised and unsupervised classifications)
• Classification scheme: the set of classes used to classify an image – Classes chosen to be appropriate for use of data – Sufficient number of classes chosen
• Don’t want pixels that don’t resemble any samples (how would we classify these?)
– Good to use or adapt standard classification schemes • Helpful to compare your analysis with other analyses • Standard hierarchical schemes exist, and you can choose
the level of detail appropriate for your goals – Common classification systems used in Europe:
» CORINE (Coordination of Information on the Environment) » LCCS (Land Cover Classification System)
![Page 18: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/18.jpg)
Key advantages of supervised classification (relative to unsupervised)
• Images using supervised classification are more easily comparable to other images (of different dates or neighboring regions).
• Class definitions are based on data about training samples of known landcover type. – Potential for higher accuracy
• Avoids the difficult attribution stage of unsupervised classification.
![Page 19: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/19.jpg)
Key disadvantages of supervised classification (relative to unsupervised)
• Supervised classification assigns pixels to one of the user-defined classes. Some pixels may represent a landcover type that doesn’t fit one of those classes.
• One class may have a highly variable signature (possibly bimodal or ‘multimodal’) – this can hinder the classification process.
• Getting good, representative training data is important but may not be possible.
• Process is substantially slower & more expensive than unsupervised.
![Page 20: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/20.jpg)
Break for statistics brief
![Page 21: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/21.jpg)
Concept of Variance
DN value DN value
Low variance High variance
1,000
2,000
3,000
4,000
5,000
3,000
6,000
9,000
12,000
15,000
# of pixels
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
![Page 22: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/22.jpg)
A Distribution
DN value
frequency
0
5000
10,000
15,000
20,000
0 255
![Page 23: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/23.jpg)
Another Distribution
DN value
0
5000
10,000
15,000
20,000
0 255
Does this distribution have more or less variance than the one on the previous slide?
![Page 24: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/24.jpg)
Variance – covariance
Band #1
Band #2
Band #1
Band #2
In addition to variances for each band, a covariance is calculated which measures how one band tends to vary with the other band.
Positive covariance Negative covariance
![Page 25: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/25.jpg)
Band #1
Band #2
Bivariate distribution Another look
Zero covariance
![Page 26: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/26.jpg)
Distances
Band #1
Band #2
Band #1
Band #2
Distances can be calculated in terms of variances and covariances for each class.
Mean (center) of class Mean (center) of class
The concentric ellipses are densities of points. The more “rings” out from the center a point is, the farther away it is from the center in terms of the variance-covariance of that class.
In terms of variance-covariance, for which class is the green dot near to the mean of that class?
Class 1 (e.g. forest) Class 2 (e.g. urban)
![Page 27: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/27.jpg)
Back to remote sensing: ISODATA method of
Unsupervised Classification
The most commonly used unsupervised method.
![Page 28: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/28.jpg)
ISODATA Algorithm
Red BV’s
Infrared BV’s
0
0 255
255
v
Assume these are the data of a satellite image to be classified.
Here we only use 2 bands of data, but the process can be generalized for any number of bands.
![Page 29: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/29.jpg)
ISODATA Algorithm
Red BV’s
Infrared BV’s
0
0 255
255
v
Step 1: ISODATA algorithm creates arbitrary classes to “seed” the process
Class 3 Class 1
Class 5
Class 4
Class 2
Class 6
Class 7
![Page 30: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/30.jpg)
ISODATA Algorithm
Red BV’s
Infrared BV’s
0
0 255
255
v
Step 2: ISODATA algorithm calculates the mean, variance, and covariance of the data within each class.
Class 3 Class 1
Class 5
Class 4
Class 2
Class 6
Class 7
Class 1 mean
Class 2 mean
Class 3 mean
Class 4 mean
Class 5 mean
Class 6 mean
Class 7 mean
![Page 31: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/31.jpg)
ISODATA Algorithm
Red BV’s
Infrared BV’s
0
0 255
255
v
Step 3: ISODATA algorithm finds the nearest class mean to each pixel (in terms of variance-covariance distance).
Class 3 Class 1
Class 5
Class 4
Class 2
Class 6
Class 7
Class 1 mean
Class 2 mean
Class 3 mean
Class 4 mean
Class 5 mean
Class 6 mean
Class 7 mean
Pixel in question
![Page 32: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/32.jpg)
ISODATA Algorithm
Red BV’s
Infrared BV’s
0
0 255
255
v
Step 4: ISODATA algorithm reallocates each pixel to its new class, based on which class mean it’s nearest to.
Class 3 Class 1
Class 5
Class 4
Class 2
Class 6
Class 7
Class 1 mean
Class 2 mean
Class 3 mean
Class 4 mean
Class 5 mean
Class 6 mean
Class 7 mean
Closest to Class 2!
![Page 33: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/33.jpg)
ISODATA Algorithm
Red BV’s
Infrared BV’s
0
0 255
255
v
Step 5: ISODATA algorithm re-draws the class boundaries to accommodate the new class assignments.
Class 3 Class 1
Class 5
Class 4 Class 2
Class 6
Class 7
Class 1 mean
Class 2 mean
Class 3 mean
Class 4 mean
Class 5 mean
Class 6 mean
Class 7 mean
![Page 34: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/34.jpg)
ISODATA Algorithm
Red BV’s
Infrared BV’s
0
0 255
255
v
Go back to Step 2: ISODATA algorithm calculates the mean, variance, and covariance of the data within each new class.
Class 3 Class 1
Class 5
Class 4 Class 2
Class 6
Class 7
Class 1 mean
Class 2 mean
Class 3 mean
Class 4 mean
Class 5 mean
Class 6 mean
Class 7 mean
![Page 35: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/35.jpg)
This iterative process continues until either:
1. A predefined number of iterations finish, or
2. the proportion of pixels that change classes between iterations becomes less than a predefined threshold.
ISODATA Algorithm
![Page 36: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/36.jpg)
Result of unsupervised classification
• A map with each pixel assigned to a particular (numbered) class.
• Next step: attribution – A particular landcover category is determined
and assigned to each numbered class.
– Usually done by the analyst.
![Page 37: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/37.jpg)
Final Result: A Classified Image
![Page 38: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/38.jpg)
Supervised classification
• Parallelepiped method
• Maximum likelihood method
![Page 39: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/39.jpg)
Parellepiped Method
Red BV’s
Infrared BV’s
0
0 255
255
Sample pixels are plotted in spectral space using a subset of bands (in this case, 2 bands, so the plot is 2-dimensional).
Field sample pixels
Forest sample pixels
Water sample pixels
Urban sample pixels
![Page 40: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/40.jpg)
Parellepiped Method
Red BV’s
Infrared BV’s
0
0 255
255
Boxes (parallelepipeds) are drawn around the sample pixles of each class.
Field sample pixels
Forest sample pixels
Water sample pixels
Urban sample pixels
![Page 41: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/41.jpg)
v
Parellepiped Method
Red BV’s
Infrared BV’s
0
0 255
255
Then the full dataset is plotted in this spectral space with the parallelepipeds (all pixels – not just sample pixels).
Field sample pixels
Forest sample pixels
Water sample pixels
Urban sample pixels
![Page 42: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/42.jpg)
v
Parellepiped Method
Red BV’s
Infrared BV’s
0
0 255
255
All of the image pixels that fall within each parallelepiped are classified as the same class as the sample pixels within that parallelepiped.
Field sample pixels
Forest sample pixels
Water sample pixels
Urban sample pixels
Note: parallelepipeds can be “drawn” in more than 3 dimensions to incorporate as many bands of data as desired.
![Page 43: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/43.jpg)
Parallelepiped Method
• One of the first methods developed for supervised classification
• Intuitive • Leaves some pixels unclassified • Other methods are more accurate • A variation on this method is still often used:
– Instead of the computer algorithm drawing the parallelepipeds, an analyst can draw them
– Shapes can be of any kind (not confined to parallelepipeds)
– This can be a good, highly accurate method to classify some or all of an image
![Page 44: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/44.jpg)
Maximum Likelihood Method
Red BV’s
Infrared BV’s
0
0 255
255
Sample pixels are plotted in spectral space using all bands. Here only 2 bands are shown.
Field sample pixels
Forest sample pixels
Water sample pixels
Urban sample pixels
![Page 45: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/45.jpg)
Maximum Likelihood Method
Red BV’s
Infrared BV’s
0
0 255
255
Class means, variances, and covariances are calculated.
Field sample pixels
Forest sample pixels
Water sample pixels
Urban sample pixels
![Page 46: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/46.jpg)
v
Maximum Likelihood Method
Red BV’s
Infrared BV’s
0
0 255
255
Then each (non-sample) pixel is assigned to the class for which its BV’s are most likely, assuming the distribution of pixels in each class is multivariate normal.
Field sample pixels
Forest sample pixels
Water sample pixels
Urban sample pixels
Pixels that fall below a certain threshold (in terms of likelihood for any class) are sometimes left unclassified, as shown.
![Page 47: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/47.jpg)
Maximum Likelihood • How is “most likely” determined?
• Consider the problem in the one band case.
DN values in a particular band (i.e., NIR) 0 255
Distribution of sample pixels in forest class
Distribution of sample pixels in ag/field class
Probability (Probability that a randomly chosen forest pixel has the given DN value)
![Page 48: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/48.jpg)
Consider curves in terms of probability
• The height of the curve at each point shows the relative probability that a random pixel in the forest class has that DN value.
• Area under curve equals 1, since all forest pixels fall somewhere under this curve
DN values in a particular band (i.e., NIR) 0 255
Probability (Probability that a randomly chosen forest pixel has the given DN value)
Distribution of sample pixels in forest class According to the sample
data, forest pixels are more likely to be have this DN value than this DN value
![Page 49: Classification - LCLUC Programlcluc.umd.edu/sites/default/files/lcluc_documents/... · Classification • Key remote sensing processing technique • Placing pixels into thematic](https://reader030.vdocuments.net/reader030/viewer/2022040608/5ec61d1b3499d74ad44d208e/html5/thumbnails/49.jpg)
How is “most likely” determined?
• Pixels with a particular DN value are assigned to the class for which the likelihood (probability) is the highest.
• Therefore, the black vertical line represents the cutoff point for class membership.
DN values in a particular band (i.e., NIR) 0 255
frequency (# of pixels with each DN value)
Distribution of sample pixels in forest class
Distribution of sample pixels in ag/field class
Highest probability = highest Maximum Likelihood
This is generally done with multiple bands (in multiple dimensions).