Training Fields Parallel Pipes Maximum Likelihood Classifier Class 11. Supervised Classification.

Post on 01-Apr-2015

214 views 0 download

Tags:

transcript

Training Fields

Parallel Pipes

Maximum Likelihood Classifier

Class 11. Supervised Classification

Unsupervised classification is a processof grouping pixels that have similar spectral values and labeling each group with a class

Definition

Supervised classification is to classify animage using known spectral information foreach cover type

1. Training Fields (minimum spectral distance)

A sample area for estimating representative spectral statistics,or spectral signatures.

A seed-pixel approach can be used (page 137, Verbyla) according to the minimum distance classifier

Verbyla 7.0

Two-band image

AB: Aspen/Birch

SM: Sedge/Medow

Lillisand & Keifer 7.0

2. Parallelpiped classifier

Define max/min for each band for each class

If a class has normally distributed spectral valuesthen 95% of pixels are within mean±2 standard deviations, i.e.,

Minimum = mean-2×SDMaximum = mean+2×SD

Max/min can be adjusted according to needs

Step-wiseparallelpipes

3. Maximum likelihood classifier

From the training field, create contours of equal likelihood for each class. The highest likelihood for a candidate pixel determines the class of the pixel

Single-band example

From training fields for cattail (CT) and smartweed (SW)

Mean digital value Standard deviation

()

Number of pixels

CT 30 5 100

SW 20 5 100

Class 12

Assessment of classification Accuracy

Error Matrix (confusion matrix)

User’s AccuracyProducer’s Accuracy

Overall AccuracyKappa Statistics

Error MatrixGround Truth

1 2 3 4 5 Row total

1 40 0 0 3 0 43

2 0 30 12 0 1 43

3 0 3 25 0 2 30

4 2 0 0 50 0 52

5 0 0 0 0 32 32

Column total

42 33 37 53 35 200

Pre

dic

ted

class

class

Verbyla 8.0

Overall Classification Accuracy

It is the total number of correct class predictions(the sum of the diagonal cells) divided by the total number of cells.

In this case, it is (40+30+25+50+32)/200 =88%

Producer’s and user’s accuracy by cover type class

Class Producer’s Accuracy User’s Accuracy

1 40/42=95% 40/43=93%

2 30/33=91% 30/43=70%

3 25/37=68% 25/30=83%

4 50/53=94% 50/52=96%

5 32/35=91% 32/32=100%

Kappa Statistic

KHAT=Overall Classification Accuracy – Expected Classification Accuracy

1 – Expected Classification Accuracy

The expected classification accuracy is the accuracy expected based on chance,Or the expected accuracy if we randomly assigned class values to each pixel. In this case (see the next slide), it is (1806+1419+1110+2756+1120)/40,000=21%

In this case, KHAT=(0.88-0.21)/(1-0.21)=0.85

Products for KHATGround Truth

1 2 3 4 5 Row total (error matrix)

1 1806 1419 1591 2279 1505 43

2 1806 1419 1591 2279 1505 43

3 1260 990 1110 1590 1050 30

4 2184 1716 1924 2756 1820 52

5 1344 1056 1184 1696 1120 32

Column total (error matrix)

42 33 37 53 35 200

Pre

dic

ted

classclass