8/7/2019 Lecture 6R - App. RS
1/22
15/2/0515/2/05 App. RSApp. RS 11
Unsupervised Classification: In many cases you dont have the necessary
beforehand knowledge of the land cover needed toperform a supervised classification in which case anunsupervised classification can be used.
- the definition, identification, labeling, andmapping of natural classes with respect to brightnessin several spectral channels;
General steps Automatic grouping of pixels into similar spectral
classes (e.g. the ISODATA clustering algorithm). Label spectral classes into information classes (i.e.Land cover classes).
Perform accuracy assessment
Lecture 6
Image Classification
8/7/2019 Lecture 6R - App. RS
2/22
15/2/0515/2/05 App. RSApp. RS 22
Classification involving algorithms that examine the
unknown pixels in an image; Pixels then aggregated into a number of classes based on
the natural groupings or clusters present in the imagevalues;
Values within a given cover type should be close togetherin the measurement space, whereas data in different classesshould be comparatively well separated - basic premise ;
The resulting classes are spectral classes; Initial identification of spectral classes are not known
because they are based sole on natural grouping;
8/7/2019 Lecture 6R - App. RS
3/22
15/2/0515/2/05 App. RSApp. RS 33
Analysis must compare classified data with somereference data e.g. large scale images or maps;
In supervisedclassification useful informationcategories are first defined then examination oftheir separability is done;
In unsupervisedclassification separability is firstdetermined then definition of their informationalutility;
8/7/2019 Lecture 6R - App. RS
4/22
15/2/0515/2/05 App. RSApp. RS 44
Approaches to Unsupervised Classification:
Classification Algorithm: Distance measurement:
Classification of an entire image must considerthousands of pixels;
To find grouping of pixels, the distance between
pairs of pixels are calculated; The simplest method is Euclidean distance based
on Pythagorean theoremc = a2 + b2
Finding the distance a, b, and c which are measured
in units of the two spectral channels;
8/7/2019 Lecture 6R - App. RS
5/22
15/2/0515/2/05 App. RSApp. RS 55
In relation to other distancesthis form a means ofdefining similarities betweenpixels;
For example, if the distanceab = 40.5 and ac = 86.3 thewe know A is closer andtherefore similar to B than itis to C;
The group should be formed
between A and B rather thanA and C
The measure can be applied to
as many dimensions (spectral
channels as might be available,by addition of distances e.g.
above
8/7/2019 Lecture 6R - App. RS
6/22
15/2/0515/2/05 App. RSApp. RS 66
Thousands of distance
calculations made as a
means of determining
similarities; Proceeds in an interactive
fashion to search for an
optimal allocation of pixel
categories, given the
constraints specified bythe analysis;
8/7/2019 Lecture 6R - App. RS
7/22
15/2/0515/2/05 App. RSApp. RS 77
8/7/2019 Lecture 6R - App. RS
8/22
15/2/0515/2/05 App. RSApp. RS 88
K-means Approach:
The number of clusters are set by the analyst location
of that number of cluster centers in multidimensionalspace is then calculated; Pixel are assigned to the cluster whose arbitrary mean
vector is closest; After assigning pixels to a cluster the mean vectors for
each clusters are revised and computed; Revised means used as a basis to reclassify the image
minimization of within group variation and maximizingbetween group variation;
This procedure is continued until there is no significant
changes in the location of class mean vectors betweensuccessive iterations as the classes meet all theconstraints required by the operator;
8/7/2019 Lecture 6R - App. RS
9/22
15/2/0515/2/05 App. RSApp. RS 99
At this point the analyst determines the land coverfor each spectral class;
Often applied to sub-areas rather than full scenes;
Training statistics developed for combiningclasses are used to classify the entire scene (e.g.minimum distance or maximum likehood
algorithm) called hybrid classification usingboth supervised and unsupervised techniques;
Hybrid classification helps when there arecomplex variability in spectral response due to
variation in species (i.e. cover type) and differentsite condition (soils, slope, aspect);
8/7/2019 Lecture 6R - App. RS
10/22
15/2/0515/2/05 App. RSApp. RS 1010
Accuracy Assessment:
A classification is not finished before the accuracy ofthe map has been assessed;
Visually
A simple visual inspection does the map look
like the real world;Confusion matrix (Error Matrix)
A table that shows how individual classes isclassified in relation to a set of reference data
(test data);
8/7/2019 Lecture 6R - App. RS
11/22
15/2/0515/2/05 App. RSApp. RS 1111
Sources ofClassification Error:
Classification error is the assignment of a pixelbelonging to one category to another category duringthe classification process
The simplest causes of error are related tomisassignment of information categories to spectral
categories; Mixed pixels occur as resolution elements fall on the
boundaries between landscape parcels;
Compiling the Error Matrix:
Comparison of two images the reference image andthe image to be evaluated on a point-by-point basis;
8/7/2019 Lecture 6R - App. RS
12/22
15/2/0515/2/05 App. RSApp. RS 1212
Determination of how each site on the reference imageis represented in the classification;
Comparison between image and a map depicting fieldverified data;
A network of uniform cells will form the units ofcomparison, small enough to provide enough cell for
statistically valid samples and avoiding mixed cells; Images are superimposed for compilation manually or
digitally using a computer;
If differences are due to one being more detailed, the
more detailed map can be collapsed into more generalclasses
8/7/2019 Lecture 6R - App. RS
13/22
15/2/0515/2/05 App. RSApp. RS 1313
8/7/2019 Lecture 6R - App. RS
14/22
15/2/0515/2/05 App. RSApp. RS 1414
Overall Accuracy:
The simple expression is:
The number of correctly classified pixels inrelation to the total number of pixels;
Should be in the order of at least 70-80%
A good classification has an overall accuracygreater than 85%
8/7/2019 Lecture 6R - App. RS
15/22
15/2/0515/2/05 App. RSApp. RS 1515
Producers Accuracy:
The Producer is you;
Interested in how well a certain class is classified;
Therefore the producer wants to know how manyreference pixels for a certain class that has beencorrectly classified;
8/7/2019 Lecture 6R - App. RS
16/22
15/2/0515/2/05 App. RSApp. RS 1616
Users Accuracy:
The user is the final reader of the map;
Interested in how well a certain class on the map
actually represent that class on the ground; Therefore the user wants to know how many
reference pixels for a certain class has beencorrectly classified;
- a guide to the reliability of the map as a predictivetool;
8/7/2019 Lecture 6R - App. RS
17/22
15/2/0515/2/05 App. RSApp. RS 1717
The Confusion Matrix
8/7/2019 Lecture 6R - App. RS
18/22
15/2/0515/2/05 App. RSApp. RS 1818
Advantages ofUnsupervisedClassification:
No extensive prior knowledge of the region is
required: Opportunity for human error is minimized:
Unique classes are recognized as distinct units:
Disadvantages andLimitations:
May identify spectrally homogeneous classes in thedata that does not necessarily correspond to theinformation of interest to the analysis;
Limited control over the menu of classes and theirspecific identity;
Spectral properties of specific informational classeswill change over time;
8/7/2019 Lecture 6R - App. RS
19/22
15/2/0515/2/05 App. RSApp. RS 1919
FieldData:
Defining the relationship between image data and
corresponding points on the ground; Acquiring data suitable for specific tasks and
establishing with confidence the relationship betweenthe image and condition on the ground;
Kinds ofFieldData:
Field data serves one of three purposes;i. To verify, evaluate or assess the results of RS
investigation;ii. Provision of reliable data to guide the
analytical process e.g. creating training fieldto support supervised classification;
8/7/2019 Lecture 6R - App. RS
20/22
15/2/0515/2/05 App. RSApp. RS 2020
iii. Provision of information used to modelspectral behavior of specific landscape
features e.g. plants, soil, water bodies; Field data must contain at least three kinds of
information;i. attributes or measurements that describe
ground conditions at a specific placee.g. identification of specific crop or landuse;
ii. Observations must be linked to locationand size e.g. slope, aspect, and
elevation;- matching attributes to correspondingpoints in the image;
8/7/2019 Lecture 6R - App. RS
21/22
15/2/0515/2/05 App. RSApp. RS 2121
iii. Observation must be describe withrespect to time and date;
Complete field data also includes other data e.g.weather, illumination, calibration information for
instruments etc. The purpose of field data is to facilitate
reconstruction in as much detail as possible, ofground and atmospheric conditions at the time andplace the image was taken;
8/7/2019 Lecture 6R - App. RS
22/22
15/2/0515/2/05 App. RSApp. RS 2222
Biophysicaldata:
Measurement of physical characteristics collected inthe field e.g. type, size, form, spacing of plants etc.;
Specific to the purpose of the study typicallyinclude characteristics such as, leaf area index (LAI),biomass, soil texture, and soil moisture;
Measurements vary over time therefore carefulrecords of time, date, location, and weathercondition is necessary;