EX2 - Image Classification

Question 1

What is image classification?

Answer 1

Digital image classification is the process of assigning pixels to classes.

Pixels are compared to one another and to pixels of known identity. Then, similar pixels are grouped into classes of interest.

These classes form regions on a map or image.

Pixels within classes are spectrally more similar to one another than they are to pixels in other classes

Theoretically, each class is homogeneous
Practically, each class has some diversity

Question 2

Spectral /point classifiers

Answer 2

Classifier refers loosely to a computer program that implements a specific procedure for image classification

consider each pixel individually, assigning it to a class based on its several values measured in separate spectral bands

Pros: simplicity
Cons: doesn’t exploit the relationship contained in relationships between each pixel and those that neighbor it.

Human interpreters, for example, derive little information using this point by point approach; humans derive information from context and patterns of
brightness of groups of pixels

Question 3

Supervised Classification

4 advantages and 5 disadvantages

Answer 3

Advantages:

1) Analyst controls selection of information classes.
2) Classification tied to areas of known training areas
3) No need to match spectral classes to information
4) Easier to detect serous errors in training data

Disadvantages:
1) May imposes structure on data that may not match natural classes in data

2) Training areas are defined primarily with respect to informational category and secondarily with respect to spectral properties (100% forest?)
3) If the area to be classified is large and complex, training data may not be representative of conditions encountered throughout the image
4) Selection of training data can be time consuming, tedious and expensive
5) Supervised classification may not be able to recognize and represent special or unique categories not represented in the training data (because they are unknown to the analyst or they occupy very small areas on the image)

Question 4

Unsupervised Classification

5 advantages and 3 disadvantages

Answer 4

Advantages

1) No extensive prior knowledge required
2) Objective;minimum human influenceor bias or error
3) Unique classes (big and small)
4) Works fast
5) Works the same way always

Dis-advatages:

1) Identifies spectrally homogeneous classes that may not correspond to the informational categories of interest to the analyst
2) Analyst has little control
3) Spectral properties of specific informational classes will changeover time (seasonally and over years). Relationships defined for one image cannot extend to others

Question 5

Informational Classes and Spectral subclasses

Answer 5

FOREST > shadowed > pine > etc etc

Categories of interest to users, Object of analysis
Classes that we wish to derive from the data. We derive these using BV

Each information class is composed of numerous spectral 
subclasses.

In classification, we treat spectral subclasses as distinct 
units during classification but then display several spectral 
classes under a single informational class for the final 
image

Question 6

Unsupervised classification

Basic strategy

Answer 6

Euclidean Distance Measure

K-means algorithm

Question 7

Unsupervised classification

Euclidean Distance Measure

Answer 7

• differences between pixel a and pixel b in each band
• square the differences
• Total the differences squared
• get the square root of the total
• You have euclidean distance measure

if pixels AB is less than the same operation for AC, we know the pixel A is closer to B than to C.

Question 8

Unsupervised classification

K-means algorithm (5 Steps)

Answer 8

• choose the number of clusters
• randomly assigns initial positions on cluster centroids
• Assigns points to nearest centroids
• Recompute for closer centroids
• if solution coverges then stop!

Key components of the algorithm:

1. Effective methods of measuring distances in data space
2. Identifying class centroids
3. Testing the distinctness of classes

Objective, but nor completely objective because the analyst decides:

1. The data to be examined
2. The algorithm to be used
3. The # of classes to be found
4. (Possibly) the uniformity and distinctness of classes

Question 9

Assignment of spectral classes to informational classes ( 2 serious practical problems with unsupervised classification)

Answer 9

 Some informational categories may not have direct spectral counterparts, and vice versa. clear matches are not always possible.

 Analyst cannot control the nature of categories generated (hence, comparison between places and over time are hard to make). The same set of informational
categories are not always generated on both images

Question 10

Supervised Classification

3 Basic Strategies

Answer 10

1) Analyst uses prior knowledge to guide the classification.
2) Analyst identifies “training areas” to represent the typical spectral classes that make up the informational classes. Training areas are digitized polygons
3) The classification algorithm then classifies each pixel in the rest of the image based on comparisons to training data

Question 11

Supervised Classification

Training fields

Answer 11

Training fields are vector polygons that are digitized over pixels of known identity.

pixels of unknown identity are identified or assigned to a particular informational class by comparing their spectral signature to the spectral signature of the pixels within the training field

Question 12

Supervised Classification

6 Key characteristics of training fields

Answer 12

 # of pixels: at least 100 pixels for each
category

 Size: Large enough to provide accurate estimates of spectral characteristics of each category; but not large enough to include spectral in homogeneity

 Location: Each informational category must be represented by several training areas positioned throughout the image (even distribution)

 Number: Better to define many small training areas than few large ones.Ideally, 5-10 training areas for each
category (to ensure representation of spectral subclasses)

 Placement: Training field boundaries must be placed well away from the edges of contrasting parcels so that they do not encompass edge pixels (avoid mixels!)

 Uniformity: Data within each training
area should exhibit a unimodal frequency
distribution for each spectral band to be
used.

Question 13

Parametric Classification

Answer 13

Decision rules are based entirely on the statistics

(min, max, mean, # of pixels in training field, # of bands in input image) produced from the training field.

Question 14

non-parametric classification

Answer 14

Decision rules are spatial; based on location

of pixel in multidimensional spectral space

Question 15

Common classifiers for supervised classification

a.Parallelpiped classification (STD DEV)

Answer 15

Non parametric

Classifies solely on the ranges (or on standard deviations) of spectral values in the training data to define regions within multidimensional space

pixels that match the training data range (1 or 2 standard deviations from the mean) are assigned to the appropriate categories

can be extended to as many bands or as many categories as needed

Question 16

Common classifiers for supervised classification

b.Minimum Distance Classification (Centroid -> eucd dist)

Answer 16

Spectral data from training fields are plotted in multidimensional space

1) Similar to unsupervised classification: but, pixels in training field defined by the analyst
2) Each training area cluster has its own centroid (defined as its mean value)
3) unassigned pixels are assigned to the closest centroid based on minimum “Euclidean Distance” measures

Advantages
 Computational efficiency (slowerthan parallelepiped; faster than maximum likelihood)
 Good to use when number of pixels in each training field is small and the training field are not very
well defined
 All pixels are classified, so no holes in the image
 No overlapping of classes

Disadvantages
 Not widely used
 Not very accurate
 Does not take into account the variability within classes
 Assumes that spectral variability is the same in all directions, which is not the same

Question 17

Common classifiers for supervised classification

c.ISODATA Classification (kmeans algo)

Answer 17

An improvement on the Minimum Distance Classification

A hybrid classification : shares characteristics of both supervised and unsupervised classification methods

1. Training data is delineated and represented as clusters in multidimensional place (S)
2. All unassigned pixels are then assigned to the nearest centroid (S)
3. Centroids are then recomputed to include effects of those pixels that have been assigned in step 2
   (U)
4. If any centroid changes in value from Step 2 to Step

3, then the process returns to step 2 and
repeats the assignment of pixels to the closest centroid. The process is repeated until there is no
change or very little change in class centroids from one iteration to the next (U)

Question 18

Common classifiers for supervised classification

d.Maximum Likelihood Classification

Answer 18

Parametric

training data is a way of estimating means and variances of the classes, which are then used to estimate the probability of a pixel belonging to a particular class

Parallelepiped classifier, overlap is a serious problem
because spectral data space cannot then be neatly divided into discrete units

Maximum Likelihood Classifier, instead, makes good use of this overlap

Question 19

Common classifiers for supervised classification

e.Fuzzy Clustering

Answer 19

Fuzzy classifier permits partial membership

SCALE:
0 [for non membership]
1 [full membership]).

Other classifiers, all pixels (including mixels and edge pixels) are assigned to a single discrete class and many pixels are incorrectly or illogically labeled

 Fuzzy classifier can assign a pixel a membership grade of 0.3 for “water” and 0.7 for “forest”

 Fuzzy clustering is a soft classifier 
each pixel has a ‘degree of membership” in each class as opposed to complete allocation in

Question 20

Hard Classifiers vs. Soft Classifiers

Answer 20

hard classifiers: classifiers where each pixel is assigned to only one of a number of classes

soft classifier: classifiers where each pixel has a ‘degree of membership” in each class as opposed to complete allocation in one class

Question 21

Spatial /neighborhood classifiers

Answer 21

classifiers consider groups of pixels within their spatial setting within the image as a means of using the textural information (which humans understand better).

These classifiers examine small areas within the image
using both spectral and textural information to classify the image

Pros: In some cases, this classifier has demonstrated improved accuracy

Cons: complex; difficult to program; not routine

Question 22

supervised classification: popular methods

Answer 22

parallel piped
max-like
minimum distance
iso DATA
fuzzy clustering