Remote Sensing

3/23/2020

1

Classification of Remotely Sensed Data

General Classification Concepts Unsupervised Classifications

What is Image Classification?

• Process of converting image pixels or regions to classes that represent self-similar features or “themes”

• Using images to create “thematic maps”

How?

• Two general approaches: – Manual interpretation (e.g., photointerpretation,

“heads-up digitizing”) – Digital classification (per pixel)

• Many digital techniques developed – Unsupervised classification – Supervised classification – Classification and Regression Trees (CART) – Neural Networks – Etc., etc., etc.

3/23/2020

2

General Classification Steps

Field reconnaissance Develop a classification scheme (legend) Enhance imagery (as needed) Use classification algorithms Incorporate ancillary data (as needed) Check accuracy of product Refine iteratively

Field Reconnaissance

• Critical for understanding the distribution of your theme in the real world

• Helps you choose useful ancillary data • Useful for understanding satellite imagery

back at the office • Nice to get out once in a while

3/23/2020

3

What characteristics of this landscape might be important for making a map using satellite data?

Developing Classification Schemes (Legends)

• How many types do you want to map? • How should you divide up the feature you are

interested in? • Can be very controversial!

Classification Schemes (List of types to map)

• What thematic classes are you going to assign pixels to?

1) Must be useful 2) Must be detectable using the data you have 3) Should be hierarchical 4) Categories must be mutually exclusive 5) Require explicit definitions of each class

3/23/2020

4

Classification Scheme -- Example

I. Vegetated A. Forest

1. Evergreen a. Spruce-fir forest

i. Spruce-fir with winterberry understory

b. Lodgepole pine forest c. etc.

2. Deciduous

B. Shrubland

II. Non-Vegetated

Basic Classification Steps

1) Field reconnaissance 2) Development of classification scheme 3) Image enhancements (Veg indices, etc.) 4) Run classification algorithm 5) Incorporate ancillary data 6) Check accuracy of product 7) Refine iteratively

Classification Algorithms

• Procedures for grouping pixels and/or areas into the classes from your classification scheme

3/23/2020

5

Manual processing

• Aerial photos, print-outs of images, photos

– Transparent overlays – Delineate features (interpretation) – Compile maps – Generate reports

Digital processing

• Satellite images, digital or scanned photo • Digital on-screen interpretation (“Heads-up

digitizing”)

– Display geo-referenced image/photo on-screen – Digital line-drawing (with mouse or digital pen) to

delineate features • Analogous to drawing on an overlay

– Processed lines are converted to features in a Geographic Information System (GIS)

– Generate maps and reports

Detailed view of Wyoming GAP Land Cover Map

3/23/2020

6

Digital classification

• Conversion of pixel values into thematic classes – Statistical clustering of the data (lumping

spectrally similar pixels into the same class) – Spectral vs. informational classes – Sometimes combine spectral classes together

to make informational classes – Converting digital satellite data into meaningful

maps—the heart of remote sensing!

Use many bands at once to create a map of classes

Classification

General Types of Classifications

• Unsupervised – computer clusters pixels together based only on the similarity of their DNs.

• Supervised – computer uses training data— examples of target classes—and assigns pixels to the training class that they are most similar to.

• Others (neural networks, fuzzy logic etc).

3/23/2020

7

Classification Analogy

– Truck-load of fruits (pixels): Apples, oranges, kiwis, nectarines, bananas, pineapples, tangerines, plums, peaches, lemons (hundreds of each)

– Goal: separate them by type and put them in separate baskets (classes)

– Using a person (= computer) who has never seen these fruits before (!) or doesn’t know the difference between them

Unsupervised Classification

• Software Identifies natural groups (spectral classes) within multi-spectral data based on limited input from the analyst

• Pixel values are grouped based similarity of their DNs, radiance or reflectance – Pixels => Clusters

• Analyst has to match each cluster to a thematic class

Unsupervised Classification 2-bands

. .

.

.

.

.

. .

.. .

.. .

.

.

. . ..

.

.

. ..

. ..

... .

... .

.

.

.

.. . .

. .

.

.

.

.

.

. .

.. .

.. .

.

.

. . ..

.

.

. ..

. ..

... .

... .

.

.

.

.. . .

.

.

.

.

.

.

. .

.. .

.. .

.

.

. . ..

.

.

. ..

. ..

... .

.. .

.

.

.

Band X0 Max

M a x

Class

B an

d Y

3/23/2020

8

Unsupervised Classification 3-bands

1 pixel

1 Class

Unsupervised Classification

Unsupervised Classification

• Choose bands, indices, enhancements, etc. that highlight differences in your classes

• Decide how many classes to separate • Choose a grouping algorithm

– Simple clustering, K-means, etc.

• Classify the image • Group and evaluate the results

3/23/2020

9

Advantages of Unsupervised Classifications

• No extensive prior knowledge of map area required (but you have to label the classes!)

• Repeatable – objective classes are based only on spectral information

• Unique spectral classes recognized as units

Disadvantages of Unsupervised Classifications

• Spectral classes do not always correspond to informational classes

• Limited control over the output classes you end up with

• Spectral properties of informational classes change over time so you can’t always use same class statistics when moving from one image to another

Grouping Algorithms

• Statistical routines for grouping similar pixels together

• Differ in how they: – Determine what is similar (distance measures) – Determine the statistical center (centroid) of a

class – Test the distinctivness of classes

3/23/2020

10

Common Algorithms

• ISODATA • K-means Clustering

ISODATA

• ISODATA = Iterative Self-Organizing Data Analysis – Choose how many classes you want – The algorithm chooses class centers (centroid) by

spreading them evenly through the “data cloud” – Groups each pixel with nearest centroid – Calculates the centroid of the new cluster – Regroups each pixel with nearest new centroid – Keeps doing this until centroids don’t move much

K-means Clustering

• Like ISODATA but starts by picking centroids as far apart as possible from one another in the data cloud

• Iteratively groups pixels with the centroid and then re-calculates centroid

• Iterates until centroid stops moving

3/23/2020

11

K-means Clustering

Seed the Clusters

Assign Pixels

Move Centroids

Recalculate Cluster

Members

Unsupervised Classification – Summary

• Classification is the statistical clustering of pixels into groups

• Clusters => Thematic classes • Results should be checked and the

classification revised if necessary

3/23/2020

12

Task 1 - Unsupervised Classification

3/23/2020

13

37

Image Classification

Thematic information can be mapped and further analyzed.

• Satellite images “clustered” based on spectral similarities.

• These “clusters” are then assigned into a “theme” or class

38

Image Classification

Derives thematic information from spectral information • Reduces data volume

• Permits analysis of features

Class 1 Vegetation

Class 3 Water

Class 2 Urban

Water

Vegetation

Urban

39

Image Classification: K Means

This process organizes groups (clusters) of pixels with similar spectral responses

Spectral clusters (like land covers) are identified

K Means requires minimal input • # of desired classes

• Iterations

• Convergence threshold

3/23/2020

14

40

Image Classification: K Means

2. Minimum Distance calculations: Each pixel is associated with closest mean

Band 1

..... . . .

..

.. .

.

.

.

.

...

. .... .

. .

. .

.. . ..

.

. . ..

. . . . .

.. .. . . .

.

. . . . ...

..

.

. . . .

.

.

.

.. . .

. . ...

.

. . . .

. .

.

. .

.

Cluster Means 1. Means are initialized along diagonal

3. New mean calculated for each cluster and means migrate to new locations

2

4. Iterations continue until convergence or maximum iterations is reached

5. Each cluster associated with a value. Each pixel given this value

Task 2 - Renaming Classes

42

Labelling Classes

The process of identifying land cover classes and naming them

Label Water Forest Grass Agriculture Urban

ISODATA Class 1 Class 2 Class 3 Class 4 Class 5

Class Names

3/23/2020

15

43

Labelling Classes

Raster attribute editor Color assignments

Naming Classes

Making educated guesses

Forest

Water

Grass

Wheat

Forest1 (Decid)

Forest2 (Conif)

Forest3 (Mix)

Water1 (Deep)

Water2 (Shallow)

Task 3 - Recoding Classes

45

Recode or Merging Classes

This allows for the combining similar classes. For example:

Water 1(Silted) and

water 2 (clear)

may be RECODED to a new class of WATER

3/23/2020

16

46

Recode (Merge) Classes

Greater class delineation by selecting more classes than ultimately desired

The likelihood of mixed classes is reduced

Water

Land

Water Vegetation1

Vegetation2

Vegetation3

Vegetation4

Vegetation5