Upload
lenhan
View
217
Download
0
Embed Size (px)
Citation preview
9/24/2014
1
1
Classification
= Semi-automated grouping of similar pixels using their multispectral DNs
Remote Sensing Classification Automated grouping of similar pixels using multispectral DNs
Developed since 1972 (Landsat 1)
Digital alternative to manual mapping
9/24/2014
2
3
Remote sensing Classification
Semi-automated grouping of similar pixels using their multispectral DNs Slide Number
4
Manual interpretation e.g. air photos
Features are classified = simplified
Human interpretation / classification relies on attributes such as: Shape, pattern, texture, shadows, size, association, tone, colour
9/24/2014
3
5
Using just one band to classify ?
A band (or colour composite) could be treated as an air photo (interpretation) Digital Numbers from one band alone are rarely enough – features are not unique
Band 3 Band 4
6
The role of multispectral sensing in classification use of multiple bands – 2 features similar in one band, differ in another
9/24/2014
4
7
The role of multispectral sensing in classification
DN Band A
DN band B
DNs in Band A are similar for Corn and Wheat DNs in Band B are similar for Corn and Soybeans … but if we use both Bands A and B, then ……
http://fas.org/irp/imint/docs/rst/Sect1/Sect1_16.html
8
Classification – land cover
Image classification uses multispectral digital numbers (‘colour’) Algorithms are ‘per pixel’ classifiers
http://fas.org/irp/imint/docs/rst/Sect1/Sect1_16.html
3 bands … 3D (and we have 7 bands)
9/24/2014
5
9
Band / channel selection
Landsat TM has 7 bands: You would NOT select 3 visible bands to classify The visible bands are similar – and thus the composite is low in contrast
10
sample band correlation coefficients
Example: PG Landsat data (r values between bands)
TM1 2 3 4 5 6 TM1 TM2 .97 TM3 .96 .96 TM4 .07 .16 .11 TM5 .66 .72 .76 .46 TM6 .77 .77 .81 .14 .80 TM7 .83 .86 .90 .25 .93 .86
The Visible bands are highly correlated (similar) .. (r = .96 to .97)
.. so are bands 5 and 7 (r = .93)
band 4 (near-IR) is not very correlated with Visible or MIR (nor thermal)
Note: these values will vary for different environments e.g. urban, desert, forested
9/24/2014
6
11
Classification: Band / Channel Selection
How to choose which ones to use: 1. Low correlation e.g. TM 3-4-5 or 2-4-7 (Visible-NIR-MIR) (example below shows high ‘r’ Red v Green and lower ‘r’ Red v NIR)
2. Past experience, visual examination, logical thinking 3. Channels that separate the features we want to identify (based on DNs / spectral curves)
12
A> Unsupervised classification
Characteristics
-user needs no 'a priori' knowledge of area - software clusters pixels by natural DN groupings (based on similarity and contrast – ‘natural breaks’) --------------------------------------------------------------------------- Steps - determine input bands / channels and - determine how many classes
- run classifier : K-means or Isodata
-assign names to classes (merge classes if needed)
9/24/2014
7
13
Unsupervised result – 10 classes (clusters)
K-Means and Isodata: K-means minimises within cluster range Isodata can also merge / split clusters
Unsupervised – how it works Algorithm starts with statistical seed points
Assigns each pixel to the closest seed
Calculates group mean
Re-assigns pixels to the closest group mean
Re-calculates group mean
Iterates (10?) until relatively little change and fixes groupings
9/24/2014
8
Iterations in unsupervised classification
Final step .. Assigning names to clusters (and merge some)
16
PG classification report 1 iteration
DN values for bands 3,4,5
9/24/2014
9
17
After 16 iterations
Unsupervised classification
Input bands selected – minimum 3 or 4 bands; more acceptable
9/24/2014
10
Merging and adding classes
Merging – clusters are not really separate features Clusters can be merged if they overlap spatially or are not distinguishable spectrally.
Splitting / adding If one cluster covers too much area – run again with more clusters Can generate many clusters, and then group merge later
20
9/24/2014
11
21
22
Iskut 543 – Water /ice
9/24/2014
12
23
K-means
24
Fuzzy K-means
9/24/2014
13
25
Isodata
Mt. Edziza – classification and sieve (‘peneira’)
- recognises connectivity of adjacent pixels in the same class - special classes e.g. wetlands can be specified and preserved - removes small sub-areas; does not ‘blur’ edges like filtering
9/24/2014
14
Sieve - filter
Classification ALWAYS produces a 'salt and pepper' effect with isolated pixels This is a result of a. the local variations in DNs and b. using ‘per-pixel’ classifiers
Mt. Kilimanjaro
28
Vector conversion After classes are finalised, the adjacent class pixels can be transformed into vector polygons, using polygon growing or raster to vector options (see feature extraction and lab 7)
9/24/2014
15
29
Thursday lecture and Lab next week :
supervised classification
Monday 6pm: wee outing
“pequena turnê”
(weather permitting)
30
Challenges in classification There are many causes of spatial variations in reflectance (a range of DNs for a feature) NATURAL RESOURCES (e.g. forest stands) purity of stand, understory, age/maturity, density, health/disease, sun angle, topography
9/24/2014
16
31
There are many causes of spatial variations in reflectance (a range of DNs for a feature) URBAN / HUMAN (e.g. roads or residences) - amount of grass, types of material, roofing colour, weathering, sun angle (building shape)
32
There are many causes of spatial variations in reflectance (a range of DNs for a feature)
All imagery: moisture, edge (mixed) pixels, sun angle, time of year