Image classification

From GRASS-Wiki
Revision as of 16:47, 2 May 2010 by ⚠️MilenaN (talk | contribs) (extended)
Jump to navigation Jump to search

Image classification

Classification methods in GRASS

radiometric
unsupervised
radiometric
supervised 1
radiometric
supervised 2
radiometric & geometric
supervised
Preprocessing i.cluster i.class (monitor digitizing) i.gensig (using training maps) i.gensigset (using training maps)
Computation i.maxlik i.maxlik i.maxlik i.smap
Remarks automated run based requires digitalization requires digitalization requires digitalization
on image statistics of training areas of training areas of training areas

You can digitize training areas with either r.digit (not recommended) or v.digit + v.to.rast (recommended)

Interactive setup

  • i.class - Generates spectral signatures for an image by allowing the user to outline regions of interest.
The resulting signature file can be used as input for i.maxlik or as a seed signature file for i.cluster.

Processing

  • i.cluster - Generates spectral signatures for land cover types in an image using a clustering algorithm.
The resulting signature file is used as input for i.maxlik, to generate an unsupervised image classification.

Unsupervised classification

  • i.maxlik - Classifies the cell spectral reflectances in imagery data.
Classification is based on the spectral signature information generated by either i.cluster, i.class, or i.gensig.

Supervised classification

  • i.smap - Performs contextual (image segmentation) image classification using sequential maximum a posteriori (SMAP) estimation.

Small classification tutorial

Required: GRASS >= 6.4.0

We are using Landsat satellite data prepared for the North Carolina sampling site. It is a multispectral dataset where each channel is stored in a separate map.

We are using the North Carolina location here.

Warming up with Landsat TM

 # look what maps are there
 g.list rast
 # set to one of the color channels and print current region settings
 g.region rast=lsat7_2002_10 -p
 # look at some Landsat map (1: blue, 2: green, 3: red, 4: near infrared etc.)
 d.mon x0
 d.rast lsat7_2002_10
 d.rast lsat7_2002_20
 d.rast lsat7_2002_30 
 d.rast lsat7_2002_40

We can easily query the spectrum for individual points:

 d.what.rast lsat7_2002_10,lsat7_2002_20,lsat7_2002_30,lsat7_2002_40

Histogram of a channel:

 d.histogram lsat7_2002_10

Create RGB view on the fly, true and false color:

 d.rgb b=lsat7_2002_10 g=lsat7_2002_20 r=lsat7_2002_30
 d.rgb b=lsat7_2002_70 g=lsat7_2002_20 r=lsat7_2002_30
 d.rgb b=lsat7_2002_70 g=lsat7_2002_50 r=lsat7_2002_30
 d.rgb b=lsat7_2002_70 g=lsat7_2002_50 r=lsat7_2002_10

Check thermal channel (encoded to fit into 0.255 data range):

  1. http://landsathandbook.gsfc.nasa.gov/handbook/handbook_htmls/chapter11/chapter11.html
  2. Look up LMIN/LMAX: "Table 11.2 ETM+ Spectral Radiance Range watts/(meter squared * ster * µm)"
  3. we won't go into details here, r.mapcalc would be used to convert to Kelvoin/degree Celsius.
 # look at encoded thermal channel
 d.rast lsat7_2002_61

NDVI as example for map algebra:

 r.mapcalc "ndvi = 1.0 * (lsat7_2002_40 - lsat7_2002_30)/(lsat7_2002_40 + lsat7_2002_30)"
 d.rast.leg ndvi
 r.colors ndvi color=ndvi
 d.rast.leg ndvi

Show selected NDVI "classes":

 d.rast ndvi val=0.3-1.0
 d.rast ndvi val=-1.0--0.1
 d.rast ndvi val=0.0-0.2
 d.rast ndvi val=0.0-0.3

Look at panchromatic channel, compare resolution:

 g.region rast=lsat7_2002_80 -p
 d.erase -f
 d.rast lsat7_2002_80
 d.zoom
 d.rast lsat7_2002_30
 d.rast lsat7_2002_80

Image Classification

Set region setting to red image:

 g.region rast=lsat7_2002_30 -p

Create a group

 i.group group=lsat7_2002 subgroup=lsat7_2002 \
   input=lsat7_2002_10,lsat7_2002_20,lsat7_2002_30,lsat7_2002_40,lsat7_2002_50,lsat7_2002_70

Unsupervised classification

Generate unsupervised statistics

 i.cluster  group=lsat7_2002 subgroup=lsat7_2002 sigfile=lsat7_2002_sig classes=10 reportfile=lsat7_2002.txt
 # look at report file
 gedit lsat7_2002.txt 

Assign pixels to classes, check quality of assignment:

 i.maxlik group=lsat7_2002 subgroup=lsat7_2002 sigfile=lsat7_2002_sig class=lsat7_2002_class reject=lsat7_2002_reject
 d.rast.leg lsat7_2002_class
 d.mon x1
 d.rast.leg lsat7_2002_reject
 d.mon x2
 d.histogram lsat7_2002_reject

Play with different cluster separations:

 i.cluster group=lsat7_2002 subgroup=lsat7_2002 sigfile=lsat7_2002_sig classes=10 reportfile=lsat7_2002.txt separation=1.5
 i.maxlik group=lsat7_2002 subgroup=lsat7_2002 sigfile=lsat7_2002_sig class=lsat7_2002_class2 reject=lsat7_2002_reject2
 d.mon x4
 d.rast.leg lsat7_2002_class2
 d.rast.leg lsat7_2002_reject2
 d.rast.leg lsat7_2002_reject
 d.rast.leg lsat7_2002_reject2
 d.rast.leg lsat7_2002_reject
 d.rast.leg lsat7_2002_reject2
 d.rast.leg lsat7_2002_class2

Compare to RGB:

 d.rgb b=lsat7_2002_70 g=lsat7_2002_50 r=lsat7_2002_10
 d.rast lsat7_2002_class2 cat=1 -o
 d.rgb b=lsat7_2002_70 g=lsat7_2002_50 r=lsat7_2002_10
 d.rast lsat7_2002_class cat=1 -o
 d.rgb b=lsat7_2002_70 g=lsat7_2002_50 r=lsat7_2002_10
 d.rast lsat7_2002_class cat=1 -o

We do a 3rd attempt with more classes

 i.cluster group=lsat7_2002 subgroup=lsat7_2002 sigfile=lsat7_2002_sig classes=15 reportfile=lsat7_2002.txt 
 i.maxlik group=lsat7_2002 subgroup=lsat7_2002 sigfile=lsat7_2002_sig class=lsat7_2002_class3 reject=lsat7_2002_reject3
 d.rgb b=lsat7_2002_70 g=lsat7_2002_50 r=lsat7_2002_10
 d.rast lsat7_2002_class3 cat=1 -o


Supervised classification

Rlaunch GUI for digitizer:

 g.gui wxpython

 # ... digitize training areas of water, asphalt, forest, uncovered soil; polygons/centroids with attribute
 # advantage: we know what the classes are

Assign also numerical ID for each class in new column

 v.db.select lsat7_training
 v.db.addcol lsat7_training col="id integer"
 v.db.select lsat7_training
 v.db.update lsat7_training col=id where="name = 'water'" val=1
 v.db.update lsat7_training col=id where="name = 'forest'" val=2
 v.db.update lsat7_training col=id where="name = 'asphalt'" val=3
 v.db.update lsat7_training col=id where="name = 'soil'" val=4
 v.db.select lsat7_training

Transform vector training map to raster model:

 v.to.rast in=lsat7_training out=lsat7_training use=attr col=id labelcol=name --o
 
 d.mon x0
 d.rast.leg lsat7_training 

Generate statistics from training areas

 i.gensigset group=lsat7_2002 subgroup=lsat7_2002 sig=lsat7_2002_smap training=lsat7_training

Perform supervised classification

 i.smap group=lsat7_2002 subgroup=lsat7_2002 sig=lsat7_2002_smap  out=smap
 d.rast.leg smap
 d.rgb b=lsat7_2002_10 g=lsat7_2002_20 r=lsat7_2002_30

Vectorize result:

 r.to.vect smap out=smap feat=area
 d.rgb b=lsat7_2002_10 g=lsat7_2002_20 r=lsat7_2002_30
 d.vect smap  type=boundary col=red

That's it :)


Further reading classification with GRASS