Image classification: Difference between revisions

Latest revision as of 10:48, 4 December 2018

Image classification

Classification methods in GRASS

	radiometric unsupervised	radiometric supervised 1	radiometric supervised 2	radiometric & geometric supervised
Image Preprocessing	r.smooth.seg (optional, in case of noisy original data)
Segmentation	i.segment
Preprocessing	i.cluster	i.class (monitor digitizing), g.gui.iclass (in GRASS 70)	i.gensig (using training maps)	i.gensigset (using training maps)
Classification	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~~ GRASS Digitizing tool + v.to.rast (recommended)

Image Preprocessing (optional)

r.smooth.seg - produces a smooth approximation of the data and performs discontinuity detection. The module is based on the Mumford-Shah variational model for image segmentation. Used as pre-analysis for a subsequent classification.

For details, see the r.smooth.seg manual (formerly called: r.seg)

i.segment - identifies segments (objects) from imagery data based (currently) on a region growing and merging algorithm. The image segmentation results can be useful (on their own or) as a preprocessing step for image classification, i.e. reduce noise and speed up the classification.
- Classification of these segments: see below in "Unsupervised classification".

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.

g.gui.iclass - Tool for supervised classification of imagery data.

Preprocessing

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.

i.gensig - Generates statistics for i.maxlik from raster map layer.
i.gensigset - Generate statistics for i.smap from raster map layer.

Background information concerning i.cluster:

Discussion: Is "i.cluster" an implementation of the ISODATA algorithm?
About Clustering

Unsupervised classification

For an introduction, see for example Cluster_analysis.

i.maxlik - Classifies the cell spectral reflectances in imagery data.

Classification is based on the spectral signature information generated by i.cluster (see above)

GRASS GIS 7: i.segment
- Classification of these segments could be achieved with these addons: v.class.mlR, v.class.mlpy, v.class.ml.

See example below.

Supervised classification

i.maxlik - Classifies the cell spectral reflectances in imagery data.

Classification is based on the spectral signature information generated by either i.class, or i.gensig.

g.gui.iclass - Tool for supervised classification of imagery data.
i.smap - Performs contextual (image segmentation) image classification using sequential maximum a posteriori (SMAP) estimation.
r.learn.ml - Supervised Machine Learning classification and regression of GRASS rasters using the python scikit-learn package
In case of panchromatic maps or limited amount of channels, it is often recommended to generate synthetic channels through texture analysis (r.texture)

See example below.

Object-based classification and image segmentation

Created based on Summary of object-based classification possibilities in GRASS GIS (June 2014). Feel free to merge this with other content or with different page.

Related commands:

i.segment - Identifies segments (objects) from imagery data.
i.segment.hierarchical performs a hierarchical segmentation on an image (group). The module uses internally i.segment
i.segment.stats calculates statistics describing raster area.
extra: r.smooth.seg - Generates a piece-wise smooth approximation of the input raster and a discontinuity map.
extra: i.superpixels.slic - Perform image segmentation using the SLIC segmentation method.

Background information:

Publications on GRASS GIS and Object Based Image Analysis (OBIA) techniques (via Google Scholar)
Old references
Another discussion about what variables to use: Object-based classification (nabble link)

Source code and scripts:

The source code of the classification tools (for inspiration in writing your own scripts):
- v.class.ml: http://svn.osgeo.org/grass/grass-addons/grass7/vector/v.class.ml/ (Pietro's tool)
- v.class.mlpy: http://svn.osgeo.org/grass/grass-addons/grass7/vector/v.class.mlpy/ (Vaclav's tool)
- i.mlpy.py: http://svn.osgeo.org/grass/sandbox/turek/i.mlpy.py (Stepan's tool)
List of Python machine learning libraries:
- scikit-learn, http://scikit-learn.org/ (used by v.class.ml)
- mlpy, http://mlpy.sourceforge.net/ (used by v.class.ml, v.class.mlpy and i.mlpy)
- Pandas, http://pandas.pydata.org/
- Milk, http://luispedro.org/software/milk

Small classification tutorial

Required: GRASS >= 6.4

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):

http://landsathandbook.gsfc.nasa.gov/handbook/handbook_htmls/chapter11/chapter11.html
Look up LMIN/LMAX: "Table 11.2 ETM+ Spectral Radiance Range watts/(meter squared * ster * µm)"
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