i.svm.train - GRASS GIS manual

NAME

i.svm.train - Train a SVM
Train a Support Vector Machine

KEYWORDS

SYNOPSIS

i.svm.train

i.svm.train --help

i.svm.train [-sp] group=name [subgroup=name] trainingmap=name signaturefile=name [type=name] [kernel=name] [cache=cache size] [degree=value] [gamma=value] [coef0=value] [eps=value] [cost=value] [nu=value] [p=value] [--overwrite] [--help] [--verbose] [--quiet] [--ui]

Flags:

-s: Do not use the shrinking heuristics; Defaults to use the shrinking heuristics
-p: Train a SVC or SVR model for probability estimates; Defaults to no probabilities in model
--overwrite: Allow output files to overwrite existing files
--help: Print usage summary
--verbose: Verbose module output
--quiet: Quiet module output
--ui: Force launching GUI dialog

Parameters:

group=name [required]: Maps with feature values (attributes)
subgroup=name: Name of input imagery subgroup
trainingmap=name [required]: Map with training labels or target values
signaturefile=name [required]: Name for output file containing result signatures
type=name: Type of SVM; Options: c_svc, nu_svc, one_class, epsilon_svr, nu_svr; Default: c_svc; c_svc: C-SVM classification; nu_svc: nu-SVM classification; one_class: one-class SVM; epsilon_svr: epsilon-SVM regression; nu_svr: nu-SVM regression
kernel=name: SVM kernel type; Options: linear, poly, rbf, sigmoid; Default: rbf; linear: u'*v; poly: (gamma*u'*v + coef0)^degree; rbf: exp(-gamma*|u-v|^2); sigmoid: tanh(gamma*u'*v + coef0)
cache=cache size: LIBSVM kernel cache size in MB; Options: 1-; Default: 512
degree=value: Degree in kernel function; Options: 0-; Default: 3
gamma=value: Gamma in kernel function; Default: 1
coef0=value: coef0 in kernel function; Default: 0
eps=value: Tolerance of termination criterion; Defaults to 0.00001 for nu-SVC and 0.001 for others
cost=value: Cost of constraints violation; The parameter C of C-SVC, epsilon-SVR, and nu-SVR; Default: 1
nu=value: The parameter nu of nu-SVC, one-class SVM, and nu-SVR; Default: 0.5
p=value: The epsilon in epsilon-insensitive loss function of epsilon-SVM regression; Default: 0.1

DESCRIPTION
NOTES
PERFORMANCE
EXAMPLE
SEE ALSO
REFERENCES
AUTHORS

DESCRIPTION

i.svm.train finds parameters for a Support Vector Machine and stores them in a signature file for later usage by i.svm.predict.

Internally the module performs input value rescaling of each of imagery group rasters by mean normalisation based on minimum and maximum value present in the raster metadata. Rescaling parameters are written into the signature file for use during prediction.

NOTES

i.svm.train internally is using the LIBSVM. For introduction into value prediction or estimation with LIBSVM, see a Practical Guide to Support Vector Classification by Chih-Wei Hsu, Chih-Chung Chang, and Chih-Jen Lin.

It is strongly suggested to have semantic labels set for each raster map in the training data (feature value) imagery group. Use r.support to set semantic labels.

PERFORMANCE

SVM training is done by loading all training data into memory. In a case of large input raster files, use sparse label rasters (e.g. raster points or small patches instead of uninterrupted cover).

During the training process there is no progress output printed. Training with large number of data points can take significant time - just be patient.

By default the shrinking heuristics option of LIBSVM is enabled. It should not impact the outcome, just the training time. On some input parameter and data combinations training with the shrinking heuristics disabled might be faster.

The cache parameter determines the maximum memory allocated for kernel caching to enhance computational speed. It's important to note that the actual module's memory consumption may vary from this setting, as it solely impacts LIBSVM's internal caching. The cache is utilized on an as-needed basis, so it's unlikely to reach the specified value.

EXAMPLE

This is the first part of classification process. See i.svm.predict for the second part.

Train a SVM to identify land use classes according to the 1996 land use map landuse96_28m and then classify a LANDSAT scene from October of 2002. Example requires the nc_spm_08 dataset.

# Align computation region to the scene
g.region raster=lsat7_2002_10 -p

# store VIZ, NIR, MIR into group/subgroup
i.group group=lsat7_2002 subgroup=res_30m \
    input=lsat7_2002_10,lsat7_2002_20,lsat7_2002_30,lsat7_2002_40,lsat7_2002_50,lsat7_2002_70

# Now digitize training areas "training" with the digitizer
# and convert to raster model with v.to.rast
v.to.rast input=training output=training use=cat label_column=label
# If you are just playing around and do not care about the accuracy of outcome,
# just use one of existing maps instead e.g.
# r.random input=landuse96_28m npoints=10000 raster=training -s

# Train the SVM
i.svm.train group=lsat7_2002 subgroup=res_30m \
    trainingmap=training signaturefile=landuse96_rnd_points

# Go to i.svm.predict for the next step.

REFERENCES

Please cite both - LIBSVM and i.svm.

For i.svm.* modules:
Nartiss, M., & Melniks, R. (2023). Improving pixel-based classification of GRASS GIS with support vector machine. Transactions in GIS, 00, 1–16. https://doi.org/10.1111/tgis.13102
For LIBSVM:
Chang, C.-C., & Lin, C.-J. (2011). LIBSVM : a library for support vector machines. ACM Transactions on Intelligent Systems and Technology, 2:27:1--27:27.

AUTHORS

Maris Nartiss, University of Latvia.

SOURCE CODE

Available at: i.svm.train source code (history)

Latest change: Tuesday Apr 23 10:45:15 2024 in commit: f8115df1219e784a7136e7609f4c9bb16d928e2f