Pktools - Open Source Geospatial Tools: Applications in Earth Observation - page 187

-c|--class name and -r|--reclass number . The number of pairs -c
name -r number should exactly match the number of distinct labels found in the
training vector file. For the urban land use mapping of the IEEE data fusion contest
2013 (see Appendix A.2.3), we can classify the hyperspectral image based on pktools
with the following command:

pksvm -t training_features_names.sqlite -i
→ 2013_IEEE_GRSS_DF_Contest_FEATURES.tif -o

→ 2013_IEEE_GRSS_DF_Contest_map.tif -c grass_healthy -r 1

→ -c grass_stressed -r 2 -c grass_synthetic -r 3 -c tree -r

→ 4 -c soil -r 5 -c water -r 6 -c residential -r 7 -c

→ commercial -r 8 -c road -r 9 -c highway -r 10 -c railway

→ -r 11 -c parking_lot1 -r 12 -c parking_lot2 -r 13 -c

→ tennis_court -r 14 -c running_track -r 15

The name-value pairs can also be used to reclass multiple class labels to the same
class number in case you want to merge classes. The following command creates
a binary forest (1) non-forest (2) map, based on a training set with three different
forest types and three non-forest land cover types:

pksvm -t training_types.sqlite -i image.tif -o
→ forest-noforest.tif -c broadleaved -r 1 -c coniferous -r

→ 1 -c mixed -r 1 -c grass -r 2 -c urban -r 2 -c water -r 2

→

-ct ct.txt

12.5.3 No-Data Values

Pixel in the input raster dataset (or features in case of an input vector dataset) with
all their bands equal to 0 are ignored by the classification algorithm and obtain a
no-data value in the output. As a default, the output no-data value is set to 0, but
you can change it with the option -nodata . In addition, you can set a mask image
with no-data values using the options -m|--mask and -msknodata . Pixels in the
input image matching for which the corresponding mask image has a no-data value
will not be classified. Instead, for these pixels, the mask will be copied to the ouput
map.

12.5.4 Optimizing the SVM Parameters

The support vector machine depends on several parameters. Ideally, these parameters
should be optimized for each classification problem. In case of a radial basis kernel
function, two important parameters are cost and gamma . The utility pkoptsvm can
optimize these two parameters, based on an accuracy assessment (the Kappa value).