Case Study on Multispectral Land Cover Classification - Open Source Geospatial Tools: Applications in Earth Observation - page 299

Training
forest
no-forest

0 25 50 km

Fig. 17.3 Forest/non-forest training vector

17.2.3 Add Band Attributes

We must prepare the training sample for the classification utility implemented in
pktools ( pksvm ) by adding information from the satellite image bands. Using
pkextract we will add the attributes for the band information to the Spatialite
vector file that we created above. We will calculate the median value in each band
and each selected polygon. Other choices are valid too, but a nice feature of the
median is that it is less affected by outlier pixels than, for instance, the average.

pkextract -i LC82070232013_features.tif -s training.sqlite -r
ₒ

median -polygon -bndnodata 0 -srcnodata 0 -t 50 -f SQLite

ₒ

-o training_features.sqlite