Case Study on Multispectral Land Cover Classification - Open Source Geospatial Tools: Applications in Earth Observation - page 300

-i LC82070232013_features.tif
Name of the input raster dataset.
-s training.sqlite
Name of the sample vector dataset we want to extract.
-r median
Select the median value of all pixels in input raster dataset that covered by a
polygon in the sample.
-polygon
Create the output vector dataset with polygon features.
-bndnodata 0 -srcnodata 0
Ignore pixels with values that have a value 0 in band 0 from the input raster
dataset.
-t 50
Extract only 50 % of the features.
-f SQLite
Output vector type is SQLite.
-o training_features.sqlite
Name of the output vector dataset.

We set the option -srcnodata 0 to make sure pixels with a low vegetation
index are not selected. The vegetation mask is the first band of our input image, so we
set -bndnodata to 0 (see Sect. 12.5.3 for more on no-data values in pksvm ). The
resulting Spatialite vector dataset can then directly be used to train our supervised
classifier implemented in pksvm .

17.3 Image Classification

17.3.1 Unsupervised Classification

We will first perform an unsupervised classification on the Landsat 8 composite as
a preliminary way of dividing the image into clusters. This clustering can assist
with image interpretation. The tool otbcli_KMeansClassification from
the Orfeo Toolbox provides an effective way of carrying out an unsupervised classifi-
cation. Using the same input feature image fromSect. 17.1.2 , we run the unsupervised
classification with 15 clusters. The result is shown in Fig. 17.4 .

otbcli_KMeansClassification -in LC82070232013_features.tif -nc
ₒ 15 -maxit 1500 -ct 0.0001 -out

ₒ LC82070232013_unsupervised.tif