Case Study on Multispectral Land Cover Classification - Open Source Geospatial Tools: Applications in Earth Observation - page 303

-m LC82070232013_features.tif
Mask overlaid on the input image. In this case the input image is also used as a
mask (only first band is read for the mask).

-msknodata 0
Ignore pixel values in the input image for which the mask has a value 0.
-nodata 0
Ignored pixels will have a nodata (value 0) in the output map.
-o LC82070232013_fmap_supervised.tif
Name of the output raster dataset (classified map).

17.3.3 Post-processing

We will now introduce the concept of a minimum mapping unit (MMU) by applying
an extra filter on the pixel based classification result. A sieving filter can merge small
objects in the land cover map to the majority class of their neighbors (see Sect. 9.1 ) .
As an example, we will define a minimum mapping unit of 5 ha by setting the size
threshold to 55 pixels (one pixel represents an area of 900 m 2 ). We do this for both
the supervised and unsupervised forest maps.

gdal_sieve.py LC82070232013_fmap_unsupervised.tif LC82070232013_fmap_
ₒ unsupervised_sieved.tif -st 55
gdal_sieve.py LC82070232013_fmap_supervised.tif LC82070232013_fmap_
ₒ

supervised_sieved.tif -st 55

LC82070232013_fmap_unsupervised.tif
Name of the input map.
LC82070232013_fmap_unsupervised_sieved.tif
Name of the sieved output map.
-st 55
Size threshold: objects smaller than 55 pixels will be merged into largest neighbor
object.

Pixel values within the acquisition area where the NDVI was below the vegeta-
tion threshold were masked from the classification for the supervised classifier and
obtained a value 0. For instance, inland water bodies typically have very low NDVI
values. We want these values to be labeled as non-forest in the final forest map.
Also for the unsupervised classification result, we want to make sure these values
are labeled as non-forest.