Case Study on Multispectral Land Cover Classification - Open Source Geospatial Tools: Applications in Earth Observation - page 309

We can check if both layers have indeed been created using the utility ogrinfo .

ogrinfo -so reference.sqlite

INFO: Open of 'reference.sqlite'
using driver 'SQLite' successful.
1: valid
2: border

With the reference sample being labeled, we can use the utility pkdiff (see
Sect. 12.6 ) to assess the accuracy of the forest maps (both unsupervised and super-
vised). First, we concentrate on the valid layer (i.e. no border pixels) only.

pkdiff -ln valid -ref reference.sqlite -i
ₒ

LC82070232013_fmap_unsupervised_masked.tif -cm -nodata 0 -lr label

ₒ -lc class -o reference_fmap_unsupervised.sqlite

processing layer valid
0...10...20...30...40...50...60...70...80...90...100 - done.
12
1 8 6
2 14 842

class #samples userAcc prodAcc

1 54 67 (12.8) 52 (6.9)

2 856 97 (3.3) 98 (0.4)

Kappa: 0.56
Overall Accuracy: 96 (0.7)

-ln valid
Name of the layer to select.
-ref reference.sqlite
Name of the labeled reference vector dataset.
-i LC82070232013_fmap_unsupervised_masked.tif
Name of the input raster dataset.
-cm
Report with a confusion matrix.
-nodata 0
Do not take 0 values in input raster dataset into account.
-lr label
Name of the attribute for reference label in reference vector dataset.
-lc class
Name of the new attribute for classified label (map) in output vector dataset.
-o reference_fmap_unsupervised.sqlite
Name of output vector dataset.

Though the overall accuracy is 96%, there is need for improvement. This becomes
clear when we look at the confusion matrix in more detail. It shows the reference