Probability map
Probability maps are rasters containing each band as a class. Each band represents the probability of each pixel belonging to a given class in a [0; 1000] range. Band names are sorted, for example:
band name |
1 |
12 |
13 |
1597 |
2 |
25 |
236 |
3 |
4 |
band number |
1 |
2 |
3 |
4 |
5 |
6 |
7 |
8 |
9 |
Parameters involved
Parameter Key |
Parameter section |
Parameter Type |
Default value |
Parameter purpose |
|---|---|---|---|---|
enable_probability_map |
arg_classification |
Boolean |
False |
enable the probability map generation |
Parameters compatibility
The probability maps can only be generated if the shark classifier is used during the run.
Warning
if enable_probability_map is True then classifier must be 'sharkrf'
Additionnal outputs
Each classifications will generate it own probability map called PROBAMAP_TT_model_MM_seed_SS.tif
in classif output iota2 directory where TT is the tile’s name MM the model’s name
and SS the seed number. Once all generated, they are merged under the
ProbabilityMap_seed_SS.tif name into the final directory.
Internal choices
In some case, there is many classifier decisions to a given pixel. This section detail internal choices in order to provide a probability map without NoData labels.
Post-classification fusion
By enabling the dempster_shafer_SAR_Opt_fusion parameter flag, iota2 will
class each pixels invoking the model built thanks to SAR data and the one built
by using optical data. Then, the dempster-shafer is used to attribute the final decision.
Here is the rules to attribute the vector of probability to each pixels
- Consider :
\(p\): the pixel of interest
\(ProbaMapSAR\): the probability map provided by the SAR model
\(ProbaMapOpt\): the probability map provided by the optical model
\(ProbaMapOut\): the output probability map
\(DS\): the dempster-shafer choice
\(DS(p)\) in \({0, 1, 2, 3}\)
0 : no decision
1 : choice is both
2 : choice is SAR
3 : choice is Optical
- Then :
\(ProbaMapOut(p) = 0\) if \(DS(p) = 0\)
\(ProbaMapOut(p) = ProbaMapSAR(p)\) if \(DS(p) = 2\)
\(ProbaMapOut(p) = ProbaMapOpt(p)\) if \(DS(p) = 3\)
\(ProbaMapOut(p) = ProbaMapOpt(p)\) if \(DS(p) = 1\) and \(max(ProbaMapOpt(p)) > max(ProbaMapSAR(p))\)
\(ProbaMapOut(p) = ProbaMapSAR(p)\) if \(DS(p) = 1\) and \(max(ProbaMapSAR(p)) > max(ProbaMapOpt(p))\)
Too huge regions
Using both parameters classifMode to 'separate' and mode_outside_RegionSplit to an integer value,
many models will be built to a given region. Then each models will votes to each pixels inside the region and
some fusion rules are involved. Here is the probability maps fusion rule :
the probability of a given class is the mean of all probability provided by each classifier.
Developers corner
Some unit tests are involved by probability maps generation.
Tests
unittests/iota2_tests_classifications.py:Iota2TestClassifications:test_reorder_proba_map unittests/iota2_tests_optical_sar_fusion.py:Iota2TestOpticalSARFusion:test_compute_probamap_fusion