SAR and Optical post-classification fusion
Purpose of the approach
The main goal of this feature is to allow iota2 to perform a post-classification fusion of SAR and optical data.
This feature was proposed in issue #67 open on September 5, 2018.
An unique field has been added to the configuration file to enable the feature : dempster_shafer_sar_opt_fusion
- Default value
dempster_shafer_SAR_Opt_fusion : True
different from 'None' and an
optical sensor has to be set.
If this conditions are not met, a exception is thrown and the execution of iota2
Step created in order to compute the confusion matrix using a set of validation samples, to evaluate SAR classifications and optical ones. These confusion matrices are computed by tile.
Fusion of confusion matrices by tile in order to obtain a confusion by model
Fusion of classifications comming from SAR and optical models.
Some vector data are created in order to compute confusion matrices and being able to chose the label coming from SAR or optical models.
TTTT : tile’s name
RRRR : region’s name
SSSS : seed number
This functionality requires the production of two classifications by region, one by model.
/classif/TTTT_model_RRRR_confidence_seed_SSSS_DS.tifto classification map
/classif/TTTT_model_RRRR_confidence_seed_SSSS_DS.tifto confidence map
Thanks to the fusion of classification results, we can produce a map which allows users to know which label has been chosen by the fusion of classifications.
It contains 4 possible values resumed in the following table :
SAR + optical
The fusion of classifications is perfomed using the OTB’s Application
The Dempster-Shafer method is the one chosen to decide which label will be the
one in the final classification.
The FusionOfClassification OTB application does not provide a confidence map. The confidence map corresponding to the fusion of classfications is generated thanks to the map of choices with the following rules:
- SAR label has been chosen :
SAR confidence is used
- Optical label has been chosen :
Optical confidence is used
- SAR and optical models voted for the same label :
the maximum confidence is used
The unittest script called
iota2_tests_optical_sar_fusion.py has been created to test this