4.3.4 Dempster Shafer based Classifier Fusion

This framework is dedicated to perform cartographic validation starting from the result of a detection (for example a road extraction), enhance the results fiability by using a classifier fusion algorithm. Using a set of descriptor, the processing chain validates or invalidates the input geometrical features.

Fuzzy Model (requisite)

The DSFuzzyModelEstimation application performs the fuzzy model estimation (once by use case: descriptor set / Belief support / Plausibility support). It has the following input parameters :

The application can be used like this:

otbcli_DSFuzzyModelEstimation -psin     PosSamples.shp  
                              -nsin     NegSamples.shp  
                              -belsup   "ROADSA"  
                              -plasup   "NONDVI" "ROADSA" "NOBUIL"  
                              -desclist "NONDVI" "ROADSA" "NOBUIL"  
                              -out      FuzzyModel.xml

The output file FuzzyModel.xml contains the optimal model to perform informations fusion.

First Step: Compute Descriptors

The first step in the classifier fusion based validation is to compute, for each studied polyline, the choosen descriptors. In this context, the ComputePolylineFeatureFromImage application can be used for a large range of descriptors. It has the following inputs :

The output is a vector data containing polylines with a new field containing the descriptor value. In order to add the ”NONDVI” descriptor to an input vector data (”inVD.shp”) corresponding to the percentage of pixels along a polyline that verifies the formula ”NDVI >0.4” :

otbcli_ComputePolylineFeatureFromImage -in   NDVI.TIF  
                                       -vd  inVD.shp  
                                       -expr  "b1 > 0.4"  
                                       -field "NONDVI"  
                                       -out   VD_NONDVI.shp

NDVI.TIF is the NDVI mono band image of the studied scene. This step must be repeated for each choosen descriptor:

otbcli_ComputePolylineFeatureFromImage -in   roadSpectralAngle.TIF  
                                       -vd  VD_NONDVI.shp  
                                       -expr  "b1 > 0.24"  
                                       -field "ROADSA"  
                                       -out   VD_NONDVI_ROADSA.shp

otbcli_ComputePolylineFeatureFromImage -in   Buildings.TIF  
                                       -vd  VD_NONDVI_ROADSA.shp  
                                       -expr  "b1 == 0"  
                                       -field "NOBUILDING"  
                                       -out   VD_NONDVI_ROADSA_NOBUIL.shp

Both NDVI.TIF and roadSpectralAngle.TIF can be produced using Monteverdi feature extraction capabilities, and Buildings.TIF can be generated using Monteverdi rasterization module. From now on, VD_NONDVI_ROADSA_NOBUIL.shp contains three descriptor fields. It will be used in the following part.

Second Step: Feature Validation

The final application (VectorDataDSValidation) will validate or unvalidate the studied samples using the Dempster-Shafer theory . Its inputs are :

The output is a vector data containing only the validated samples.

otbcli_VectorDataDSValidation -in      extractedRoads_enriched.shp  
                              -descmod FuzzyModel.xml  
                              -out     validatedSamples.shp