The use of Orfeo Toolbox in the context of map updating
- 1. The use of Orfeo Toolbox in the Context of Map Updating Christophe Simler Royal Military Academy Brussels, Belgium
- 2. Main part of the ARMURS project VHR satellite image (Ikonos or raster of an old vector Quickbird) or aerial image RGB geographical database pansharpening XS pansharpened multispectral pixel description and mean shift segmentation segmented image region feature extraction and SVM classification classified image (road/building/other) comparaison change map (roads and buildings) database update
- 3. Softwares Freeware -ORFEO Toolbox Extensible Handle most image format (use GDAL) -Development Image processing for remote sensing -Proprietary code -Open source code
- 4. Main part of the ARMURS project VHR satellite image (Ikonos or raster of an old vector Quickbird) or aerial image RGB geographical database pansharpening XS pansharpened multispectral pixel description and mean shift segmentation segmented image region feature extraction and SVM classification classified image (road/building/other) comparaison change map (roads and buildings) database update
- 5. Main part of the ARMURS project VHR satellite image (Ikonos or raster of an old vector Quickbird) or aerial image RGB geographical database pansharpening: otb::SimpleRcsPanSharpeningFusionImageFilter XS pansharpened multispectral pixel description and mean shift segmentation: otb::MeanShiftVectorImageFilter segmented image region feature extraction and SVM classification: otb::SVMModel and otb::SVMClassifier classified image (road/building/other) comparaison change map (roads and buildings) database update
- 6. Mean shift segmentation results Part of an Ikonos satellite image in the region of Jodoigne (Belgium) Roads and buildings are generally precisely extracted
- 7. Régions feature extraction The regions obtained from the segmentation are described by the following feature vector: area eccentricity mean R mean G mean B mean NIR
- 8. Image classification The feature vectors are classified into classes « roads », « building » or « other » Support Vector Machine (SVM)
- 9. Training set Our training set is composed of about 1000 mean shift regions manually assigned to class « road », « building » or « other » two componants of our feature vector:
- 10. Training two 2-class SVM with Gaussian kernel are trained independently road/other building/other Parameters to tune: - kernel standard deviation - penalisation of the misclassifications
- 11. Training INPUT : training set INPUT : 2D grid value for (road/other or building/other) the 2 parameters to tune training set test set new couple of permutation parameter values learning decision boundaries FN FP balanced loss= = + VP + FN VN + FP (loop) optimal parameter values 1- optimisation of the two parameters by cross validation learning 2- learning on the whole set OUTPUT : decision boundaries 3- classifier performance quantification
- 12. Optimal tuning Energie to minimise
- 13. Optimal tuning Minimisation with coarse-to-fine approach
- 14. Optimal tuning Minimisation with coarse-to-fine approach
- 15. Classification: SVM input image Part of an aerial RGB image of a region of Bruxelles (Belgium)
- 16. SVM Classification results Overlap of the two 2-class SVM classification results roads buildings other both building and road (the existence of such conflict areas is due to the fact the two 2-class SVM are trained separately)
- 17. Conclusion The ORFEO ToolBox has been considered as a basic component in our application of map updating within the ARMURS project. The provided image segmentation and classification functions speeded up the implementation and test of the approach. As far as the demonstrator is concerned, the integrated file formats for image access and vector read are important assets. We are also currently considering the potential of the recent OTB application Urban Area Extraction (from OTB 3.0) as a component on which to base building and road extraction.