SCENE CLASS RECOGNITION USING HIGH RESOLUTION SAR/INSAR SPECTRAL DECOMPOSITION METHODS Anca Popescu, Inge Gavat University Politehnica Bucharest (UPB) Mihai Datcu German Aerospace Center (DLR) IGARSS July 2011, Vancouver, Canada University POLITEHNICA Bucharest Faculty of Electronics, Telecommunications and Information Technology

Motivation: High resolution scene category indexing, large number of structures visible in urban sites Non – parametric feature extraction methods

Summary Introduction Data descriptors Spectral features Spectral components Experimental data Classification results Conclusions

Introduction Study of value adding processing methods for SLC and InSAR data, for scene class recognition

Introduction Study of value adding processing methods for SLC and InSAR data, for scene class recognition Concept: Make use of the information contained in the phase of the SAR signal

Data descriptors – spectral features Direct estimation from spectra based on spectral differences

Data model The 2-D signal model: Where complex amplitude of the k th sinusoid unknown frequencies of the k th sinusoid 2-D noise Data descriptors spectral components estimation Problem: Estimate the parameters of the sinusoidal signals:

J. Li, P. Stoica: “Efficient Mixed-Spectrum Estimation with Applications to Target Feature Extraction”, IEEE Transactions on Signal Processing, 44, 1996, Model choice: minimize NLS criterion: Peak of the 2-D periodogram Algorithm preparations: if α k and f k are known, minimize cost function for the k th sinusoid is: Height of the peak (complex) Data descriptors spectral components estimation

Data descriptors spectral components estimation

AKAIKE Information Criterion – model order selection Estimates the expected Kullback-Leibler information between the model generating the data and a candidate model Model selection: = parameter to be estimated from empirical data y Best Model: Minimum AIC value y = generated from f(x), X is a random variable Log likelihood function for model selection: Data descriptors spectral components estimation

Goal: asses parameter’s capability to discriminate scene classes Evaluation: Accuracy = (TP+TN) / (TP+TN+FP+FN) Methodology for scene class indexing

Test Site – Bucharest, Romania TerraSAR-X High Resolution Spotlight: LAN 130

Frequency of classes in database: dominant classes: tall blocks, green areas, urban fabric, commercial and industrial sites Experimental data – TSX LAN-130 Test Site – Bucharest, Romania

1650 SLC patches, 200 x 200 m Water course/ water body Stadion Very tall buildingTall Block Industrial Site Test Site – Patch database formation Interferometric data

Spectral Centroid, Flux, and Rolloff (azimuth and range) Mean and variance of most significant 3 cepstral coefficients Results – Spectral and cepstral parameters

Relax parameter α k (modulus representation), selection of six random components from the estimated stack Results – Spectral components Dense urban area, mostly tall blocks Green area, small vegetation

Reconstructed data from estimated spectral components and original SAR patch Results – Spectral Components

Results – Scene Class Recognition SLC/InSAR True Negative Rate indicator 23 scene classes discoverable with SLC database 15 scene classes discoverable with InSAR database

Results – Scene Class Recognition SLC – Spectral Features and Spectral Components Accuracy indicator Class Spectral components Spectral features

Influence of interferogram spectral features on classfication accuracy Results – Scene Class Recognition SLC / InSAR InSAR SLC

Evaluation of the capability of spectral parameters to discriminate scene classes for complex HR TerraSAR-X data Spectral estimation method for high resolution data characterization and reconstruction, based on RELAX algorithm. Evaluation of the capability of spectral components to discriminate scene classes for complex HR TerraSAR-X data Accuracy of recognition better than 80% for main class training Acknowledgements to DLR team for providing the interferometric data: Nico Adam, Christian Minet, Nestor Yague-Martinez, Helko Breit. Conclusions

Thank you for your attention!