High Resolution Complex SAR Image Analysis Using Azimuth Sub-Band & Eigenspace Decompositions

CNES/DLR Workshop on Information Extraction and Scene Understanding for Meter Resolution Images 29/03/07 Oberpfaffenhofen, Germany High Resolution Complex SAR Image Analysis Using Azimuth Sub-Band & Eigenspace Decompositions Houda Chaabouni-Chouayakh(1) and Mihai Datcu(1,2) 1 German Aerospace Center (DLR), Oberpfaffenhofen, Germany 2Ecole Nationale Supérieure des Télécommunications (ENST), Paris, France

Motivation With the increase of SAR sensor resolution, we should expect a more detailed analysis and a finer description of the scene. BUT, In high resolution SAR images, we are confronted by: • High diversity of real man-made targets (buildings, parking, roads, vehicles, vegetation,…) • Very complicated structures (various designs and shapes) • Different behaviors to SAR (according to the sensor angle…) • Complex images (amplitude + phase) In the literature, among the image understanding methods, there are: • Azimuth sub-band decomposition (mainly for relevant scatterers detection in high resolution SAR images) • Eigenspace decomposition (mostly for real valued images for face classification and complex-valued images for very specific military targets classification) ≠ Low diversity of military targets High diversity of urban targets

Outline • Overview on SAR • Description of the method • Azimuth sub-band decomposition • Eigenspace formalism • Results & Discussion • Conclusions & Perspectives

SAR Data Acquisition and Image Formation • Transmitting pulsed signals (chirp) in the range direction, when the platform travels in the azimuth direction • Coherently adding the successively reflected and received pulses The platform is moving during the illumination time Doppler Effect

Doppler Centroid = Doppler shift of a target positioned in the antenna boresight direction Zero Doppler Non-Zero Doppler

Description of the method Azimuth Sub-band Decomposition = a decomposition of the complex spectrum in the azimuth direction = a selection of an azimuth sub-aperture corresponds to a selection of some viewing angles or sensor positions Could be useful to: • study the variation of the signal from one band to another in order to get a finer description of the targets according to different sensor angles • analyze the behavior of the strong scatterers that may exist in the scene

Description of the method Azimuth Sub-band Decomposition Evidence of some details which were not in the original image. Loss or fading of some structures In the sub-band images. Low directivity of the corner reflectors. Original image 1 2 3 Sub-band left Sub-band right

Description of the method Azimuth Sub-band Decomposition Steps: • Doppler centroid estimation and compensation of Doppler shift • Unweighting in azimuth • Spectrum division into 2 sub-bands • Centering the obtained sub-images • Zero-padding and hamming weighting of each sub-band

X= Classification Description of the method Eigenspace Decomposition or Covariance Formalism Training Data Covariance Matrix Σ=XX*T Vectors Eigenspace Decomposition Principal Components Analysis (PCA) C3 C1 C2 Projection & Comparison

X= Averaging Description of the method Eigenspace Decomposition or Covariance Formalism Training Data Eigenspace Vectors C3 C1 C2 Projection WCk Average Projection Vector of the training data of Ck Ck Projection (Ck) Training data of the classCk

X= Classification Description of the method Eigenspace Decomposition or Covariance Formalism Training Data Covariance Matrix Σ=XX*T Vectors Eigenspace Decomposition Principal Components Analysis (PCA) C3 C1 C2 Projection & Comparison Test Data Vector XT= Projection Vector of XTIn the Eigenspace Average Projection Vector of the training data of CkIn the Eigenspace

Classification Description of the method Eigenspace Decomposition or Covariance Formalism Training Data Covariance Matrix Vectors Eigenspace Decomposition Principal Components Analysis (PCA) Projection & Comparison Test Data Vector How to use the rich information provided by the azimuth sub-band decomposition in the covariance formalism to better classify the strong scatterers of the high resolution SAR images?

Matrix to Vector Converter 2-Azimuth Sub-band Decomposition Matrix to Vector Converter Description of the method Azimuth Sub-band Decomposition & Eigenspace Decomposition Original Image Sub-band 1 X1 = X= & Eigenspace Decomposition X2= Sub-band 2 CLASSIFICATION

Results & Discussion V W Description of the database • SAR image over the city of Dresden in Germany, acquired with the Experimental SAR system (E-SAR) of the German Aerospace Center (DLR) • 5 classes (50 images from each class): • Big Buildings (BB) • Average Buildings (AB) • Small Buildings (SB) • Vegetation (V) • Water (W) BB AB SB

Results & Discussion How to test the performance of our new classifier? Percentage of Good Classification PGC • RCk: number of the correctly Recognized test images of the Class k • TCk: Total number of the test images of the Class k • 1,2,…,5correspond respectively to BB, AB, SB, V and W Image size n n 20<n<60 n

Results & Discussion Big Building Classification • The covariance with azimuth sub-band decomposition algorithm outperforms the general covariance algorithm for almost all the image sizes. • In fact, the azimuth decomposition provides finer characterization of the strong scatterers with which the big buildings are mainly constructed (corner reflectors, antennas on the roofs,…).

Results & Discussion Average Building Classification • By using only the covariance formalism, the algorithm is not able to recognize the average buildings (less than 30% of good classification in most of the cases). • With the azimuth decomposition, the recognition becomes much better, specially for the large image sizes (more than 40% of well-classified average buildings when n>28).

Results & Discussion Small Building Classification • A good classification requires a large image size. The surrounding area in this case seems to react as a relevant characteristic. Indeed, an image of small buildings, includes, in general, different small sub-classes (vegetation, cars, roads, lights,...) which have several backscattering behaviors, and thus requires a sufficiently large number of pixels for a good description.

Results & Discussion Vegetation Classification • The azimuth decomposition improves advantageously the classification. • The image size is determining (more than 40x40 pixels are needed for a recognition over 70%). • The fact that the vegetation could sometimes be considered as a sub-class for the buildings, results in a kind of confusion between classes for the low image sizes.

Results & Discussion Water Classification • The azimuth decomposition has no effective amelioration on water classification. • However, combing it with the covariance formalism provides more flexibility to find the optimal image size.

Results & Discussion What is the optimal image size? • PGCk:Percentage of Good Classification of the class k. • 1,2,…,5correspond respectively to BB, AB, SB, V and W.

Results & Discussion Classification results in terms of confusion matrix when n=nopt=48 p a • p: predicted class • a: actual class

Results & Discussion Classification results in terms of confusion matrix when n=nopt=58 p a • p: predicted class • a: actual class

Training Data no Test Data Eigenspace & Azimuth Decomposition Best Classification? yes STOP Results & Discussion: Training Data Selection Classification Quality Test TD: Training data Classification Quality function

Training Data no Test Data Eigenspace & Azimuth Decomposition Best Classification? yes STOP Results & Discussion: Training Data Selection Steps of the training data selection algorithm Step 0: initialization of the training data: TD=TD0 Step 1: classification of the whole database using the covariance with azimuth sub-band decomposition algorithm. Step 2: computation of the classification quality function C(TD) and evaluation of the stopping condition of the algorithm: C(TD)>70% If the stopping condition is not yet met, a new training data TD* is chosen by changing randomly the training images of the class k* which has the lowest percentage of good classification. The algorithm is then run again from Step 1.

Training Data no Test Data Eigenspace & Azimuth Decomposition Best Classification? yes STOP Results & Discussion: Training Data Selection p p a a Confusion matrix when n=nopt=58

Conclusions &Perspectives • A preliminary classification of high resolution SAR images has been performed on a five-class database (BB; AB, SB, V and W). • The proposed method aims at combining the rich information provided by the azimuth decomposition with the promising properties of the covariance formalism, to get a superior classification performance. • To evaluate the performance of our classifier, a study on the optimal image size was carried out. • It was demonstrated that: • The covariance with azimuth decomposition algorithm outperforms the general covariance algorithm for all the classes for almost all the image sizes. • The image size is an important and determining parameter for the classification. • The azimuth decomposition provides more flexibility in the choice of the optimal image size. • As perspectives, we can propose: • To test the performance of the proposed method for unsupervised classifications. • To compare the method to some non-linear classification methods (ICA,…)

CNES/DLR Workshop on Information Extraction and Scene Understanding for Meter Resolution Images 29/03/07 Oberpfaffenhofen, Germany High Resolution Complex SAR Image Analysis Using Azimuth Sub-Band & Eigenspace Decompositions Houda Chaabouni-Chouayakh(1) and Mihai Datcu(1,2) 1 German Aerospace Center (DLR), Oberpfaffenhofen, Germany 2Ecole Nationale Supérieure des Télécommunications (ENST), Paris, France

High Resolution Complex SAR Image Analysis Using Azimuth Sub-Band & Eigenspace Decompositions