ship detectors exploiting spectral analysis of sar images - IEEE Xplore
SHIP DETECTORS EXPLOITING SPECTRAL ANALYSIS OF SAR IMAGES Armando Marino1 , Maria J. Sanjuan-Ferrer2 , Irena Hajnsek1,2 and Kazuo Ouchi3 1
ETH Zurich, Institute of Environmental Engineering, Zurich, Switzerland German Aerospace Center (DLR), High Frequency Department, Oberpfaffenhofen, Germany 3 Korean Institute of Ocean Science and Technology, Korea Ocean Satellite Center, Ansan, South Korea 2
I. INTRODUCTION Ship detection is an important topic for security and surveillance of maritime and costal areas. A solution exploiting satellite-borne SAR sensors is particularly interesting, because it offers wide scale surveillance capabilities, which are not reliant on solar illumination and are rather independent of weather conditions ([1], [2], [3]). In SAR images, the main feature of a ship is a relatively large backscattering signal, which is usually brighter in comparison with the sea background. This led to the idea of using the intensity contrast as a feature to discriminate between targets and sea clutter. Several methodologies were proposed ([1], [3]). Most of these techniques set a statistical test between target and clutter background. Recently, the several ship detectors were proposed that exploits the property of SAR images to perform detection. In this work, two methodologies used for coherent scatterer detection are tested for the first time for ship detection and a comparison of ship detectors based on spectral analysis is performed over L-band ALOS date accompanied by a ground survey.
Gaussian process) [5], [4]. Interestingly, the sea present such behavior and therefore a procedure that is able to check for stability of targets over the spectrum could be exploited to understand if a target is present in the scene [6], [7]. In this section the different detectors used in the following comparison are presented. As a general recommendation in the formation of the sub-images is that the spectrum in range and azimuth is weighted by some windowing procedure aimed at reducing side-lobes in the spatial domain. The windowing can be easily removed considering a multiplication for the inverse of the power spectrum. II-A. Sub-look coherence This is the first detector exploiting sub-images to perform ship detection (please note different names were proposed in the papers) [6], [7]. The idea is to split the spectrum two nonoverlapping portions and consider the coherence between the two: γ=p
II. SHIP DETECTION WITH SPECTRAL ANALYSIS This section aims at presenting several algorithms exploiting properties of the spectrum of SAR images to perform ship detection. The idea behind SAR spectral analysis considers the evaluation of the Fourier transform of images (in azimuth and range) and the extraction of portions of the spectrum (following some rational) [4]. The spectrum sections are then anti-transformed (in order to obtain images with lower resolution, here defined as sub-images) and combined together with some procedure. In the following the image spectrum will be indicated with S, while the image in the spatial domain is s. A sub-image will be referred as si (with i and index indicating the image), corresponding to a portion of the spectrum called Si . The idea behind using spectral analysis for detection is that point targets (i.e vessels) are supposed to have a spectral response that is quite stable and constant in the spectral axes and therefore they will show a relatively high coherence. On the other hand a target with presents a completely developed speckle (the single look pixel can be modeled as a complex
978-1-4799-5775-0/14/$31.00 ©2014 IEEE
978
|hs1 · s∗T i| p2 ∗T hs1 · s1 i hs2 · s∗T 2 i
(1)
where ∗ and T stand for conjugate and transpose respectively and hi defines a spatial average considering a boxcar filter on the sub-images. II-B. Sub-look cross-correlation A methodology based on the same principle of the sublook coherence is presented in [8]. There are two main differences with the previous methodology, firstly the cross correlation between sub-images is not normalized: ρ = |hs1 · s∗T 2 i|.
(2)
secondly, the sub-images are obtained by portion on the spectrum that can overlap in the spectral domain. This has the advantage that vessels which do not present a very strong point target behavior will be detected as well, but the sea will not be rejected as strongly as in the case of non-overlapping spectral portions. Nevertheless, it is showed in [8] that the contrast sea-vessel in ρ images can be improved overlapping the portions of spectrum.
IGARSS 2014
II-C. Sub-look entropy The coherence is a useful operator to express the correlations between two images. however, when more images are considered, several coherences can be estimated (for each couple of images). A methodology to express the correlation between a stack of images is considering the covariance matrix of all the images and derive the entropy of the eigenvalues. If the entropy is zero (only one eigenvalue is different from zero) than the rank of the covariance matrix is unitary and all the images represent the same target (i.e. in the pixel there is a target stable over the different spectral portions) [9]. If x represents pixel by pixel a vector of images obtained with spectral portions (i.e. x = [x1 , ..., xn ]T , where n subimages are considered). The covariance matrix is calculated as [X] = x · x∗T . If λi are the n eigenvalues obtained by the diagonalisation of [X], their entropy is calculated as: H=−
n X
pi logn pi , where pi =
i=1
λi . T race([X])
(3)
II-D. GLRT Recently, a methodology to detect coherent targets using a Generalized Likelihood Ratio Test was proposed [10]. The algorithm push further the idea of overlapping spectral portions and set a statistical test based on a statistical model for the sea clutter (supposed with a fully developed speckle) and vessels. The analytical expression of the GLRT is: LG =
|a∗T M −1 x|2 (a∗T M −1 a) (x∗T M −1 x)
(4)
where x represent a vector of images obtained with spectral portions. a is a vector that depends on the location of the scatterer in the resolution cell (in this detections the scatterers are assumed at the center of the resolution cell), M is a matrix considering the correlation for a complex Gaussian sea clutter due to the overlapping of the spectral portions (the methodology to calculate it is presented in [10]). The bilinear form LG is one for perfectly coherent (i.e. spectrally stable) targets. III. ALGORITHM COMPARISON III-A. Presentation of the dataset Six Fine Quad-polarimetric images were acquired during winter 2013/2014 in the North Sea and Kattegat sea. The data were collected under the SOAR project EI-5145. Figure 1 shows the locations of the images as polygons on Google Earth. One aim of the SOAR project EI-5145 is to test ship detectors under severe weather conditions. In order to amplify the effects of strong sea clutter, large incidence angles (higher than 350 ) were avoided. One acquisition (03/12/2013) also includes a very steep incidence angles (e.g. ∼ 21.50 ). In
979
order to increase the probability of observing rough sea conditions the test areas were selected with relatively high latitudes and during the winter season. On the other hand, care was taken to select areas with large traffic of vessels. The scenes in the North Sea are close to the harbours of Rotterdam and Amsterdam, while the ones in the Kattegat Sea are around the Anholt Island, along the shipping route from the North Sea to the Baltic Sea. In total, 69 validated ships were observed with a variety of dimensions (ranging between 30 m to 200 m in length) and typology (e.g. fishing boats, cargos, etc). 69 vessels translates in a quantisation error for the estimation of Pd equal to 0.015. III-B. Comparison The first test considers . Figure 2 presents the ROC curves with a boxcar window of 5x9 pixels for the three polarisation channels HH, V V , HV . The cross-polarisation (HV ) channel is able to provide better ROC curves compared with the co-polarisation ones (HH and V V ). This result was also observed by other researchers [1], [11], [12]. Accordingly to the Bragg model, the sea clutter has a very low (or even null) backscattering in the cross-channel, leading in many circumstances to an increased contrast. Several works deal with exploiting polarimetric information to improve the detection performance [13], [14], [15]. Regarding the comparison of different detectors, the GLRT provides the best detection performance. Please note, in the following the comparisons will be mostly done considering rather high values of Pd (e.g. greater than 0.8), since this are the operation scenarios that are most interesting. Following the GLRT, there are the sub-look coherence and entropy. Regarding the sub-look correlation, two versions were tested in this paper following the observation that a different overlap of sub-portions can provide different results [8]. For this reason, one version (solid black line) considers non overlapping sub-portions and the other (black dashed line) exploits sub-portion with 30% overlap. As explained previously, the algorithms are compared with two detectors based on the intensity considering SLC (without filtering) or multi-look (boxcar filtering) images [1]. Interestingly, the intensity detector on the cross-polarisation show large performance loss only if compared with the GLRT. This is an indicator that, if the cross channel is available, the gain of acquiring SLC instead than detected images would be evident only if a powerful methodology is exploited. This is not true for the co-polarisation channel, where the sub-look process provides benefits for all the sublook detectors at exception than the cross-correlation. ACKNOWLEDGMENTS c MacDonald, DetRADARSAT-2 Data and Products twiler and Associates Ltd. (2013-2014) - All Rights Re-
Fig. 1. Locations (red polygons) of the six RADARSAT-2 Fine Quad-pol acquisitions used in the comparison. Google Earth.
served. RADARSAT is an official trademark of the Canadian Space Agency. IV. REFERENCES [1] D. J. Crisp, “The State-of-the-Art in ship detection in Synthetic Aperture Radar imagery,” Australiane Government Department of Defence, 2004. [2] C. Liu, P. W. Vachon, and G. W. Geling, “Improved ship detection using polarimetric SAR data,,” IGARSS Geoscience and Remote Sensing Symposium, vol. 3, pp. 1800–1803,, 2004. [3] P. W Vachon, “Ship detection in synthetic aperture radar imagery.,” Proceedings OceanSAR, St. John s, NL, Canada, 2006. [4] J.-C. Souyris, C. Henry, and F. Adragna, “On the use of complex sar image spectral analysis for target detection: Assessment of polarimetry,” IEEE Trans. on Geos. & Rem. Sen., vol. 41(12), pp. 2725–2734, 2003. [5] C. Oliver and S. Quegan, Understanding Synthetic Aperture Radar Images., SciTech Publishing, Inc, 2004. [6] A. Arnaud, “Ship detection by SAR interferometry,” IGARSS’99, vol. 5, pp. 2616–2618, 1999. [7] K. Ouchi, S. Tamaki, H. Yaguchi, and M. Iehara, “Ship detection based on coherence images derived from cross correlation of multilook SAR images,” IEEE Geoscience and Remote Sensing Letters, vol. 1(3), 2004. [8] C. Brekke, S. N. Anfinsen, , and Y. Larsen, “Subband extraction strategies in ship detection with the subaperture cross-correlation magnitude,” IEEE Geoscience and Remote Sensing Lettersg, vol. 10(4), pp. 786–790, 2013. [9] R. Z. Schneider, K. P. Papathanassiou, I. Hajnsek, and A. Moreira, “Polarimetric and interferometric characterization of coherent scatterers in urban areas,” IEEE Trans. on Geos. & Rem. Sen., vol. 44(4), pp. 971–984, 2006.
980
[10] M. J. Sanjuan-Ferrer, Detection of coherent scatterers in SAR data: algorithms and applications, Ph.D. thesis, ETH Zurich, 2013. [11] R. Touzi, “On the use of polarimetric SAR data for ship detection,” IGARSS Geoscience and Remote Sensing Symposium, vol. 2, pp. 812–814, 1999. [12] M. Yeremy, G. Geling, M. Rey, B. Plache, and M. Henschel, “Results from the crusade ship detection trial: polarimetric sar.,” Proceeding on IGARSS 2002, 2002. [13] A. Marino, “A notch filter for ship detection with polarimetric sar data,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 6(3), 2013. [14] A. Marino, M. Sugimoto, K. Ouchi, and I. Hajnsek, “Validating a notch filter for detection of targets at sea with alos-palsar data: Tokyo bay ,,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. early access, 2013. [15] F. Nunziata, M. Migliaccio, and C.E. Brown, “Reflection symmetry for polarimetric observation of manmade metallic targets at sea,” IEEE Journal of Oceanic Engineering, vol. 37(3), pp. 384–394, 2012.
(a) RGB Pauli
(b) ROC with HV channel
(c) ROC with HH channel
(d) ROC with VV channel
Fig. 2. Detectors ROC curves for sub-look analysis in the range direction. RADARSAT-2 Fine quad-pol data. Boxcar filter: [5,9]. (a) HV channel; (b) HH channel; (c) VV channel. Vessels analysed: 69; Pixels used for Pf : ∼ 107 . Black: Sublook correlation (Solid: filtered with no overlap; Dashed: filtered 30% overlap; Dotted: unfiltered with 30% overlap); Blue: Sub-look coherence; Green: GLRT; Violet: Intensity (Dotted: no average; Solid: averaged); Red: Sub-look entropy.
981
ETH Zurich, Institute of Environmental Engineering, Zurich, Switzerland German Aerospace Center (DLR), High Frequency Department, Oberpfaffenhofen, Germany 3 Korean Institute of Ocean Science and Technology, Korea Ocean Satellite Center, Ansan, South Korea 2
I. INTRODUCTION Ship detection is an important topic for security and surveillance of maritime and costal areas. A solution exploiting satellite-borne SAR sensors is particularly interesting, because it offers wide scale surveillance capabilities, which are not reliant on solar illumination and are rather independent of weather conditions ([1], [2], [3]). In SAR images, the main feature of a ship is a relatively large backscattering signal, which is usually brighter in comparison with the sea background. This led to the idea of using the intensity contrast as a feature to discriminate between targets and sea clutter. Several methodologies were proposed ([1], [3]). Most of these techniques set a statistical test between target and clutter background. Recently, the several ship detectors were proposed that exploits the property of SAR images to perform detection. In this work, two methodologies used for coherent scatterer detection are tested for the first time for ship detection and a comparison of ship detectors based on spectral analysis is performed over L-band ALOS date accompanied by a ground survey.
Gaussian process) [5], [4]. Interestingly, the sea present such behavior and therefore a procedure that is able to check for stability of targets over the spectrum could be exploited to understand if a target is present in the scene [6], [7]. In this section the different detectors used in the following comparison are presented. As a general recommendation in the formation of the sub-images is that the spectrum in range and azimuth is weighted by some windowing procedure aimed at reducing side-lobes in the spatial domain. The windowing can be easily removed considering a multiplication for the inverse of the power spectrum. II-A. Sub-look coherence This is the first detector exploiting sub-images to perform ship detection (please note different names were proposed in the papers) [6], [7]. The idea is to split the spectrum two nonoverlapping portions and consider the coherence between the two: γ=p
II. SHIP DETECTION WITH SPECTRAL ANALYSIS This section aims at presenting several algorithms exploiting properties of the spectrum of SAR images to perform ship detection. The idea behind SAR spectral analysis considers the evaluation of the Fourier transform of images (in azimuth and range) and the extraction of portions of the spectrum (following some rational) [4]. The spectrum sections are then anti-transformed (in order to obtain images with lower resolution, here defined as sub-images) and combined together with some procedure. In the following the image spectrum will be indicated with S, while the image in the spatial domain is s. A sub-image will be referred as si (with i and index indicating the image), corresponding to a portion of the spectrum called Si . The idea behind using spectral analysis for detection is that point targets (i.e vessels) are supposed to have a spectral response that is quite stable and constant in the spectral axes and therefore they will show a relatively high coherence. On the other hand a target with presents a completely developed speckle (the single look pixel can be modeled as a complex
978-1-4799-5775-0/14/$31.00 ©2014 IEEE
978
|hs1 · s∗T i| p2 ∗T hs1 · s1 i hs2 · s∗T 2 i
(1)
where ∗ and T stand for conjugate and transpose respectively and hi defines a spatial average considering a boxcar filter on the sub-images. II-B. Sub-look cross-correlation A methodology based on the same principle of the sublook coherence is presented in [8]. There are two main differences with the previous methodology, firstly the cross correlation between sub-images is not normalized: ρ = |hs1 · s∗T 2 i|.
(2)
secondly, the sub-images are obtained by portion on the spectrum that can overlap in the spectral domain. This has the advantage that vessels which do not present a very strong point target behavior will be detected as well, but the sea will not be rejected as strongly as in the case of non-overlapping spectral portions. Nevertheless, it is showed in [8] that the contrast sea-vessel in ρ images can be improved overlapping the portions of spectrum.
IGARSS 2014
II-C. Sub-look entropy The coherence is a useful operator to express the correlations between two images. however, when more images are considered, several coherences can be estimated (for each couple of images). A methodology to express the correlation between a stack of images is considering the covariance matrix of all the images and derive the entropy of the eigenvalues. If the entropy is zero (only one eigenvalue is different from zero) than the rank of the covariance matrix is unitary and all the images represent the same target (i.e. in the pixel there is a target stable over the different spectral portions) [9]. If x represents pixel by pixel a vector of images obtained with spectral portions (i.e. x = [x1 , ..., xn ]T , where n subimages are considered). The covariance matrix is calculated as [X] = x · x∗T . If λi are the n eigenvalues obtained by the diagonalisation of [X], their entropy is calculated as: H=−
n X
pi logn pi , where pi =
i=1
λi . T race([X])
(3)
II-D. GLRT Recently, a methodology to detect coherent targets using a Generalized Likelihood Ratio Test was proposed [10]. The algorithm push further the idea of overlapping spectral portions and set a statistical test based on a statistical model for the sea clutter (supposed with a fully developed speckle) and vessels. The analytical expression of the GLRT is: LG =
|a∗T M −1 x|2 (a∗T M −1 a) (x∗T M −1 x)
(4)
where x represent a vector of images obtained with spectral portions. a is a vector that depends on the location of the scatterer in the resolution cell (in this detections the scatterers are assumed at the center of the resolution cell), M is a matrix considering the correlation for a complex Gaussian sea clutter due to the overlapping of the spectral portions (the methodology to calculate it is presented in [10]). The bilinear form LG is one for perfectly coherent (i.e. spectrally stable) targets. III. ALGORITHM COMPARISON III-A. Presentation of the dataset Six Fine Quad-polarimetric images were acquired during winter 2013/2014 in the North Sea and Kattegat sea. The data were collected under the SOAR project EI-5145. Figure 1 shows the locations of the images as polygons on Google Earth. One aim of the SOAR project EI-5145 is to test ship detectors under severe weather conditions. In order to amplify the effects of strong sea clutter, large incidence angles (higher than 350 ) were avoided. One acquisition (03/12/2013) also includes a very steep incidence angles (e.g. ∼ 21.50 ). In
979
order to increase the probability of observing rough sea conditions the test areas were selected with relatively high latitudes and during the winter season. On the other hand, care was taken to select areas with large traffic of vessels. The scenes in the North Sea are close to the harbours of Rotterdam and Amsterdam, while the ones in the Kattegat Sea are around the Anholt Island, along the shipping route from the North Sea to the Baltic Sea. In total, 69 validated ships were observed with a variety of dimensions (ranging between 30 m to 200 m in length) and typology (e.g. fishing boats, cargos, etc). 69 vessels translates in a quantisation error for the estimation of Pd equal to 0.015. III-B. Comparison The first test considers . Figure 2 presents the ROC curves with a boxcar window of 5x9 pixels for the three polarisation channels HH, V V , HV . The cross-polarisation (HV ) channel is able to provide better ROC curves compared with the co-polarisation ones (HH and V V ). This result was also observed by other researchers [1], [11], [12]. Accordingly to the Bragg model, the sea clutter has a very low (or even null) backscattering in the cross-channel, leading in many circumstances to an increased contrast. Several works deal with exploiting polarimetric information to improve the detection performance [13], [14], [15]. Regarding the comparison of different detectors, the GLRT provides the best detection performance. Please note, in the following the comparisons will be mostly done considering rather high values of Pd (e.g. greater than 0.8), since this are the operation scenarios that are most interesting. Following the GLRT, there are the sub-look coherence and entropy. Regarding the sub-look correlation, two versions were tested in this paper following the observation that a different overlap of sub-portions can provide different results [8]. For this reason, one version (solid black line) considers non overlapping sub-portions and the other (black dashed line) exploits sub-portion with 30% overlap. As explained previously, the algorithms are compared with two detectors based on the intensity considering SLC (without filtering) or multi-look (boxcar filtering) images [1]. Interestingly, the intensity detector on the cross-polarisation show large performance loss only if compared with the GLRT. This is an indicator that, if the cross channel is available, the gain of acquiring SLC instead than detected images would be evident only if a powerful methodology is exploited. This is not true for the co-polarisation channel, where the sub-look process provides benefits for all the sublook detectors at exception than the cross-correlation. ACKNOWLEDGMENTS c MacDonald, DetRADARSAT-2 Data and Products twiler and Associates Ltd. (2013-2014) - All Rights Re-
Fig. 1. Locations (red polygons) of the six RADARSAT-2 Fine Quad-pol acquisitions used in the comparison. Google Earth.
served. RADARSAT is an official trademark of the Canadian Space Agency. IV. REFERENCES [1] D. J. Crisp, “The State-of-the-Art in ship detection in Synthetic Aperture Radar imagery,” Australiane Government Department of Defence, 2004. [2] C. Liu, P. W. Vachon, and G. W. Geling, “Improved ship detection using polarimetric SAR data,,” IGARSS Geoscience and Remote Sensing Symposium, vol. 3, pp. 1800–1803,, 2004. [3] P. W Vachon, “Ship detection in synthetic aperture radar imagery.,” Proceedings OceanSAR, St. John s, NL, Canada, 2006. [4] J.-C. Souyris, C. Henry, and F. Adragna, “On the use of complex sar image spectral analysis for target detection: Assessment of polarimetry,” IEEE Trans. on Geos. & Rem. Sen., vol. 41(12), pp. 2725–2734, 2003. [5] C. Oliver and S. Quegan, Understanding Synthetic Aperture Radar Images., SciTech Publishing, Inc, 2004. [6] A. Arnaud, “Ship detection by SAR interferometry,” IGARSS’99, vol. 5, pp. 2616–2618, 1999. [7] K. Ouchi, S. Tamaki, H. Yaguchi, and M. Iehara, “Ship detection based on coherence images derived from cross correlation of multilook SAR images,” IEEE Geoscience and Remote Sensing Letters, vol. 1(3), 2004. [8] C. Brekke, S. N. Anfinsen, , and Y. Larsen, “Subband extraction strategies in ship detection with the subaperture cross-correlation magnitude,” IEEE Geoscience and Remote Sensing Lettersg, vol. 10(4), pp. 786–790, 2013. [9] R. Z. Schneider, K. P. Papathanassiou, I. Hajnsek, and A. Moreira, “Polarimetric and interferometric characterization of coherent scatterers in urban areas,” IEEE Trans. on Geos. & Rem. Sen., vol. 44(4), pp. 971–984, 2006.
980
[10] M. J. Sanjuan-Ferrer, Detection of coherent scatterers in SAR data: algorithms and applications, Ph.D. thesis, ETH Zurich, 2013. [11] R. Touzi, “On the use of polarimetric SAR data for ship detection,” IGARSS Geoscience and Remote Sensing Symposium, vol. 2, pp. 812–814, 1999. [12] M. Yeremy, G. Geling, M. Rey, B. Plache, and M. Henschel, “Results from the crusade ship detection trial: polarimetric sar.,” Proceeding on IGARSS 2002, 2002. [13] A. Marino, “A notch filter for ship detection with polarimetric sar data,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 6(3), 2013. [14] A. Marino, M. Sugimoto, K. Ouchi, and I. Hajnsek, “Validating a notch filter for detection of targets at sea with alos-palsar data: Tokyo bay ,,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. early access, 2013. [15] F. Nunziata, M. Migliaccio, and C.E. Brown, “Reflection symmetry for polarimetric observation of manmade metallic targets at sea,” IEEE Journal of Oceanic Engineering, vol. 37(3), pp. 384–394, 2012.
(a) RGB Pauli
(b) ROC with HV channel
(c) ROC with HH channel
(d) ROC with VV channel
Fig. 2. Detectors ROC curves for sub-look analysis in the range direction. RADARSAT-2 Fine quad-pol data. Boxcar filter: [5,9]. (a) HV channel; (b) HH channel; (c) VV channel. Vessels analysed: 69; Pixels used for Pf : ∼ 107 . Black: Sublook correlation (Solid: filtered with no overlap; Dashed: filtered 30% overlap; Dotted: unfiltered with 30% overlap); Blue: Sub-look coherence; Green: GLRT; Violet: Intensity (Dotted: no average; Solid: averaged); Red: Sub-look entropy.
981
