1. Introduction
Synthetic Aperture Radar (SAR) provides useful remote sensing data for monitoring vessel activities due to its capacity to make observation regardless of sunlight and weather conditions (Ouchi, 2004; Chan and Koo, 2008; Brusch et al., 2011). Thus, there are many studies have been made on ship detection using SAR images (Brush et al., 2011; Greidanus et al., 2004).
A commonly used ship detection method is a constant false alarm rate (CFAR) detector, which detects targets brighter than their surroundings (Crisp, 2004). Another major approach is to use phase information from SAR dual-polarization complex data (Pelich et al., 2019). Other studies use a neural network to classify ships(Wang et al., 2018). Each approach has its pros and cons, but regardless of approach, noise can cause false alarms. Another popular algorithm is the adaptive threshold method (ATM) using the same moving window.
While thermal noise, affecting the entire imagery, is reported lower noise level than noise equivalent sigma zero (NESZ) and can be easily de-noised by noise vectors in the product metadata (Meadows, 2019), while radio frequency interference (RFI) and azimuth smearing cause numerous false alarms in ship detection due to the high noise level.
RFI is the interference from human activities such as radio, mobile communication, television, and other satellites(Meyer et al., 2013; Santamaria et al., 2017). Since the receiving power of cross-polarization from the sea surface is weaker than co-polarization, the effect of RFI on cross-polarization is stronger (Meyer et al., 2013).
Azimuth smearing also hampers ship detection results. Since moving objects produce phase errors in the SAR signal, multiple nearby peaks are generated (Fienup and Kowalczyk, 1995; Liu, 2015). It also results in a longer length estimation of ships than the true length (Greidanus and Kourti, 2006). For a similar reason for RFI, azimuth smearing is also apparent in cross-polarization.
Although both RFI and azimuth smearing also appear in co-polarization, the signal-to-noise ratio (SNR) is much high compared to cross-polarization. However, cross-polarization is still advantageous in ship detection applications because it is less sensitive to wind speed and direction (Vachon and Wolfe, 2010; Hannevik et al., 2015).
There are two main approaches for handling false alarms from RFI. One is the deep learning technique, like a convolution neural network (CNN). According to Wang’s research (Wang et al., 2018), CNN can distinguish ships from false alarms caused by RFI. However, CNN can be time-consuming when the number of objects to classify is large.
Another approach is mitigating or filtering RFI signals from raw or single look complex (SLC) data. Although there are many studies about this approach, all of them require SAR phase information that is not exist in the Sentinel-1 ground range detected (GRD) product (Tao et al., 2019; Meyer et al., 2013).
Although there is no automated method to mitigating RFI from GRD product due to the absence of phase information, GRD product is still useful for ship detection because of its open access policy within 3 hours after sensing.
Therefore, there is no choice but to mask areas where RFI frequently occurs (Santamaria et al., 2017). This approach has a limitation of missing vessels in the masked area. Azimuth smearing caused by phase error cannot be corrected in the Sentinel-1 GRD product for the same reason.
Hence, the purpose of this study is to reduce false alarms caused by RFI and azimuth smearing for ship detection with ATM in Sentinel-1 GRD data. In particular, it will be shown that VH polarization is suitable for ship detection in comparison with VV polarization, although it is vulnerable to RFI and azimuth smearing i.e., the false alarms can be mitigated when both polarizations blended properly.
Thus, we introduce an effective method combining two polarizations to reduce false alarms from RFI and azimuth smearing. The key strength and novelty of this work are its simplicity and efficiency to reduce numerous false alarms using only a single pixel of intensity value for each polarization.
2. Dataset
In this work, we used the Sentinel-1 GRD products over the waters off Busan, South Korea. The period of the dataset acquisition is from Dec. 28, 2017 to Dec. 17, 2018 as shown in Table 1. The Sentinel-1 data are often radio interfered in this spectrum of wavelength, i.e., C-band; therefore, RFI noise Is frequently observed. During this period, a total of 43 images of Sentinel-1A/B covering the study area were downloaded from the Copernicus Open Access Hub (https://scihub. copernicus.eu/dhus/).Among the collected 43 images, 20 images have RFI noise with the peak intensity higher than -22 dB. Also, azimuth smearing appears in every Sentinel-1 data.
Table 1. Sentinel-1 IW GRD dataset used in this study
Fig. 1 depicts the area of interest(AOI)in this study. The yellow rectangle indicates the Sentinel-1A AOI, and the green rectangle is AOI of Sentinel-1B. Fig. 2 is an example of the false-color Sentinel-1 image, showing RFI and azimuth smearing noises where VH-polarization is in red and green, i.e., yellow and VV-polarization in blue. As noted in the introduction section, RFI and azimuth smearing are strong in VH-polarization as seen in yellow color.
Fig. 1. Area of interest (AOI) in this study. The yellow rectangle is AOI of Sentinel-1A data, and the green rectangular is AOI of Sentinel-1B data.
Fig. 2. (Top) Sentinel-1B false-color image acquired on Sep. 24, 2018. Red: VH, Green: VH, and Blue: VV. (Bottom left) the yellow rectangular area in the top image showing RFI in yellow and azimuth smearing in the green box. (Bottom right) Green rectangle in the bottom left. Since yellow is a combination of red and green, radio frequency interference (RFI) and azimuth smearing, which is strong at VH polarization, have a yellow color.
3. Methods
For the quantitative analysis of noise mitigation, we first calculated the median intensity of sea surface, RFI, and azimuth smearing for 20 data. As shown in Fig. 2, RFI usually appears in extended areas, and azimuth smearing appears around the small areas close to moving ships.
1) Statistics
The image format is, initially, the normalized radar cross-section (NRCS) in the square meter unit, but in this paper, it is translated into decibel value for better presentation.
Usually, in the sea surface images without RFI and smearing noises, the intensity of VV polarization is higher than VH polarization by about 4 dB to 12 dB. The median difference between the two polarizations of a clean sea surface is 8.17 dB.
Since the effect of RFI in VH polarization is stronger than VV polarization, the intensity difference between the two polarizations varies in the RFI and smearing areas. The median difference in intensity between the VH and VV data is 1.81 dB for RFI affected area. Similarly, azimuth smearing has a greater effect on VH polarization than VV. The median difference between both polarization data is 0.82 dB.
For the RFI, only 10% of VV pixels are 5.93 dB higher than VH pixels in intensity. Similarly, only 10% of VV pixels are 6.53 dB higher than VH pixels in azimuth smearing. In other words, more than 90% of pixels affected by RFI or azimuth smearing have less than 6.53 dB difference in intensity between VV and VH polarizations.
Also, we investigated the correlation between pixel intensity, incidence angle, and wind speed in the sea surface without RFI and azimuth smearing.
Although VH polarization has noticeable thermal noise (see the left of Fig. 3), the coefficient of variation (CV) of VH is less than VV. The absolute CV of VV polarization is 0.06, while the CV of VH polarization is 0.04. It means that VV polarization is more sensitive to the incidence angle. It is reported that cross-polarization is a more stable incidence angle characteristic rather than co-polarization in ship detection (Hannevik et al., 2015).
Fig. 3. (Left) The relationship between the incidence angle and intensity of background sea in the scene number 18 of Table 2. (Right) The relationship between the wind speed from Weather Research and Forecasting (WRF) model (Skamarock et al., 2008) generated in the Korea Operational Oceanographic System (KOOS) (Park et al., 2015) and median intensity of background sea in the scene number 15 in Table 2.
Cross-polarization is less affected by wind speed than co-polarization and independent of wind direction (Vachon and Wolfe, 2010). We confirmed this fact by correlating the median image intensity of background sea surface with wind speeds from the Weather Research and Forecasting (WRF) (Skamarock et al., 2008) model calculated in the Korea Operational Oceanographic System (KOOS) (Park et al., 2015). The right of Fig. 3 shows the verification result. Unlike VH polarization, VVpolarization is strongly influenced by wind.
2) Assumptions
The image format is, initially, the normalized radar cross-section (NRCS) in the square meter unit, but in this paper, it is translated into decibel value for better presentation.
We made two assumptions based on the statistics as follows. 1) VH polarization is less affected by incidence angle and wind speed, and therefore it is suitable for ship detection in comparison with VV polarization when RFI and azimuth smearing do not exist. 2) Only 10% of RFI and azimuth smearing pixels at VV polarization are 5.93 dB and 6.53 dB higher than VH polarization on average, respectively. Therefore, more than 90% of the noise pixels meet the following condition.
VVi, j – VHi, j < 6.53 dB (1)
In equation (1), VVi,j and VHi,j represent the intensity values of the pixels located in ith row and jth column of VV and VH images, respectively.
The principle is that false alarms can be mitigated if VH polarization is used as a basis, where a VH pixel value contaminated by noise is converted into the adjusted VV polarization intensity values. Thus, under the condition of equation (1), the following band-math formula can be made.
\(\text { value }_{i j}=\left\{\begin{array}{c} V V_{i j}-6.53 d B, \text { if } V V_{i j}-V H_{i j}<6.53 d B \\ V H_{i j}, \text { otherwise } \end{array}\right.\) (2)
Equation (2) can be applied to the entire image, and pixel value conversion, which changes VH pixel intensity into adjusted VV polarization intensity, is performed only for pixels that satisfy the condition of equation (1).
3) Assumption test
To verify the equation (2), we performed the ship detection on the dataset mentioned in Table 2 using Adaptive Threshold Method (ATM). Equation (2) is expected to reduce false alarms because the noise pixels of VH polarization are replaced with pixels of VV polarization with less noise effect.
Table 2. Median intensity of ocean, RFI and azimuth smearing (unit: decibel representation of normalized radar cross section)
The workflow of the entire ship detection procedure for testing is depicted in Fig. 4. For comparison, we performed the ship detection in Fig. 4 and did one more time except for the proposed method as a control group.
Fig. 4. Overall ship detection flow. The proposed bandmath stage is located in the fourth step. Subscripts of i and j indicate pixels located in ith row and jth column of the image.
The preprocessing steps include radiometric calibration (Miranda and Meadows, 2015) and converting digital numbers into NRCS. Thermal noise reduction (Hong et al., 2018) has been added to mitigate the thermal noise described in Fig. 3. This is a statistical method that adds a constant depending on the incidence angle. The Range-Doppler terrain correction method (Small and Schubert, 2008) is also included to correct geometric errors.
After the preprocessing, noise reduction is carried out by the proposed method and also by the speckle filter. For the latter, we adopt Lee filter (Lee, 1980) that uses local statistics and has an edge-preserving feature. Sliding window based ATM is performed after applying the land masking (Yang et al., 2018) to detect locally bright pixels as follows.
xtarget > μbackground + σbackground * t (3)
\(t=\left\{\begin{array}{c} 4, \text { if }-12.21 d B<x_{\text {target }} \\ 10, \text { if }-15.23 d B<x_{\text {target }} \leq-12.21 d B \\ 12, \text { if }-20.00 d B<x_{\text {tanget }} \leq-15.23 d B \\ 14, \text { if } x_{\text {target }} \leq-20.00 d B \end{array}\right.\) (4)
In equation (3), xtarget is the average intensity of the target window, μbackground is the average intensity and σbackground is the standard deviation of the background window. t is the threshold parameter. In the test, we configured the target window size to 3 pixels, the guard window size 21 pixels, and the background window size 31 pixels.
Based on this configuration, the ATM filter extracts targets which meets equation (3). In equations (3) and (4), t is designed to be different depending on xtarget as an empirical interim measure. In general, ATM detector sometimes detects the low-intensity objects at the edge of the image or a calm sea. In this case, even a weak noise can cause a false alarm. To prevent this, in the case of a dark target, a higher contrast against the background is required.
Finally, the automatic identification system (AIS) data are interpolated in the time-space to match the AIS and SAR data acquisition times for the identification of the detected ships.
4) Test result
Without equation (2), ship detection using VH polarization resulted in 9,059 false alarms from the RFI and 220 false alarms from the azimuth smearing, respectively, as in Table 3.After applying the proposed method, the ship detection result showed 2,320 false alarms (74.39% decrease) in RFI and 17 false alarms (92.27% decrease) in azimuth smearing, respectively.
Table 3. Number of false alarms in ship detection result
Among the 20 images, scenes 8 and 19 still have numerous false alarms caused by RFI even after applying the proposed method. It is because the effect of the RFI is stronger than of other images. That is, as in Table 2, in scenes 8 and 19, the intensity of RFI of VV was 9.0 dB and 16.98 dB higher than the sea surface, respectively, compared to other images (on average, difference of 2.55 dB).
Verification using the AIS data was made for 18 scenes, except for scenes 8 and 17, which had no AIS data. Except for scene 20, the total number of detected vessels are similar or slightly lower in the columns of AIS matching count for detections of both VH only and proposed method in Table 3. It is related to the land masking method used in this study (Yang et al., 2018). Because the land masking method used in this study detects and masks shoreline dynamically, the pixel values of the shoreline changed by the proposed method make the impact on the shoreline detection result. Therefore, land masking method over-masked ships that are close to the land.
Fig. 5 shows how the proposed method worked on an image affected by RFI and azimuth smearing. Before using the proposed method, lots of false alarms appeared in the area masked by black box which intensity of noise is higher than -20 dB, colored by yellow and red. Since these were spread over large portion of the image, the number of false alarms often exceeds thousand in severe cases, like scene 1, 4, 8, and 19. After applying the proposed method, the noise intensity was lowered. Therefore, the difference in median intensity between ship and azimuth smearing has increased by 4.92 dB. Meanwhile, the median difference between ship and RFI increased by 7 dB. It means improved SNR for the ship and contrast with the background.
Fig. 5. Pseudocolor images before and after the application of the proposed method for bottom part of Fig. 2. Left two images show before the application of the proposed method. The top and bottom areas in black boxes are affected by radio frequency interference and azimuth smearing, respectively. These noises show yellow to red color due to the high intensity. The right two images show the after application results. The intensity of pixels in the black boxes has decreased.
Additionally, we also investigated the impact of the proposed method on ship images using intensity statistics of 100 ships with an average length of 145 m as in Fig. 6. The average intensity was changed in 80 ships, and the median intensity was changed in 71 ships. The average intensity change for all 100 vessels was –0.45 dB.
Fig. 6. Comparison of the median intensity of 100 ships in the VH polarization image and the proposed method.
4. Conclusion
In ship detection using the Sentinel-1VHpolarization images, RFI and azimuth smearing often cause numerous false alarms. In contrast, VVpolarization has a high SNR for RFI and azimuth smearing. Nevertheless, if the RFI and azimuth smearing is not exist, VH polarization is more suitable for ship detection.
To reduce false alarms from RFI and azimuth smearing in the VH polarization, we designed a band-math formula using VV polarization data. Test on the 20 Sentinel-1 images showed a 74.39% reduction in RFI caused false alarms, and 92.27% reduction in azimuth smearing caused false alarms. The proposed method has the side effect of decreasing the average intensity by 0.45 dB for ships, however, missing target from this adverse effect is very few. Also, this method consumed only about 10 seconds of computation time on systems with the i7-6800K. It is a key strength of this work because the ship detection system often requires a near-real-time operation.
Acknowledgements
This research is a part of the projects entitled “Development of Satellite-based System on Monitoring and Predicting Ship Distribution in the Contiguous Zone”, “Construction of Ocean Research Stations and their application Studies”, “Technology Development for Practical Applications of Multi-Satellite Data to Maritime Issues”, funded by the Ministry of Oceans and Fisheries, Korea.
참고문헌
- Brusch, S., S. Lehner, T. Fritz, M. Soccorsi, A. Soloviev, and B. V. Schie, 2011. Ship surveillance with TerraSAR-X, IEEE Transactions on Geoscience and Remote Sensing, 49(3): 1092-1103. https://doi.org/10.1109/TGRS.2010.2071879
- Chan, Y. K. and V. C. Koo, 2008. An introduction to synthetic aperture radar (SAR), Progress In Electromagnetics Research B, 2: 27-60. https://doi.org/10.2528/PIERB07110101
- Crisp, D. J., 2004. The state-of-art in ship detection in synthetic aperture radar imagery, Defense Science and Technology Organization, Fairbairn, Canberra, Australia.
- Fienup, J. R. and A. M. Kowalczyk, 1995. Detecting moving targets in SAR imagery by using a phase-error correction algorithm, Proc. of Algorithms for Synthetic Aperture Radar Imagery II, Bellingham, WA, Jun. 5, vol. 2487, pp. 337-344.
- Greidanus, H., P. Clayton, M. Indregard, G. Staples, N. Suzuki, P. Vachoir, C. Wackerman, T. Tennvassas, J. Mallorqui, N. Kourti, R. Ringrose, and H. Melief, 2004. Benchmarking operational SAR ship detection, Proc. of IEEE International Geoscience and Remote Sensing Symposium, Anchorage, AK, Sep. 20-24, vol. 6, pp. 4215-4218.
- Greidanus, H. and N. Kourti, 2006. Findings of the DECLIMS project-detection and classification of marine traffic from space, Proc. of SEASAR: Advances in SAR Oceanography from Envisat and ERS Missions, Frascati, Italy, Jan. 23-26.
- Hannevik, T. N., K. Eldhuset, and R. Olsen, 2015. Improving ship detection by using polarimetric decompositions, Norwegian Defense Research Establishment, Kjeller, Norway.
- Hong, D. B., J. J. Bae, and C. S. Yang, 2018. Automatic mapping technique of sea ice in the coastal waters of the Arctic Ocean using Sentinel-1 data, Journal of Coastal Research, 85: 556-560. https://doi.org/10.2112/SI85-112.1
- Lee, J., 1980. Digital image enhancement and noise filtering by use of local statistics, IEEE Transactions on Pattern Analysis & Machine Intelligence, PAMI-2(2): 165-168. https://doi.org/10.1109/TPAMI.1980.4766994
- Liu, C., 2015. A dual-polarization ship detection algorithm, Defense Research and Development, Ottawa, Canada.
- Meadows, P. J., 2019. S-1A & S-1B annual performance report for 2018, European Space Agency, Paris, France.
- Meyer, F. J., J. B. Nicoll, and A. P. Doulgeris, 2013. Correction and characterization of radio frequency interference signatures in l-band synthetic aperture radar data, IEEE Transactions on Geoscience and Remote Sensing, 51(10): 4961-4972. https://doi.org/10.1109/TGRS.2013.2252469
- Miranda, N. and P. J. Meadows, 2015. Radiometric calibration of S-1 Level-1 products generated by the S-1 IPF, European Space Agency, Paris, France.
- Ouchi, K., 2004. Principles of synthetic Aperture Radar for remote sensing, Tokyo Denki University, Tokyo, Japan.
- Park, K.-S., K.-Y. Heo, K. Jun, J.-I. Kwon, J. Kim, J.-Y. Choi, K.-H. Cho, B.-J. Choi, S.-N. Seo, Y.-H. Kim, S.-D. Kim, C.-S. Yang, J.-C. Lee, S.-I. Kim, S. Kim, J.-W. Choi, and S.-H. Jeong, 2015. Development of the operational oceanographic system of Korea, Ocean Science Journal, 50(20): 353-369. https://doi.org/10.1007/s12601-015-0033-1
- Pelich, R., M. Chini, R. Hostache, P. Matgen, C. Lopez- Martinez, M. Nuevo, P. Ries, and G. Eiden, 2019. Large-scale automatic vessel monitoring based on dual-polarization Sentinel-1 and AIS data, Remote Sensing, 11(9): 1078. https://doi.org/10.3390/rs11091078
- Santamaria, C., M. Alvarez, H. Greidanus, V. Syrris, P. Soille, and P. Argentieri, 2017. Mass processing of Sentinel-1 images for maritime surveillance, Remote Sensing, 9(7): 678-697. https://doi.org/10.3390/rs9070678
- Skamarock, W. C., J. B. Klemp, J. Dudhia, D. O. Gill, D. M. Barker, M. G. Duda, X.-Y. Huang, W. Wang, and J. G. Powers, 2008. A description of the Advanced Research WRF version 3, National Center for Atmospheric Research, Boulder, CO, USA.
- Small, D. and A. Schubert, 2008. Guide to ASAR geocoding, European Space Agency, Paris, France.
- Tao, M., J. Su, Y. Huang, and L. Wang, 2019. Mitigation of radio frequency interference in synthetic aperture radar data: current status and future trends, Remote Sensing, 11(20): 2438. https://doi.org/10.3390/rs11202438
- Vachon, P. W. and J. Wolfe, 2010. C-band crosspolarization wind speed retrieval, IEEE Geoscience and Remote Sensing Letters, 8(3): 456-459. https://doi.org/10.1109/LGRS.2010.2085417
- Wang, Y., C. Wang, and H. Zhang, 2018. Ship discrimination with deep convolutional neural networks in SAR Images, Proc. of 2018 IEEE International Geoscience and Remote Sensing Symposium, Valencia, Jul. 22-27, pp. 8444-8447.
- Yang, C. S., J. H. Park, and H.-A. Rashid, 2018. An improved method of land masking for synthetic aperture radar-based ship detection, The Journal of Navigation, 71(4): 788-804. https://doi.org/10.1017/S037346331800005X