1. Introduction
The radiation measured at a satellite sensor can be attenuated by atmospheric effects such as scattering and absorption while transmitting from the land surface to the sensor (Proud et al., 2010). Because such atmospheric effects in remote sensing cause uncertainty in surface observation, accurate atmospheric correction is an essential preprocessing step for the analysis of surface characterization and environmental monitoring (Kaufman, 1987). The physical model-based method has the advantage of calculating the atmospheric contribution numerically using precise Radiative Transfer Model (RTM) (Lee and Yum, 2019) such as Moderate Resolution Atmospheric Transmission (MODTRAN) (Kneizys, 1996), High-Resolution Transmission Molecular Absorption Database (HITRAN) (Rothman et al., 2013), and Second Simulation of a Satellite Signal in the Solar Spectrum (6S) (Vermote et al., 1997).
Since the gas concentration in the atmospheric particles has a slight variation in space and time, scattering and absorption by gas molecules are not difficult to calculate. However, the distribution of aerosol and water vapor is quite variable in space and time, so the elimination of the effects of aerosol and water vapor in the atmospheric correction is essential (Lee, 2019). Especially in Korea affected by the aerosol inflow from the continent by a westerly wind, atmospheric correction can have more importance. Understanding the optical properties of aerosols is necessary for the accurate observation of land surfaces (Lee and Kim, 2008). The particle size of aerosol is similar to the wavelength of sunlight, which can cause Mie scattering (forward scattering is greater than backward scattering). The attenuation of radiation by Mie scattering makes the differences between the radiation calculated at the satellite sensor and the radiation measured at the land surface.
Thus, it is necessary to use high-quality aerosol optical depth (AOD) data for an appropriate atmospheric correction of high-resolution satellite images. MODIS AOD has been used for the atmospheric correction of Landsat images (Xie et al., 2010; Nazeer et al., 2014; Yusuf et al., 2018) and Sentinel-2 images (Martins et al., 2017). Kim et al. (2021) has compared multiple AOD products (MODIS, VIIRS, Himawari-8, and Sentinel-3) with AERONET sunphotometer observations in South Korea, which showed the best performance of MODIS AOD. However, for more reasonable use of AOD data for atmospheric correction, the atmospheric correction results using the raster AOD like MODIS data and the single-point AOD like AERONET data should be compared objectively. In this study, we examined the 6S-based atmospheric correction results for the Sentinel-2 images in South Korea using raster AOD (MODIS) and single-point AOD (AERONET).
2. Data
1) Sentinel-2 image
The experiment was conducted for the area of 33 to 38.5°N and 125.8 to 129.8°E. Fig. 1 shows 21 AERONET stations operated in South Korea during the period from January 2015 to December 2019. Table 1 shows the date and time for the Sentinel-2 images shown in Fig. 1.
Table 1. Information of the Sentinel-2 images used in the study
Fig. 1. Location of the scenes used in this study.
European Satellite Agency (ESA) Sentinel-2 satellite has a wide swath of 290 km and provides high resolution (10, 20, and 60 m) images. The twin satellites, Sentinel-2A and -2B, were launched on June 23, 2015 and March 7, 2017, respectively, in the same sun-synchronous orbit phased at 180° to each other. Each satellite carries a multispectral instrument (MSI) with 13 spectral bands, including visible (VIS), red-edge (RE), near-infrared (NIR), and short-wave infrared (SWIR) wavelengths (Table 2). Fig. 2 shows the Spectral Response Functions (SRF) for each band. We used the Level 1C products with top-of-atmosphere (TOA) reflectance. Blue, green, red, red-edge, and NIR bands in the cloud-free images were selected.
Table 2. Spectral bands and spatial resolutions of Sentinel-2 MSI
Fig. 2. Spectral response functions of Sentinel-2 MSI (Li et al., 2018).
2) MODIS AOD product
MODIS onboard Terra and Aqua satellites conducts the Earth observations at about 10:30 (Terra) and 13:30 (Aqua) local time. Multiangle Implementation of Atmospheric Correction (MAIAC) algorithm is for the aerosol retrieval and atmospheric correction for both light desert and dark vegetation surfaces (Lyapustin et al., 2011a; 2011b). It improves the accuracy of cloud detection, aerosol retrieval, and atmospheric correction through the combination of time-series analysis and image processing. A sliding window for the past several days is used to create a time-series composite, and a surface bidirectional reflectance distribution function (BRDF) derived from satellite observations of various geometries is used (Lyapustin et al., 2018) for more accurate surface reflectance. Terra and Aqua data were combined together for a cross-correction. We used the MAIAC-based AOD product with 1 km resolution (MCD19A2).
3) Auxiliary data
Multiple auxiliary data were used for atmospheric correction (Table 3). We extracted solar zenith angle (SZA), solar azimuth angle (SAA), viewing zenith angle (VZA), and viewing azimuth angle (VAA) from sentinel-2 metadata. For atmospheric conditions, total column water vapor (TCWV) from ERA5 reanalysis and total column ozone (TCO) from CAMS global reanalysis (EAC4) from the European Center for Medium-Range Weather Forecasts (ECMWF) were used. Digital surface model (DSM) with 30 m resolution from ALOS (Advanced Land Observing Satellite) satellite and the land cover map from the Ministry of Environment were also prepared.
Table 3. Data for 6S parameters
3. Methods
After the comparisons among the satellite AOD products in South Korea, we selected MODIS AOD. Using the 6S model with the input parameters such as geometric, atmospheric, spectral, and aerosol conditions, the surface reflectance for Sentinel 2 images was calculated through the atmospheric correction (Fig. 3).
Fig. 3. Flowchart for study.
1) Comparison between AOD products
Kim et al. (2021) conducted the accuracy comparison for the AOD products by MODIS, VIIRS, Himawari- 8, and Sentinel-3, by referring to the AERONET observations. Due to the difference in data availability, the accuracy characteristics according to seasonal and geographical differences were analyzed for MODIS and VIIRS in 2015-2019, Himawari-8 in 2016-2019, and Sentinel-3 in 2019. Unlike the daily data of AERONET, MODIS, and VIIRS, the hourly Himawari-8 was aggregated on an hourly basis. Sentinel-3 data has an interval of two days, and the nearest time operation was carried out for the daily matchup. The result showed that the AOD accuracy of the MODIS product was superior to others (CC=0.836).
2) 6S RTM for atmospheric correction
6S RTM was used to correct atmospheric effects in TOA reflectance measurements observed by the MSI. It can calculate the scattering and absorption effects of the atmospheric components such as water vapor, ozone, and aerosol for various geometrical, atmospheric, aerosol, and spectral conditions (Lee et al., 2020). The simulation performance of 6S is usually better than the other RTMs such as MODTRAN and SHARM (Kotchenova et al., 2006). In the 6S model, surface reflectance is derived by the following equation (1).
ρTOA(θs, θv, ø) =
\(T_{g}\left(\theta_{s}, \theta_{v}\right)\left[\rho_{R+A}+T^{\downarrow}\left(\theta_{s}\right) T^{\dagger}\left(\theta_{v}\right) \frac{\rho_{s}}{1-S \rho_{s}}\right]\) (1)
where, ρTOA is the TOA reflectance, θs, θv and ø represent the SZA, VZA, and relative azimuth angle (RAA), respectively. Tgis the gaseous transmission of H2O, CO2, O2, O3for the radiance. ρR+Aindicates the total reflectance of the molecule and aerosol scattering. T↓(θs) and T↑(θv) are the total transmission of the atmosphere on the path between the sun and the surface (T↓(θs)) and the surface and the sensor (T↑(θv)). S represents the spherical albedo of the atmosphere. ρsis the equation of transfer for a Lambertian homogeneous target of reflectance. When 6S RTM runs, it is generated that the three atmospheric correction coefficients called xa, xband xc. xadenotes the path radiance in reflectance unit, xbis the scattering term of the atmosphere, and xcis the same as S. The surface reflectance (ρTOC) is calculated by equation (2) using the atmospheric correction coefficients.
\(\rho_{T O C}=\frac{x a \cdot L-x b}{1+x c \cdot(x a \cdot L-x b)}\) (2)
L(in Wm–2sr–1μm–1) is the TOA radiance measured by the satellite sensor. For the experiment of Sentinel-2 images, we employed the Py6S library, an interface to the 6S through the Python programming language (Wilson, 2013).
4. Results and Discussions
1) Case of January 15, 2019
In order to compare the difference in accuracy when using raster AOD (MODIS) and single-point AOD (AERONET) for atmospheric correction, two experiments were conducted on a day with a very high concentration of fine dust. First, on January 15, 2019, an atmospheric correction was performed for an area of 2 km × 2 km near Ulsan. Fig. 4 shows the raster AOD distribution. The nearest AERONET point was Gwangju_GIST and had an AOD of about 0.91. The average daily PM10 recorded by Air Korea was 78 μg/m3, and the PM2.5 was 59 μg/m3. Table 4 shows the comparison between the raster and single-point AOD in terms of the difference in the band reflectance after atmospheric correction. We used the Sentinel-2 level 2 product for the reference data. The NIR band reflectance has the highest agreement with the level 2 data, with a root mean square difference (RMSD) of 0.022 when using raster AOD and an RMSD of 0.025 when using single point AOD. The Blue band reflectance showed the lowest agreement with the level 2 data, with an RMSD of 0.084 for raster AOD and an RMSD of 0.154 for single-point AOD. It is presumably because the shorter the wavelength, the greater the sensitivity to AOD (Jung et al., 2020).
Fig. 4. Distribution of raster AOD values (January 15, 2019).
Table 4. Differences between the 6S result and Sentinel-2 level 2 data according to AOD products (January 15, 2019)
Fig. 5. Land cover map for the Sentinel-2 image of January 15, 2019.
Fig. 6 is the scatter plot for the comparison between the 6S result and Sentinel-2 level 2 data. Because Sentinel-2 level 2 data was created by MODTRAN, the values of band reflectance could be different from the 6S result, but the overall trend should be the same. The scatter plots were closer to the 1:1 line when using raster AOD than when using single-point AOD. The histograms in Fig. 7 show that the reflectance values calculated using the single-point AOD have a few outlier cases.
Fig. 6. Scatter plots for the 6S result and Sentinel-2 level 2 data according to AOD products (January 15, 2019).
Fig. 7. Histogram of the band reflectance according to AOD products (January 15, 2019).
Fig. 8 to 9 shows that the reflectance calculated using single-point AOD had overall low values than using raster AOD. The difference was more significant in the visible band (Fig. 8). When referring to the land cover map (Fig. 5), a substantial difference was found, especially in the forest area. This means that the effect of AOD parameters is different depending on the wavelength and land cover. The reflectance difference map (Fig. 10) shows that such a difference was greater for the forest. Also, the reflectance difference was largest for the blue band and smallest for the NIR band. It was in the order of blue, green, red, red edge, and NIR. Normalized difference vegetation index (NDVI) was calculated using the reflectance of red and NIR band, according to the AOD data (Fig. 11). The NDVI values were almost agreed with the land cover when using raster AOD, but the single-point AOD yielded many extremely high or low NDVI values.
Fig. 8. Surface reflectance of visible bands according to AOD products (January 15, 2019).
Fig. 9. Surface reflectance of red edge and NIR bands according to AOD products (January 15, 2019).
Fig. 10. Reflectance difference between the result using raster AOD and the result using single-point AOD for atmospheric correction (January 15, 2019).
Fig. 11. NDVI calculated by the red and NIR reflectance using (a) raster AOD and (b) single-point AOD (January 15, 2019).
2) Case of March 4, 2020
For another case study, the atmospheric correction using raster and single-point AOD was performed on a 2 km × 2 km area near Gyeongsangbuk-do on March 4, 2019. A map for raster AOD distribution is shown in Fig. 12. The nearest AERONET station Hankuk_UFS with an AOD of 1.1. The average daily PM10 from Air Korea was 35 μg/m3, and the PM2.5 was 23 μg/m3. The statistics for the difference from Sentinel-2 level 2 data are presented in Table 5. The NIR band had the highest agreement with an RMSE of 0.012 when using raster AOD and an RMSE of 0.032 when using single-point AOD, which indicated that the NIR sensitivity to the AOD is not critical (Jung et al., 2020). The scatter plots for the 6S results and the Sentinel-2 level 2 data (Fig. 14) also show that they were closer to the 1:1 line when raster AOD. The histogram also includes a few outliers when using single-point AOD (Fig. 15). Fig. 16 to 17 are the maps for the surface reflectance, and Fig. 16 shows the difference calculated by subtracting the single-point AOD result from the raster AOD result. The difference was greatest in the forests and waters for all bands. Many pixels created by the atmospheric correction using the single-point AOD had inappropriate values for the forests and waters. While the NDVI created using raster AOD agreed with the land cover map (Fig. 13), the NDVI from single-point AOD seemed unnatural (Fig. 19). Usually, the waterbody NDVI is close to 0. The forest NDVI in spring is relatively high but does not reach 1. Previous studies also pointed out that the correction of aerosol scattering is crucial to the NDVI value (Goward et al., 1991).
Table 5. Differences between the 6S result and Sentinel-2 level 2 data according to AOD products (March 4, 2019)
Fig. 12. Distribution of raster AOD values (March 4, 2019).
Fig. 13. Land cover map for the Sentinel-2 image of March 4, 2019.
Fig. 14. Scatter plots for the 6S result and Sentinel-2 level 2 data according to AOD products (March 4, 2019).
Fig. 15. Histogram of the band reflectance according to AOD products (March 4, 2019).
Fig. 16. Surface reflectance of visible bands according to AOD products (March 4, 2019).
Fig. 17. Surface reflectance of red edge and NIR bands according to AOD products (March 4, 2019).
Fig. 18. Reflectance difference between the result using raster AOD and the result using single-point AOD for atmospheric correction (March 4, 2019).
Fig. 19. NDVI calculated by the red and NIR reflectance using (a) raster AOD and (b) single-point AOD (March 4, 2019).
3) Comparisons with drone images
To make sure the result of the 6S atmospheric correction, we also conducted comparisons with drone images provided by Chonnam National University. The drone images were taken at the altitudes of 100~300 m, and the atmospheric correction can be ignored (Yu and Kim, 2019). The atmospheric correction using raster AOD was performed for Sentinel-2 images, and they were matched to the drone images within ±5 days. For the target area of the paddy field, the drone images were upscaled to 10 m resolution to compare with Sentinel-2 images. The red and NIR reflectance from drone images and the 6S result were first compared(Fig. 20 to 24). In general, the accuracy of the red band is lower than the NIR band, which can lead to some disagreement of NDVI. However, the upscaled drone NDVI were overall agreed with the Sentinel-2 NDVI after 6S correction (Table 6) (Fig. 25 to 29), although Sentinel-2 NDVI had relatively low values. The spatial distribution of both images seemed very similar. Fig. 21 shows the season of the rice harvest, with abrupt changes in NDVI value for a few days.
Fig. 20. Comparisons of a drone image (August 14, 2020) and the 6S result (August 15, 2020).
Fig. 21. Comparisons of a drone image (September 24, 2020) and the 6S result (September 29, 2020).
Fig. 22. Comparisons of a drone image (July 12, 2021) and the 6S result (July 16, 2021).
Fig. 23. Comparisons of a drone image (July 21, 2021) and the 6S result (July 21, 2021).
Fig. 24. Comparisons of a drone image (September 28, 2021) and the 6S result (September 24, 2021).
Table 6. Differences of the NDVI values from the drone and Sentinel-2 images
Fig. 25. NDVI maps created by a drone image (August 14, 2020) and the 6S result (August 15, 2020).
Fig. 26. NDVI maps created by a drone image (September 24, 2020) and the 6S result (September 29, 2020).
Fig. 27. NDVI maps created by a drone image (July 12, 2021) and the 6S result (July 16, 2021).
Fig. 28. NDVI maps created by a drone image (July 21, 2021) and the 6S result (July 21, 2021).
Fig. 29. NDVI maps created by a drone image (September 28, 2021) and the 6S result (September 24, 2021).
5. Conclusions
In this study, we examined the atmospheric correction results for the Sentinel-2 images in South Korea by comparisons of using raster AOD (MODIS) and single point AOD (AERONET). The 6S result was overall agreed with the Sentinel-2 level 2 data, and using raster AOD showed better performance than using single point AOD. The differences in forests and waters were greatest for all bands. The atmospheric correction using the single-point AOD yielded some inappropriate values for the forests and waters, while the atmospheric correction using raster AOD much agreed with the land cover map. The Sentinel-2 NDVI after 6S correction were overall agreed with the upscaled drone NDVI, although Sentinel-2 NDVI had relatively low values. Also, the spatial distribution of both images seemed very similar for growing and harvest seasons. Future work will be necessary to make efforts for the gap filling of AOD data and an accurate bi-directional reflectance distribution function (BRDF) model for high-resolution atmospheric correction.
References
- Goward, S.N., B. Markham, D.G. Dye, W. Dulaney, and J. Yang, 1991. Normalized difference vegetation index measurements from the Advanced Very High Resolution Radiometer, Remote Sensing of Environment, 35(2-3): 257-277. https://doi.org/10.1016/0034-4257(91)90017-Z
- Jung, D., D. Jin, N.H. Seong, K.S. Lee, M. Seo, S. Choi, S. Sim, K.S. Han, and B.R. Kim, 2020. Sensitivity Analysis of Surface Reflectance Retrieved from 6SV LUT for Each Channel of KOMPSAT-3/3A, Korean Journal of Remote Sensing, 36(5-1): 785-791 (in Korean with English abstract). https://doi.org/10.7780/KJRS.2020.36.5.1.11
- Kaufman, Y.J., 1987. Atmospheric effect on spectral Signature-measurements, Advances in Space Research, 7(11): 203-206. https://doi.org/10.1016/0273-1177(87)90313-9
- Kim, S., Y. Jeong, Y. Youn, S. Cho, J. Kang, G. Kim, and Y. Lee, 2021. A Comparison between multiple satellite AOD products using AERONET sun photometer observations in South Korea: Case study of MODIS, VIIRS, Himawari-8, and Sentinel-3, Korean Journal of Remote Sensing, 37(3): 543-557 (in Korean with English abstract). https://doi.org/10.7780/KJRS.2021.37.3.14
- Kneizys, F.X., L.W. Abreu, G.P. Anderson, J.H. Chetwynd, E.P. Shettle, A. Berk, L.S. Bernstein, D.C. Robertson, P. Acharya, L.S. Rothman, J.E.A. Selby, W.O. Gallery, and S.A. Clough, 1996. The MODTRAN2/3 Report and LOWTRAN7 MODEL, Phillips Laboratory, Hanscom AFB, Bedford, MA, USA.
- Kotchenova, S.Y., E.F. Vermote, R. Matarrese, and F.J. Klemm Jr, 2006. Validation of a vector version Of the 6S radiative transfer code for atmospheric correction of satellite data, Part I: Path radiance, Applied Optics, 45(26): 6762-6774. https://doi.org/10.1364/AO.45.006762
- Lee, K.H. and J.M. Yum, 2019. A Review on Atmospheric Correction Technique Using Satellite Remote Sensing, Korean Journal of Remote Sensing, 35(6-1): 1011-1030 (in Korean with English abstract). https://doi.org/10.7780/KJRS.2019.35.6.1.11
- Lee, K.H. and Y.J. Kim, 2008. Sensitivity of COMS/GOCI measured top-of-atmosphere reflectances to atmospheric aerosol properties, Korean Journal of Remote Sensing, 24(6): 559-569 (in Korean with English abstract). https://doi.org/10.7780/KJRS.2008.24.6.559
- Lee, K.S., 2019. Atmospheric correction issues of optical imagery in land remote sensing, Korean Journal of Remote Sensing, 35(6-3): 1299-1312 (in Korean with English abstract). https://doi.org/10.7780/KJRS.2019.35.6.3.12
- Lee, K.S., C.S. Lee, M. Seo, S. Choi, N.H. Seong, D. Jin, J.M. Yeom, and K.S. Han, 2020. Improvements of 6S Look-Up-Table Based Surface Reflectance Employing Minimum Curvature Surface Method, Asia-Pacific Journal of Atmospheric Sciences, 56(2): 235-248. https://doi.org/10.1007/s13143-019-00164-3
- Li, Y., J. Chen, Q. Ma, H.K. Zhang, and J. Liu, 2018. Evaluation of Sentinel-2A surface reflectance derived using Sen2Cor in North America, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 11(6): 1997-2021. https://doi.org/10.1109/jstars.2018.2835823
- Lyapustin, A., J. Martonchik, Y. Wang, I. Laszlo, and S. Korkin, 2011a. Multiangle implementation of atmospheric correction (MAIAC): 1. Radiative transfer basis and look-up tables, Journal of Geophysical Research: Atmospheres, 116(D3).
- Lyapustin, A., Y. Wang, I. Laszlo, R. Kahn, S. Korkin, L. Remer, R. Levy, and J.S. Reid, 2011b. Multiangle implementation of atmospheric correction (MAIAC): 2. Aerosol algorithm, Journal of Geophysical Research: Atmospheres, 116(D3).
- Lyapustin, A., Y. Wang, S. Korkin, and D. Huang, 2018. MODIS collection 6 MAIAC algorithm, Atmospheric Measurement Techniques, 11(10): 5741-5765. https://doi.org/10.5194/amt-11-5741-2018
- Martins, V.S., C.C.F. Barbosa, L.A.S. De Carvalho, D.S.F. Jorge, F.D.L. Lobo, and E.M.L.D.M. Novo, 2017. Assessment of atmospheric correction methods for Sentinel-2 MSI images applied to Amazon floodplain lakes, Remote Sensing, 9(4): 322. https://doi.org/10.3390/rs9040322
- Nazeer, M., J.E. Nichol, and Y.K. Yung, 2014. Evaluation of atmospheric correction models and Landsat surface reflectance product in an urban coastal environment, International Journal of Remote Sensing, 35(16): 6271-6291. https://doi.org/10.1080/01431161.2014.951742
- Proud, S.R., M.O. Rasmussen, R. Fensholt, I. Sandholt, C. Shisanya, W. Mutero, C. Mbow, and A. Anyamba, 2010. Improving the SMAC atmospheric correction code by analysis of Meteosat Second Generation NDVI and surface reflectance data, Remote Sensing of Environment, 114(8): 1687-1698. https://doi.org/10.1016/j.rse.2010.02.020
- Rothman, L.S.,I.E. Gordon, Y. Babikov, A. Barbe, D.C. Benner, P.F. Bernath, M. Birk, L. Bizzocchi, V. Boudon, L.R. Brown, A. Campargue, K. Chance, E.A. Cohen, L.H. Coudert, V.M. Devi, B.J. Drouin, A. Fayt, J.-M. Flaud, R.R. Gamache, J.J. Harrison, J.-M. Hartmann, C. Hill, J.T. Hodges, D. Jacquemart, A. Jolly, J. Lamouroux, R.J. Le Roy, G. Li, D.A. Long, O.M. Lyulin, C.J. Mackie, S.T. Massie, S. Mikhailenko, H.S.P. Muller, O.V. Naumenko, A.V. Nikitin, J. Orphal, V. Perevalov, A. Perrin, E.R. Polovtseva, C. Richard, M.A.H. Smith, E. Starikova, K. Sung, S. Tashkun, J. Tennyson, G.C. Toon, Vl. G. Tyuterev, and G. Wagner, 2013. The HITRAN2012 molecular spectroscopic database, Journal of Quantitative Spectroscopy and Radiative Transfer, 130: 4-50. https://doi.org/10.1016/j.jqsrt.2013.07.002
- Vermote, E.F., D. Tanre, J.L. Deuze, M. Herman, and J.J. Morcette, 1997. Second simulation of the satellite signal in the solar spectrum, 6S: An overview, IEEE Transactions on Geoscience and Remote Sensing, 35(3): 675-686. https://doi.org/10.1109/36.581987
- Wilson, R.T., 2013. Py6S: A Python interface to the 6S radiative transfer model, Computers & Geosciences, 51(2): 166. https://doi.org/10.1016/j.cageo.2012.08.002
- Xie, Y., X. Zhao, L. Li, and H. Wang, 2010. Calculating NDVI for Landsat7-ETM data after atmospheric correction using 6S model: A case study in Zhangye city, China, Proc. of 2010 18th International Conference on Geoinformatics, Beijing, China, Jun. 18-20, pp. 1-4.
- Yu, H.J. and D.S. Kim, 2019. Necessity on the Introduction of UAV and Hyperspectral Image for Fluvial Remote Sensing, Water for Future, 52(5): 42-54.
- Yusuf, F.R., K.B. Santoso, M.U.L. Ningam, M. Kamal, and P. Wicaksono, 2018. Evaluation of atmospheric correction models and Landsat surface reflectance product in Daerah Istimewa Yogyakarta, Indonesia, IOP Conference Series: Earth and Environmental Science, 169(1): 012004.