Sentinel-2 MSI Radiometric Characterization and Cross-Calibration with Landsat-8 OLI

Near-nadir observations by the Multispectral Instrument (MSI) onboard the Sentinel-2 and the Operational Land Imager (OLI) onboard Landsat 8 were collected during two Simultaneous Nadir Overpasses (SNO). Multispectral images with 10, 20, and 30 m resolution from a spatially uniform area in the Saharan desert were acquired for direct comparison of MSI and OLI TopOf-Atmosphere (TOA) reflectances. This paper presents an initial radiometric cross-calibration of the 8 corresponding spectral bands of the Sentinel-2 MSI and Landsat 8 OLI sensors. With the well-calibrated Landsat 8 OLI as a reference, the comparison indicates that 6 MSI bands are consistent with OLI within 3% in terms of spectral band adjustment factors Bi . The Near-Infra-Red (NIR) and cirrus bands are exceptions. They yield radiometric differences on the order of 8% and 15% respectively. Cross-calibration results show that the radiometric difference of the 7 corresponding bands are consistent to OLI within 1% or better, except on cirrus band. A pixel-by-pixel match between the MSI and OLI observations for different land covers showed that. This initial study suggests that the red-edge band B8A of MSI can be used to replace the NIR band B08 when conducting vegetation monitoring.


Introduction
The successful launch of the European Space Agency (ESA) Sentinel-2A on 23 June 2015 with the key instrument MultiSpectral Instrument (MSI) provides an important means to augment Earth-observation capabilities following the legacy of Landsat. After the three-month satellite commissioning campaign, the MSI onboard Sentinel-2 (S-2) is performing very well [1]. By 3 December 2015, the sensor data records have achieved provisional maturity status and have been accessed in level-1C Top-Of-Atmosphere (TOA) reflectance by the remote sensing community worldwide.
Sentinel-2 is an ESA land monitoring mission with two identical satellites that provide high resolution optical imagery. The launch of Sentinel-2B is planned for mid-2016 [2]. After that, the two identical Sentinel-2 satellites (Sentinel-2A and Sentinel-2B) will provide an exceptional revisit capability of 5 days at the equator and 2 -3 days over mid-latitudes. The twin satellites fly in the same sun-synchronous orbit but phased at 180˚ to each other. The coverage limits are between latitudes 56˚ south and 84˚ north, including all the land surfaces, coastal waters, and Mediterranean Sea. At a nominal equatorial altitude of 786 km, the swath width is 290 km. The wide swath width and high revisit time will support monitoring of changes to vegetation within the growing season. The MSI sensor aboard on Sentinel-2 capitalizes on the technology and the vast experience acquired with SPOT and Landsat over the past decades [3]. The S-2 MSI samples 13 spectral bands covering wavelengths from 0.4 to 2.2 um: four bands at 10 meters, six bands at 20 meters and three bands at 60 meters spatial resolution [4].
In order to meet the requirement for monitoring rapidly changing land processes (e.g. agriculture, wild fire, vegetation phenology, and extreme weather events), the scientists of the NASA Land-Cover and Land-Use Change (LCLUC) program and NASA Multi-source Land Imaging (MuSLI) program have been actively pursuing the synergy of Sentinel-2 and Landsat 8 (L8) data [5]. S-2 MSI and L8 Operational Land Imager (OLI) together make a potent source for higher-rate multispectral observation with global coverage and free and open access.
The biggest challenge of using remote sensing data from multiple sources, however, is inter-calibration across different instruments. This paper explores aspects of the radiometric cross-calibration of the S-2 MSI and L8 OLI sensors based on near-simultaneous imaging of common ground targets in the Saharan desert. Band adjustment factors and linear regression slopes for each band are derived from data at this site. Eight corresponding bands (including aerosol, four visible and near-Infrared bands, cirrus, and two shortwave infrared bands) of MSI are compared with that of OLI in scenes collected during Simultaneous Nadir Overpasses (SNO) and vicarious site.  [3] and summarized in Table 2.

Relative Spectral Response (RSR)
Given the similar mission concepts of the Spot and Landsat series sensors (MSS, TM, ETM+, and OLI), the spectral band configuration of the S-2 MSI was designed around the use of Landsat and Spot wavelengths [3]. The S-2 MSI has spectral bands similar to Landsat 8 OLI's, which enable the synergistic use of data from S-2A, S-2B, and Landsat OLI. Figure 1 Figure 1). Compared to the OLI vegetation bands, however, there are significant differences in RSR profiles between corresponding MSI and OLI in three visible bands (Blue, Green, Red) and NIR. The NIR band of MSI is much wider than OLI's ( Figure 1). The Lmin is the radiance corresponding to the minimum quantized and calibrated data digital number, which is typically "1" or "0" and Lmax is the radiance corresponding to the maximum quantized and calibrated data digital number typically "4095". b Lref is the reference radiances. c Signal to Noise Ratio (SNR) for the Lref defined for S-2 mission.
Nearly simultaneous observations over the Saharan desert were taken by S-2 MSI and L8 OLI on Dec. 8, 2015 (Table 3). The homogeneous desert area in  [7]. Sand dunes create variation over much of the scene, but there is a largely spatially homogeneous area over 2.4 by 4.2 km. In order to assess the radiometric characteristics of MSI vegetation bands, an additional MSI and L8 image pair from a forested region in Nigeria is also used in this study ( Table 3). The selected MSI and OLI image pairs were unaffected by clouds.  and greatly reduces the uncertainties of radiometric bias caused by the surface BRDF and radiation transfer.

Vicarious Calibration
Vicarious calibration makes use of natural or artificial sites on the Earth's surface for in-flight calibration of satellite sensors [8]- [14]. It has been used successfully for the absolute radiometric calibration of Landsat TM [15]. A reflectance-based approach is one of these vicarious methods. [13] suggested that TOA reflectance comparisons have the potential to yield the best possible calibration comparisons between two sensors with near simultaneous nadir data acquisitions, because, 1) the cosine effect of different solar zenith angles was removed and 2) proper compensation for the exo-atmospheric solar irradiance was supplied.
The key radiometric equations for cross-calibration of Landsat satellites have been developed and applied by [15] [16]. Following these equations, sensor MSI responsivity MSI i G in spectral band i is given by: Original definitions of the variables in Equation (2)  and 0 cos OLI i E θ are the exo-atmospheric solar irradiance (in Watts/(m 2 ·μm)), two θ are the solar zenith angles of MSI and OLI. Equation (2) is developed for the cross-calibration start from raw data (or level 0). (2) is the spectral band adjustment factor,

The main part of Equation
Its uncertainty is directly proportional to the uncertainty in the cross calibration [17].
In this study, bi-directional reflectance effects are not expected to be significant since the selected uniform areas have near-nadir viewing geometry and the sun-angle differences between the image pairs are small. In addition, small misregistration is not expected to have an impact on the result because of the degree of spatial homogeneiety of the test area.

Spectral Band Adjustment Factors
The spectral band adjustment factors i B were computed from pixels in the grid cells from the image pair over the desert site ( Figure 3). The spectral band difference effect less than 3%, except in NIR and cirrus bands. The i B value for NIR is on the order of 8%, which is mostly caused by the difference of spectral profile between MSI and OLI. The i B value for cirrus band is even larger, on the order of 15%, though the spectral profile of cirrus of MSI is perfectly matched with OLI's ( Figure 1 & Figure 3).
The TOA reflectance of the grid cells ( Figure 2) are used to explore the radiometric calibration factors for the corresponding bands between MSI and OLI.
The mean values across the cells were extracted and plotted to estimate the (2)). Figure 4 shows plot for the eight corresponding bands (MSI against OLI), and Table 4 lists the slope results. The scatter plot of cirrus band is placed at the bottom right corner of Figure 4 due to its scale issue. Figure 5 together with Table 4 present the derived slopes i M and their correlation coefficients. On the left panel of Figure 5, the measurements of MSI are almost always higher than that of OLI in the bands of aerosols, blue, red, and SWIR1, where the derived slope of SWIR1 is significantly lower than the others, only 0.497. The measurements of MSI in the cirrus and NIR band, however, are much lower than the OLI's. There is more scatter in the cirrus band, averaging approximately 0.377 in terms of slope and 0.413 in terms of R-squared. The slope values i M derived from the Saharan calibration site show that cross-calibration is successful in the aerosol, blue, green, and SWIR2 bands. Figure 4 & Figure 5, Table 4 also clearly shows the issue in the NIR band between the two sensors. Figure 6 presents the calibration results from Table 4 in terms of percentage of difference compared with Landsat OLI measurements during the SNO event.
The results in Figure 6 show that almost all the differences are well within the 1% range, except in the cirrus band where the difference after calibration is around 2.5%. The calibrated MSI measurement in SWIR2 is a little bit off the    In order to further assess the magnitude of the VNIR band difference effect between Sentinel-2 MSI and Landsat OLI, spatial statistics were computed across a variety of land covers: coastal water, urban, natural forest, and desert. Statistics of TOA reflectance of the two sensors are also extracted from the SNO image pair (Table 3). In general, the SNO scene imaged by the two sensors is assumed  to be with the same sun-angle and off-nadir viewing geometry. Figure 7 plots the TOA reflectance change over different land surface. The comparison for the three visible bands (Blue, Green, and Red) yield small differences with dark surface (e.g. coastal water and natural vegetation), while the spectral band difference effects are larger over bright surface objects (e.g. heavy developed areas and desert). NIR band, however, shows significant spectral differences over variety of surfaces, with except of coastal water (Figure 7). This abnormal of NIR domain may be caused by the differences in NIR RSR profiles between the two sensors ( Figure 1 & Figure 7).

Discussion
The four VNIR bands have wide applications in remote sensing and imaging spectroscopy [18]. The cross-calibration results show that the blue channel yields the best adjustment. The difference after cross-calibration is 0.082% compared to the reference blue channel of L8 OLI. The difference after cross-calibration in the Green, Red, and NIR are on the order of 0.4%. In comparison with the significant difference of the RSRs of NIR bands in MSI and OLI, the cross-calibration performance of the MSI NIR band is promising. The NIR band is well known to be critical to the biophysical factors of vegetation monitoring. Figure 8 shows a direct comparison of NIR bands of Sentinel-2 and Landsat. There are significant differences in relative spectral response profiles between Sentinel-2 MSI and Landsat sensors (ETM+ and OLI) in the NIR. The design of NIR band of Landsat 7 ETM+ followed the Landsat TM series, which spans 760 -900 nm (Figure 8). The NIR band of L8 OLI avoids heavy water vapor contamination, using a narrow spectrum 850 -880 nm. The RSR of the NIR band of MSI, however, is more complex than that of L8 OLI. Fortunately, S2 MSI provides four additional red-edge bands (B05, B06, B07, and B8A), though the resolution is 20 m. The RSR of MSI's red-edge band B8A is similar to OLI's NIR (Figure 8). Both narrow bands are centered at 865 nm.

NIR Red
Using alternative band combinations, we calculated Normalized Difference Vegetation Indices (NDVI) for L8 OLI and S-2 MSI (Equation (3)) from the Nigerian scene which is covered by natural forest with a relatively homogenous texture. Table 5 shows the performance of MSI B8A is much better than B08.
The NDVI calculated using the B8A and B04 combination shows a larger data range. Its mean value is close to the OLI's NDVI value. This result suggests that B8A of MSI removes heavy water vapor influence yet is still sensitive enough for vegetation detection.
This result raises concerns about the MSI NIR band in data harmonization between S-2 MSI and L8 OLI. It also suggests that using B8A is a good option when calculating vegetation indices (e.g. NDVI).

Conclusions
A first cross-calibration of S-2 MSI and L8 OLI has been presented in this study. Image pairs captured during an SNO event are used to perform the radiometric cross-calibration. Nearly coincident data acquisitions over common targets make it possible to use image data from the well-calibrated L8 OLI to calibrate  S-2 MSI in analogous spectral bands. Given that Landsat 8 OLI is well-understood radiometrically [5], cross-calibration between the L8 OLI and other multispectral land imaging sensors (e.g. Landsat MSS, TM, and ETM+) can be considered in future studies.
During the development of Landsat 8 and S-2A, the two agencies (ESA and NASA) had joined calibration scientists to ensure that S-2 MSI and Landsat 8 OLI data offer compatible data products, thereby bringing greater benefits to the remote sensing communities of Earth's land and coastal zones. The preliminary results from this study indicate that the overall performance of MSI is a promising addition to the longest operating Earth Observation mission (Landsat). It will significantly augment the Landsat legacy and future Landsat missions (eg. Landsat 9, 10 and beyond).