Inter-Calibration of Satellite Passive Microwave Land Observations from AMSR-E and AMSR2 Using Overlapping FY3B-MWRI Sensor Measurements

2.1. Instrument Similarities and Differences

AMSR-E on NASA’s EOS Aqua satellite was launched 4 May 2002 and operated until 4 October 2011. As the successor to AMSR-E, AMSR2 on-board the sun-synchronous satellite GCOM-W1 was launched on 18 May 2012 by JAXA. GCOM-W1 is the first of three planned satellites of the GCOM-W project, which is a 13-year mission designed for global and continuous observation of Earth’s water and energy cycles. AMSR2 has most of the characteristics of AMSR-E [31]. Both sensors are conical scanning passive microwave radiometers with similar instrument configurations. The AMSR instruments measure microwave emissions from the Earth surface twice daily for descending/ascending orbital equatorial crossings at 1:30 AM/PM local time, with vertically (V) and horizontally (H) polarized T_b retrievals at six frequencies (6.9, 10.7, 18.7, 23.8, 36.5, 89.0 GHz). Major changes in AMSR2, distinctive from AMSR-E, include an additional frequency at 7.3 GHz designed for mitigating Radio Frequency Interference (RFI), and a larger (2.0 m diameter) main reflector for enhanced spatial resolution. The MWRI onboard the FY-3B satellite was launched 5 November 2010. Similar to the channel setting and view geometry of AMSR-E, MWRI conically scans the Earth at five frequencies (10.7, 18.7, 23.8, 36.5, 89.0 GHz) and dual (H, V) polarizations. The MWRI has four major differences from the AMSR instruments, including: (a) no C-band channel relative to the AMSR 10.7 GHz channels due to widespread C-band RFI; (b) coarser spatial resolution and slightly narrower orbital swath for all frequencies relative to the AMSR sensors; (c) an approximate 10-min satellite overpass time difference between MWRI and AMSR-E; and (d) an approximate 53 degree earth incident angle (EIA) of measurement instead of 55 degrees for the AMSR sensors. Detailed sensor configurations for AMSR-E, AMSR2 and MWRI are summarized in Table 1.

2.2. Ancillary Datasets and Study Sites

Four ancillary datasets were employed for calibrating the AMSR2 T_b record to the AMSR-E baseline. The datasets used for calibration include: (a) AMSR-E L2A Version 3 (V003) re-sampled swath T_b data, which is the most recent version with complete recalibration to the Remote Sensing Systems (RSS) Version 7 standard; (b) AMSR2 L1R re-sampled swath T_b records (version 1.1); (c) the MWRI Level 1 swath T_b record; (d) daily minimum and maximum air temperature (T_mn and T_mx) measurements from global WMO weather stations that were used for independent validation of AMSR daily air temperature estimates derived from the calibrated T_b results.

All of the satellite ascending and descending orbital swath T_b data were spatially re-sampled and re-projected to a consistent 25-km resolution global Equal Area Scalable Earth (EASE) Grid (version 1) following previously established methods [32,33]. Based on the gridded data, four-month daily T_b records of spatially and temporally collocated observations for MWRI and AMSR-E (1 June 2011 to 30 September 2011), and MWRI and AMSR2 (1 June 2013 to 30 September 2013) were extracted over all global land areas. The two sets of matched datasets were then used to evaluate sensor T_b differences and develop pixel-wise empirical relationships for developing a consistent calibrated AMSR T_b record. For the sensor inter-calibration evaluation, collocated T_b observations of the three sensors involving both calibrated and uncalibrated results were extracted over five selected relatively homogeneous test regions, each approximately 1° × 1° in size, and representing the major Köppen-Geiger global climate classification zones (Polar, Continental, Tropical, Dry, and Temperate Climate types). Among the regions, the Tropical test area in Amazonia (Latitude: −6.35° to −7.14°, Longitude: −64.82° to −65.85°) was previously used for inter-sensor comparisons between AMSR-E and WindSat [34]. The T_b calibration was also assessed within major global land cover areas defined from a 0.05 degree resolution MODIS (MCD12C1) IGBP (International Geosphere-Biosphere Programme) land cover classification [35].

Surface (∼2 m height) air temperature estimates were also derived from both calibrated and uncalibrated T_b AM and PM orbital crossings over a three-year composited sensor record. The daily temperatures were derived for all global land areas using multi-frequency, ascending and descending orbit T_b inputs from the composited sensor record and temperature estimation algorithms previously developed for AMSR-E [7]. The resulting air temperatures determined from the AMSR-E baseline and calibrated AMSR2 T_b inputs were then compared with coincident in-situ daily air temperature measurements from 142 globally distributed WMO weather stations. Daily air temperatures determined from descending (∼1:30 a.m.) and ascending (∼1:30 p.m.) orbit AMSR T_b inputs were compared against collocated T_mn and T_mx station measurements based on the assumption that the satellite overpasses are approximately coincident with the respective diurnal temperature minima and maxima. These comparisons were used for evaluating potential differences in temperature estimation accuracy between the AMSR-E baseline and calibrated AMSR2 records, and as a relative quality assessment of the AMSR2 T_b calibration.

The composited three-year daily T_b record used for air temperature estimation was derived from one-year of AMSR-E data (1 September 2010–31 August 2011), one-year of calibrated AMSR2 data (1 September 2012–31 August 2013) and one-year of calibrated MWRI data (1 September 2011–31 August 2012). Both the MWRI and AMSR2 T_b records were calibrated to the AMSR-E baseline using the pixel-wise DD approach and collocated MWRI and AMSR T_b observations for all global land areas. The WMO air temperature datasets supporting the three-year comparisons were obtained from the National Climate Data Center (NCDC) Global Summary of the Day (GSOD version 7). Similar to previous work [7], the WMO validation stations were selected within major global land cover classes determined from the MODIS IGBP land cover classification [35], while the number of WMO stations selected within each land cover class was proportional to its global land area representation. The locations of the 142 global WMO temperature validation sites and the five T_b calibration test regions representing the major global climate zones are presented in Figure 1 and the corresponding inset map, respectively.

3.1. Double Differencing Method

As introduced in Section 1, Double Differencing (DD) is a straightforward and effective method for sensor inter-comparison and inter-calibration. By utilizing a third instrument as an intermediate reference, the DD method makes direct comparisons between target and reference sensors while avoiding potential uncertainty introduced by using a RTM, and mitigating negative impacts from systematic inconsistencies between the intermediate reference and inter-calibrated sensor records. The mathematical form of DD approach used in this study is expressed as:

3.2. Inter-Calibration of AMSR2 and AMSR-E

After describing observation biases between AMSR-E and AMSR2 using the above DD method, we then employed a linear calibration adjustment of the AMSR2 observations against the AMSR-E baseline. The linear relationships between collocated MWRI and AMSR T_b observations for the well-correlated pixels were described as:

For pixels where the T_b observations between AMSR and MWRI were well correlated (R > 0.95), coefficients, a and b, were calculated directly based on the pixel-wise linear regression results; otherwise an Inverse Distance Weighting (IDW) method was used for spatial interpolation estimation of a and b from nearby pixels having the same land cover type and successfully calculated parameter (a and b) values. The pixel-wise regression and IDW method are intended to account for possible spatial variations of sensor biases, which are illustrated in the following Subsection 4.1.

3.3. Air Temperature Estimation Method

Surface air temperature is an essential climate variable for determining the state of the climate system and monitoring environmental change. It also plays a critical role in regulating environmental processes and land–atmosphere interactions governing surface water, energy and carbon budgets [37,38]. In [7], T_mn and T_mx over the Northern Hemisphere were estimated based on an iterative method using multi-frequency T_b records from AMSR-E. After screening out potential precipitation, Radio Frequency Interference (RFI), snow and frozen conditions based on the procedure by [7], three T_b ratios from the AMSR-E 18.7 and 23.8 GHz channels are first employed iteratively to determine the optical thickness of atmospheric water vapor, land surface open water fraction and vegetation transmissivity. This information is then used to solve for the effective surface temperature, T_s, which was found to be well correlated with surface air temperature. By comparing with daily measurements from WMO weather stations, the resulting T_s retrievals for respective morning (AM) and afternoon (PM) overpasses were empirically related to T_mn and T_mx in the similar empirical regression form as Equations (5) and (6). The algorithm was developed for the Version 2 AMSR-E/Aqua L2A dataset and showed favorable accuracy (R² ≥ 0.79, RMSE = 3.5 K) against daily temperature measurements from independent weather stations representing a broad range of climate and land cover conditions [7].

In this study, we applied the same algorithm as [7] but with regression coefficients adjusted to reflect reprocessing updates of the AMSR-E T_b record from Remote Sensing Systems (RSS) Version 6 to Version 7 as:

4.1. Inter-Calibration between AMSR2, AMSR-E and MWRI

Stable and well-correlated T_b observations from AMSR-E, AMSR2 and MWRI form the basis of the sensor inter-calibrations. The channel specific T_b correlations between the AMSR and MWRI sensors were thus examined first for the two, four-month overlapping data records. The linear correlation coefficient R was calculated from the overlapping sensor records for each 25 km resolution pixel within the global land domain, while the mean T_b sample size used to determine R for each pixel was 50.4 ± 29.6 days for the AMSR-E and MWRI overlapping data record and 49.2 ± 29.5 days for the AMSR2 and MWRI overlapping record. Generally favorable T_b correlations were observed over the global domain from the overlapping MWRI and AMSR records for all analyzed frequencies, orbits and polarizations, though the locations and global extent of significant correlations varied according to the different channel frequencies (e.g., Tables 2 and 3). As an illustrated example, correlation maps of collocated H-Polarized T_b ascending observations for the intermediate 23 GHz channel from the overlapping MWRI and AMSR records are presented in Figure 2. The global patterns of 23 GHz results are similar to the other frequency results, while the 23 GHz channel is also important for the subsequent air temperature estimation. More detailed summaries representing all frequencies are listed in Tables 2 and 3. The 23 GHz results from the MWRI and AMSR sensor records are highly correlated (R > 0.95) over more than half of the global domain. Regions with lower T_b correlations (R < 0.85) were located near large water bodies, including coastal areas, the Great Lakes and Amazon River. This is expected due to the variable spatial sampling and coverage of the collocated elliptical sensor footprints, which reflect satellite orbital differences and discrepancies in sensor spatial resolutions and geolocation accuracy. In particular, the inclusion of different fractional water coverage in the collocated footprints can lead to significant differences in the T_b observations due to the large dielectric contrast between water and land features. To minimize the effect of these uncertainties on the sensor T_b inter-comparison and calibration, only highly correlated pixels (R > 0.95) were selected for further evaluation of sensor observation biases. The overall T_b bias of MWRI against AMSR2 or AMSR-E was directly calculated by the SD method while the bias of AMSR2 against AMSR-E was indirectly estimated by the DD method using Equation (1). The resulting global distribution of estimated AMSR2 bias in relation to the baseline AMSR-E record is illustrated in Figure 3a,b for respective 23 GHz and 18 GHz, H-Polarized ascending T_b records. The AMSR2 ascending T_b record at 23 GHz and 18 GHz is approximately 2.10 K and 2.46 K higher than the AMSR-E record over the global domain. Higher T_b bias was generally found in regions of colder radiance, such as the Polar regions and Sahara Desert, where lower T_b observations are caused by low surface temperature and/or low H-pol emissivity. This finding is consistent with the previous study from JAXA that showed much larger AMSR2 biases toward the cold end of the T_b range [28,29].

The complete T_b inter-comparison results for the three sensor records and all matching frequencies are summarized in Table 2 for ascending orbits and Table 3 for descending orbits. The statistical metrics for the overlapping data records include the mean T_b difference or bias, the standard deviation (STD) of the T_b difference and the percentage of the global land area with R > 0.95. By referring to AMSR-E as the standard or baseline, it can be seen from Tables 2 and 3 that AMSR2 tends to generate higher T_b readings for all channels, while MWRI generally produces lower T_b values. Meanwhile, MWRI and AMSR observations are highly correlated (R > 0.95) over approximately 27%–69% of the global land area, depending on T_b frequency and polarization. As discussed in the previous paragraph, variable spatial sampling and coverage of the collocated elliptical sensor footprints may lead to low correlation of the observations. The cross-sensor T_b correlation is also lowest for the 89 GHz channel as indicated from Tables 2 and 3. As introduced in Subsection 2.1, there is an approximate 10-min overpass time difference between AMSR and MWRI and changing atmosphere conditions may lead to a lower correlation. The calibration correction of AMSR2 T_b to the AMSR-E baseline was conducted individually for each channel frequency, polarization and orbit. Based on Tables 2 and 3, global mean biases between AMSR2 and AMSR-E observations can be estimated for each sensor channel. As is seen in Figure 4, the AMSR2 observations are systematically deviated from the AMSR-E baseline by approximately 0.15 K–3.06 K, depending on the T_b frequency, polarization and ascending/descending orbits.

Based on the observed strong linear correlations between the sensor observations, empirical linear relationships were derived between the AMSR T_b observations using Equations (2)–(4) to calibrate the AMSR2 record to the AMSR-E baseline. Based on the MWRI matched datasets, the intercept and slope of the relationship between AMSR2 and AMSR-E were directly calculated in Equations (2)–(4) for well correlated land pixels (R > 0.95), and spatially interpolated over the remaining land areas using the IDW method. The calculation was carried out for each channel on a per pixel basis. Example T_b slope and intercept distribution maps describing relations between the AMSR2 and AMSR-E records is given in Figure 5 for the 23 GHz, H-Polarized ascending T_b observations. Assuming the relationships between the sensors are stable over the observation record, the empirical slope and intercept maps provide effective ancillary inputs for calibrating the AMSR2 and MWRI T_b records to the AMSR-E baseline. Using this approach and the overlapping sensor T_b observations, the available AMSR2 (July 2012–September 2013) and MWRI (June 2011–September 2013) records were linearly calibrated to the AMSR-E (June 2002–October 2011) T_b record. The resulting integrated T_b observations from the three sensors spanned a continuous, calibrated daily observation record extending from June 2002–September 2013, and consistent with the AMSR-E baseline.

4.2. Evaluation of the Sensor Inter-Calibration Results

The performance of the sensor T_b inter-calibration results was examined for five, homogeneous 1° × 1° T_b calibration test regions representing major Köppen-Geiger global climate classification zones, including Polar, Continental, Tropical, Dry, and Temperate climate types (Figure 1). Spatial mean T_b values were calculated for each test region. To ensure relatively homogenous conditions suitable for inter-calibration, the test regions were selected on the basis of having standard deviations of vertical polarization T_b values not exceeding 2 K, and standard deviations of horizontal polarization T_b values not exceeding 3 K [39]. The T_b calibration test results representing all frequencies and polarizations are summarized Tables 4 to 5. These results indicate that the T_b inter-calibration method significantly mitigates sensor observation biases for all channels over the broad range of global climate conditions represented by the five test regions.

The selected tropical T_b calibration test area in central Amazonia (Figure 1) was identified as relatively homogenous at microwave frequencies and suitable for sensor inter-comparisons from a previous study [34]. Both un-calibrated and calibrated AMSR2 and MWRI T_b (H-Polarization, and Ascending orbit) values for the 23 GHz, 18 GHz and 10 GHz channels are compared with collocated AMSR-E baseline observations for this region in Figure 6. For the 23 GHz example, there are 59 point pairs evaluated in both Figure 6a,b. These results indicate that the AMSR-E and AMSR2 T_b values (Y-axis) as a whole are well correlated with the MWRI data (R = 0.90) but also show large T_b differences of approximately −3.14 K (Bias) and 3.31 K (RMSE). The resulting empirical calibration of AMSR2 and MWRI observations to the AMSR-E baseline described in Section 3 produced generally consistent T_b observations across the three sensor records (Figure 6b). The resulting calibrated AMSR2 and MWRI T_b values for the reference area correspond favorably with the AMSR-E baseline (R = 0.99, RMSE = 0.42 K, Bias = 0.04 K), and are consistent with the inter-calibration results from the other climate test regions. Similarly, instrument biases were corrected for the other frequencies. As illustrated in Figure 6c–f, R increased from 0.82–0.98 for 18 GHz and 0.76–0.98 for 10 GHz, respectively; bias also decreased from −0.45 K to −0.10 K for 18 GHz and −3.60 K to 0.24 K for 10 GHz, respectively. Note that T_b observations from AMSR sensors are generally higher than the uncalibrated MWRI observations except for 18 GHz channels where slightly lower AMSR-E T_b values were found as shown in Figure 6c and also Tables 2 and 3.

The global T_b calibration results were also summarized for the major global land cover classes according to T_b ascending (Table 6) and descending (Table 7) orbits. The statistical parameters presented include the mean bias, RMSE and correlation coefficient which are averaged for all channels for the uncalibrated and calibrated results within each land cover class. Similar to climate based T_b calibration test, the land cover assessment results also indicate that the inter-calibration process mitigated most observation biases. Further analysis of residual errors from the resulting calibration showed no apparent non-linear bias or other artifacts in the integrated T_b record. However, the RMSE differences were only partially reduced and the correlations between datasets were slightly improved. The T_b differences from AMSR and MWRI are not only caused by different instrument calibrations, but also by other factors including different sensor spatial coverage, temporal sampling and random uncertainty in the observations. With observation biases mitigated, the calibrated AMSR2 and MWRI T_b values are similar to the AMSR-E baseline, though further integration refinement of the sensor records would likely benefit by more explicit accounting of the different sensor spatial and temporal characteristics.

4.3. Air Temperature Estimation Results

The global integrated sensor T_b record was used as a daily ancillary input for estimating surface air temperatures using algorithms Equations (5) and (6) adapted from AMSR-E [7]. Daily T_mx and T_mn estimates were derived from respective ascending (PM) and descending (AM) orbit T_b inputs and used for further assessment of the integrated sensor record. The resulting satellite temperature (T_mx and T_mn) estimates were validated against coincident in-situ air temperature measurements from 142 independent weather stations encompassing a broad range of global climate and land cover conditions. The air temperature estimates were extracted for three, one-year periods representing baseline AMSR-E and calibrated MWRI and AMSR2 portions of the integrated T_b record used for the validation assessment.

As introduced in Section 3.3, the air temperature estimates directly depend on the T_b observations at 18 and 23 GHz channels; the temperature estimates also indirectly depend on other channels used to identify and screen out unfavorable retrieval conditions, including high fractional water inundation, active precipitation or frozen surface conditions [7]. Therefore, the air temperature estimation is sensitive to all T_b channels ranging from 10–89 GHz. We assumed that the overall quality of the in-situ measurements from the 142 WMO validation sites was consistent over the three sensor years, so that the resulting temperature validation reflects the quality and consistency of both the T_b calibration and temperature estimation from the different sensors. The comparisons between measured and estimated T_mx from the AMSR-E baseline and calibrated AMSR2 T_b records are shown in Figure 7 for the 142 validation sites. The global temperature estimates from the AMSR-E baseline show generally good correlation (R² = 0.86) and low error (RMSE = 3.37 °C) relative to the in situ station measurements (Figure 6a), and consistent with temperature validation results from a previous Northern Hemisphere study [7]. The temperature estimates from the calibrated AMSR2 T_b record were also favorable (R² = 0.83 and RMSE = 3.64 °C) but represented a small, but significant (p < 0.05) accuracy reduction from the AMSR-E baseline. The overall statistical results of the temperature validation comparisons are summarized in Table 8. Overall, the AMSR temperature estimates are well correlated with the in-situ measurements and show favorable accuracy (R² ≥ 0.75, RMSE ≤ 3.64 °C). These results also indicate that the calibrated AMSR2 data record can be used for extending the AMSR-E land parameter record with reasonable accuracy and consistency. However, the calibrated MWRI T_b record produced substantially lower temperature estimation accuracy relative to the station measurements (Table 8), with both lower temperature correspondence (R² = 0.80 for T_mx and R² = 0.69 for T_mn) and larger RMSE (4.12 °C for T_mx and 4.11 °C for T_mn) differences than the AMSR results. The lower MWRI temperature accuracy was primarily attributed to the coarser spatial resolution of the MWRI footprint relative to the AMSR sensors, and other factors described in the following section.

4.4. Discussion

The relative T_b bias from AMSR2 was detected, calibrated and mitigated against the AMSR-E baseline. However, slightly lower temperature estimation accuracy from the calibrated AMSR2 T_b record was also found. The lower AMSR2 temperature estimation accuracy may reflect a limited T_b calibration period represented by the available four-month matched datasets, which do not fully represent potential seasonal or annual changes in the overlapping sensor relationships. As indicated in Subsections 4.2 and 4.3, a number of factors may cause the observation differences between sensors, including different instrument calibrations, differences in spatial coverage (e.g., spatial resolution, geo-location of the footprint center, and footprint sampling), temporal sampling and random uncertainty in the observations. The DD method is capable of minimizing observation biases and partially mitigating the effects of the different sensor configurations on the T_b retrievals, though the residual impacts of these differences, as well as observation uncertainties inherent to each instrument cannot be completely eliminated using the relatively simple DD inter-calibration method.

The impacts of the inherent differences in sensor observations are also reflected from the lower MWRI performance. Coarser MWRI spatial footprints relative to the other sensors may be the primary factor contributing to lower sensor performance. For example, the 18 GHz channel footprints from AMSR-E and AMSR2 have respective mean spatial resolutions of 21 km and 18 km, whereas the corresponding MWRI resolution is 39 km, or roughly four times the AMSR footprints. Although the orbital swath observations from the sensors were gridded to a consistent 25 km resolution global EASE-grid, the resulting T_b records after bias correction by the DD method still preserve underlying differences in footprint resolution and sampling from the individual channel frequencies. Moreover, the resampling process itself inevitably causes data losses and its related uncertainties can also affect inter-calibration accuracy.

Future data records with longer overlapping periods and more sophisticated calibrations that consider the inherent differences in sensor configurations, especially different sensor spatial resolutions, would likely improve the accuracy of cross-sensor data integration efforts. Moreover, both MWRI and AMSR2 are relatively new sensors, while AMSR-E encompasses a much longer data record and has undergone multiple data reprocessing events, each with relative improvements over earlier versions. With continuing observations and similar data reprocessing and sensor calibration enhancements, both AMSR2 and MWRI are expected to show improved performance leading to more accurate integrated long-term environmental data records.