GOSAT (also known as IBUKI) was developed by the Japan Aerospace Exploration Agency (JAXA) and launched in January 2009. In fact, the first global satellite SIF observations came from GOSAT (Joiner et al., 2011; Frankenberg et al., 2011b). On board the satellite is the greenhouse gas observation sensor (Thermal And Near-infrared Sensor for carbon Observation – Fourier Transform Spectrometer – TANSO-FTS), which has a spectral resolution of 0.012 nm (0.2 cm-1). The sensor has the following four bands: 0.758–0.775, 1.56–1.72, 1.92–2.08, and 5.56–14.3 µm. It has a sun-synchronous, descending orbit with an overpass time of 13:00 LT at the Equator, a 3 d repeat cycle, and a circular footprint of ∼82 km2 per sounding (∼10 km diameter; Kuze et al., 2009).

OCO-2 is a NASA satellite that was launched in July 2014, and OCO-3 is a duplicate of the OCO-2 grating spectrometer attached to the Japanese Experiment Module Exposed Facility (JEM-EF) on the International Space Station (ISS) in May 2019 (Eldering et al., 2019). Each platform houses a three-channel grating spectrometer, with a spectral resolving power of λ/Δλ>17000 (Crisp et al., 2017; Eldering et al., 2019), centered around the following wavelengths: an oxygen A band at 0.765 µm and carbon dioxide bands at 1.61 and 2.06 µm. The swath widths are ∼10 km with eight measurements across-track. The spatial resolution at nadir is slightly different for OCO-2 and OCO-3, at about 1.3km×2.25 km and 1.6km×2.2 km (across × along track), respectively.

OCO-2 has a 98.8 min orbit with a 13:36 nodal crossing time and a 16 d ground-track repeat cycle (Crisp, 2015). The ISS has a precessing low-inclination orbit that allows OCO-3 to view Earth at absolute latitudes less than ∼52∘. The ISS orbits the Earth ∼15.5 times a day, and data acquisition is halted during ISS maintenance and docking; thus, overpass times, revisit periods, and data availability are relatively irregular. Validation of the OCO-2 SIF retrievals was conducted by Sun et al. (2017) by comparing OCO-2 SIF to coordinated airborne measurements using the Chlorophyll Fluorescence Imaging Spectrometer (Frankenberg et al., 2018).

GOSAT observation modes are described as observation mode 1 in the sunshine (OB1D), observation mode 2 in the sunshine (OB2D), and specific observation mode in the sunshine (SPOD). OB1D is the routine observation mode, whereas OB2D is a non-routine mode in which the thermal-infrared observation and pointing mechanism is stopped during low power supply. Over land, SPOD is a target observation mode designed to observe specific sites. The Thermal And Near-infrared Sensor for carbon Observation – Fourier Transform Spectrometer (TANSO-FTS) sensor has a setting for low, medium, and high gain. The medium gain data are recommended for scenes that are bright, such as deserts. Since the data used for SIF retrievals are filtered to exclude bright scenes due to deserts, ice, snow, and cloud cover, the high gain data are used for SIF retrievals.

Nadir, glint, target, and transition observation modes are common to each OCO platform. The OCO-2 target mode provides repeated spatial sampling of a given target, such as an emission source or tower site. Target mode data for OCO-2 are absent from the v10 SIF Lite files but will be included in the v11 update.

The OCO-3 target mode is a sequence of adjacent and partially overlapping segments that allow for increased spatial sampling. The target modes for both platforms provide over 103 soundings. OCO-3 has an additional observation mode using its pointing mirror assembly (PMA), which allows for snapshot area mapping (SAM) of targets of interest. SAMs are a series of scans of a target that are nearly adjacent and can cover an area of ∼80 km2 in about 2 min. The SAMs and their target locations, which include volcanoes, various vegetation land cover types, and point sources of fossil fuel emissions, can be viewed at https://ocov3.jpl.nasa.gov/sams/index.php (last access: 7 March 2022). Target- and SAM-mode scans are prioritized and scheduled several days in advance of an overpass of the ISS over the target (Taylor et al., 2020).

The target and SAM observation modes offer a unique, spatially resolved acquisition of a target during a single overpass at different sun sensor geometries, as solar illumination is relatively fixed during overpasses, and soundings are acquired over a range of viewing angles as the sensors pass over their targets. For SIF applications, these measurements can be averaged to obtain SIF estimates with a reduced standard error or be binned by sun sensor geometries to investigate the effect of observation geometry on the retrieved SIF values, as we demonstrate below.

Level 2 data are ungridded (vector) data that contain geophysical variables that are of interest and use to the broader scientific community and are at same spatial and temporal resolution of the Level 0 and Level 1 data. Level 0 data, which are data obtained as-is from the sensor (Level 0), to which and ancillary information, such as radiometric and geometric calibration coefficients and georeferencing parameters, are appended to Level 0 data to form (Level 1) data, such as radiometric and geometric calibration coefficients and georeferencing parameters. Level 3 products refer to gridded (raster) data, which can be found at https://climatesciences.jpl.nasa.gov/sif/download-data/level-3/ (last access: 7 March 2022).

The annual and monthly spatial distribution of the GOSAT and OCO Level 2 data for the globe and the continental U.S. are presented in Figs. 1 and 2 for visualization. These data are produced by the OCO-2 and OCO-3 projects at the Jet Propulsion Laboratory (Frankenberg et al., 2014), quality controlled by NASA's Making Earth System data records for Use in Research Environments (MEaSUREs) SIF team, and are publicly available on the NASA Goddard Earth Sciences Data and Information Services Center (GES DISC) website (https://disc.gsfc.nasa.gov/, last access: 7 March 2022). Recent efforts by the OCO and MEaSUREs team have focused on harmonizing the processing pipeline, attributes, and file structures of the GOSAT and OCO SIF products (Parazoo et al., 2019). Here, we present a first analysis of these harmonized products and demonstrate for the user community their key commonalities and differences.

The ungridded Level 2 SIF Lite data are provided in netCDF-4 format and contain information for each sounding from which a SIF retrieval was made. For each of the three satellite platforms, there is one file for each day in which there is at least one sounding, and each file contains information for all soundings acquired on that day, including all measurement modes (glint, nadir, and target). The SIF Lite files can be read by, but are not limited to, MATLAB, Python, R, and Julia, using their respective netCDF4 or HDF5 libraries. The filename convention is, using the filename “oco2_LtSIF_200201_20210129t071949z.nc4” as an example, platform (oco2), data product (LtSIF), date (YYMMDD), and file creation date (YYYYMMDD) and time (tHHMMSS). The SIF Lite netCDF global attributes, dimensions, variables, and variable groups are described below and listed in Table 1.

The Quality_Flag variable indicates the quality of the data for each sounding as being best (0), good (1), or failed (2). We recommend using a combination of best and good for scientific analysis. The criteria for the best and good quality flags are listed in Table 2, and soundings that do not meet either set of criteria are flagged as having failed. The rationale for the criterion is as follows: reduced chi-squared (χ2) thresholds exclude fits that do not represent the spectrum well, continuum-level radiance excludes scenes with brightness that is too high or low, solar zenith angle (θ) excludes retrievals with extreme solar zenith angles, which are more likely affected by rotational Raman scattering, and the O2 and CO2 thresholds exclude most cloudy scenes.

The SIF values provided in the SIF Lite files are based on spectral fits covering Fraunhofer lines, as SIF reduces the fractional depth of the Fraunhofer lines (Plascyk, 1975). The SIF retrieval methodologies are fully explained by Frankenberg et al. (2011b, a) and SIF is retrieved using the identical method for GOSAT and the OCO platforms at 757 and 771 nm. In brief, the main retrieval quantity in the retrieval state vector is the fractional contribution of SIF to the continuum level radiance or relative fluorescence (SIF_Relative_757nm and SIF_Relative_771nm). The absolute SIF values (SIF_757nm and SIF_771nm) are generated during post-processing in W m-2 sr-1 µm-1.

It is important to note that although the SIF values have traditionally been loosely labeled as being retrieved at 757 and 771 nm, the retrieval fit windows used to produce the SIF Lite data are centered at 758.7 and 770.1 for OCO-2 and OCO-3 and at 758 and 771 for GOSAT. However, we retain the 757 and 771 nomenclature to remain consistent with previous publications and to avoid confusion. We estimated SIF at 740 nm for each sounding using both retrieval windows as described in more detail below.

The determination of single sounding retrieval uncertainty is covered in great detail by Sun et al. (2018) and Frankenberg et al. (2014) and is provided in the SIF Lite files as SIF_Uncertainty_740nm, SIF_Uncertainty_757nm, and SIF_Uncertainty_771nm. Briefly, these values are the 1 sigma (σ) estimated single sounding measurement precision and represent the random component of the retrieval errors. It is derived through standard least-square fitting by evaluating the error covariance matrix as follows: 1Se=(KTS0K)-1, where K is the Jacobian matrix of the least-squares fit, and S0 is the measurement error covariance matrix which characterizes the instrument noise per detector pixel.

For the OCO-2/-3 data, the uncertainty for SIF757 usually ranges between 0.3 and 0.5 W m-2 sr-1 µm-1 or ∼15 %–50 % of the absolute SIF value. Uncertainties for SIF771 are slightly higher due to less fluorescence and a relatively lower reduction in the fractional depth of the radiance at 771 nm. Uncertainty for SIF740 is calculated from using the general formula for error propagation and the partial derivatives for the uncertainties for SIF757 and SIF771 as follows: 2SIFUncertainty_740=0.5⋅((1.5⋅SIFUncertainty_757)2+(2.25⋅SIFUncertainty_771)2)‾.

The spectral window in which SIF retrievals are made depends on the wavelength bands of the platform. Assuming the spectral shape of SIF is known and invariant, one can convert SIF to a standard reference wavelength. Here, we use 740 nm as a reference as it corresponds to the second SIF peak and is not as strongly affected by chlorophyll re-absorption as red SIF, thus showing a relatively stable shape at wavelengths above 740 nm (Magney et al., 2019; Parazoo et al., 2019). The differences in the retrieval windows complicate the comparison of SIF retrievals from different sensors; thus, it is useful to provide SIF at a well-defined reference wavelength.

Although the range of the wavelengths used to retrieve SIF from the various sensors is small (740–771 nm), absolute fluorescence can vary greatly, depending on the spectral window used to retrieve SIF (Joiner et al., 2013; Köhler et al., 2018; Sun et al., 2018). However, reference far-red SIF emission spectra at the leaf level indicate that far-red fluorescence spectral shapes are consistent across species (Magney et al., 2019). Thus, we provide an estimate of absolute SIF740 (SIF_740nm) in the GOSAT and OCO-2/-3 SIF Lite files derived from the empirical relationship between SIF at 740 nm and SIF at 758.7 and 770.1 nm (denoted as 757 and 771 nm; Eq. 2). The rationale for including SIF740 in the SIF Lite files is to allow for more consistent and robust comparisons of SIF and SIF-based analyses across sensors (Parazoo et al., 2019) and to reduce the retrieval error by a factor of 2 (Sun et al., 2018). We stress that the reported SIF740 values are not retrieved but are estimated under the assumption that the spectral shape of SIF is invariant. 3SIF740=0.5⋅(1.5⋅SIF757+2.25⋅SIF771). The ratios used in Eq. (3) were based on leaf-level measurements conducted by Magney et al. (2019); however, we observed a median ratio of 1.45 from OCO-2 over vegetated areas for 2015–2019 (Fig. S1 in the Supplement). The reason for this difference has not yet been discerned and requires further analysis, but the small potential bias introduced by the use of the empirical ratio does not infringe on the utility of the SIF740 data.

Biases in retrieved SIF can occur due to uncertainties in the exact instrument line shape per footprint or slight uncertainties in detector linearity. To correct for biases, we use reference targets that are non-fluorescent surfaces barren of vegetation, similar to the method described by Frankenberg et al. (2011b). In short, the background signal over reference targets is subtracted from all relative SIF values. We calculate the background signal for each day as the mean SIF over all barren surfaces within a 31 d window centered on the current day for GOSAT and a 3 d window for OCO-2/-3. These windows were chosen to obtain a robust background signal, given their respective spatiotemporal resolution. Here, we identify barren surfaces using a combination of the MODIS MCD12Q1 land cover data product (Friedl and Sulla-Menashe, 2019) and the Vegetation Photosynthesis Model (VPM; Xiao et al., 2004; Zhang et al., 2017) from the year 2018. The native spatial resolution of these datasets is 500 m, but we aggregated the data to a global 0.20∘ grid so that the barren surface reference targets had a coarser resolution than the soundings. We classified barren surfaces as those grid cells which were 100 % barren and/or snow and ice by MCD12Q1 and had zero (0) annual gross primary production, as estimated by the VPM. We also excluded coastal grid cells that overlapped with water using a global coastline shapefile and a buffer.

We provide an estimate of daily average SIF (Daily_SIF), which is instantaneous SIF scaled entirely upon the geometry of incoming solar radiation over a day. Instantaneous SIF is the absolute value of SIF for any given sounding and is a strong function of the illumination of the canopy at that instant in time. The differences in the illumination geometry of soundings at different overpass times and latitudes complicate direct comparisons of SIF at different points of Earth's surface and comparisons of SIF to other data that are more temporally coarse, such as daily estimates of gross primary production (GPP).

Downwelling solar radiation scales linearly with cos⁡(θ) under clear-sky conditions when ignoring Rayleigh scattering and gas absorption. As described by Frankenberg et al. (2011b) and Köhler et al. (2018), a first-order approximation of daily average SIF (SIFDaily) can be written as follows: 4SIFDaily=SIFt0⋅1cos⁡(θ(t0))⋅∫t=t0-12ht=t0+12hcos⁡(θ(t))⋅H(cos⁡(θ(t)))dt, where SIFt0 is absolute instantaneous SIF, θ(t0) is the solar zenith angle θ at the time of measurement t0 with a heaviside function H to zero out negative values of cos⁡(θ), and the integral is computed numerically in 10 min time steps (dt). In terms of the SIF Lite file variable names, this equation can be written for SIF at any wavelength as Daily_SIF=SIF⋅daily_correction_factor.

We should note that SIF is, to first order, only a proxy for the electron transfer rate in the light reaction of photosystem II. However, SIF is oblivious to the light-independent reactions that fix CO2. Nevertheless, many studies have reported on the linearity of SIF and GPP at bi-weekly or monthly timescales and at coarse spatial resolutions (Verma et al., 2017; Doughty et al., 2019; Yang et al., 2015). The seasonality of SIF and GPP tend to match well at such coarse temporal resolutions because both SIF and GPP are driven by changes in canopy structure, the amount chlorophyll in the canopy, and the amount of sunlight (photosynthetically active radiation; PAR) being absorbed by canopy chlorophyll (APARchl; Magney et al., 2020; Doughty et al., 2021; Dechant et al., 2019). The SIF–GPP relationship can also become more linear at the canopy scale due to the contribution of total canopy SIF by sunlit, shaded, stressed, and non-stressed leaves (Magney et al., 2019). SIF and GPP have an indirect relationship through non-photochemical quenching and the electron transport rate (Porcar-Castell et al., 2014; Gu et al., 2019), which can sometimes simultaneously downregulate photosynthesis and SIF, as has been seen in evergreen needleleaf ecosystems, but not always (Magney et al., 2019).

At the leaf level, GPP saturates before SIF in response to APAR, such that we could see increased SIF without any response in GPP at high levels of APAR (Gu et al., 2019). Conversely, vegetation stress can cause a near- or total cessation of GPP via stomatal closure with little or no change in SIF. This decoupling has been seen at the leaf scale during the forced stomatal closure of deciduous tree species (Marrs et al., 2020) and a 1-month drought experiment with eastern cottonwood (Populus deltoides) (Helm et al., 2020). However, these studies, and others of deciduous vegetation and croplands, have repeatedly found a better correlation between SIF and APAR than SIF and GPP (Yang et al., 2018; Miao et al., 2018). For SIF to be a reliable proxy of APAR, SIFyield (ratio of SIF to APAR) would need to remain constant. For a detailed inquiry into SIF and photosynthesis, see Porcar-Castell et al. (2014), and a review of SIF remote sensing applications and challenges from the leaf, tower, and satellite scale by Magney et al. (2020) and Mohammed et al. (2019).

Data users are likely to find negative SIF values, which are due to retrieval noise, but these values should generally not be discarded. The 1σ uncertainty in retrieved SIF values (SIF_Uncertainty) can be substantial, but negative values are plausible in a retrieval sense although not in physical terms (the actual SIF emission cannot be negative). Discarding negative values will introduce a high bias when averaging. Nevertheless, extremely negative values may indicate a problem with the retrieval. We recommend the following guidelines for filtering negative SIF values: accept if SIF +2-σ uncertainty ≥0; questionable if SIF +2-σ uncertainty <0 and SIF +3-σ uncertainty ≥0; and reject if SIF +3-σ uncertainty <0. These thresholds have not been incorporated into the Quality_Flag variable of the SIF Lite data.

Users of SIF data from any source should be aware that sun sensor geometry plays a role in the absolute values of SIF, in addition to vegetation canopy characteristics (Joiner et al., 2020; Köhler et al., 2018). Absolute SIF values increase rapidly when the phase angle approaches 0∘ (when the Sun and sensor are aligned), but the effect of sun sensor geometry has been shown to be small when the phase angle is greater than 20∘ (Köhler et al., 2018; Doughty et al., 2019). Thus, retrieved SIF values from target- or SAM-mode scans during a single overpass can vary greatly, despite homogeneous vegetation cover, due to changing sun sensor geometries during data acquisition. Figure 3 illustrates the phase angle and SIF757 for a SAM acquired over the Amazon rainforest, where the vegetation canopy is very homogenous. The figure also illustrates how the phase angle changes during an OCO-3 SAM scan, and that the sun sensor geometries for each individual swath are rather distinct from each other (Fig. 3a). Mean SIF for each swath is also distinctively different (Fig. 3b), despite the fact that the canopy was experiencing the same illumination geometry and environmental conditions during the 2 min in which this SAM was acquired. The effect of sun sensor geometry is also illustrated in Fig. 4, which shows the relationship between SIF for individual OCO-2 soundings and phase angle for two target scans in the Amazon. A distinctive change in the absolute values of retrieved SIF were observed due to sun sensor geometry.

There are the following two main challenges to working with all spaceborne SIF data: (1) the inherently large uncertainties for individual soundings due to retrieval noise, and (2) the effect of differences in sun sensor geometry on retrieved SIF values. Thus, we advise against using single soundings for analysis. However, averaging soundings across space and time can reduce the retrieval noise by a factor of 1/n, with n being the number of soundings comprising the average (Frankenberg et al., 2014). For platforms with a wide swath, like the TROPOspheric Monitoring Instrument (TROPOMI), the effect of sun sensor geometry can be accounted for by averaging soundings for a point of interest over the entire repeat cycle (16 d for TROPOMI), as demonstrated by Doughty et al. (2019, 2021). In the case of OCO-2/-3, as we demonstrate in Fig. 3 and in Braghiere et al. (2021), soundings can be grouped by phase angle and then averaged to reduce retrieval uncertainty. Thus, retrieval uncertainty and sun sensor geometry effects can be substantially minimized. For GOSAT, we recommend averaging SIF retrieved from both the P and S polarizations, as demonstrated in Fig. 5.

Users should also keep in mind that, when conducting analyses at large spatial scales, gridding the data prior to analysis is largely unnecessary as the ungridded Level 2 data can be used directly (Doughty et al., 2019). Doing so will allow the users to retain sounding-level information that may aid in the interpretation of the results, which would otherwise be lost when merely gridding the SIF values. For instance, as demonstrated by Doughty et al. (2019), ungridded Level 2 SIF data were used to calculate mean SIF for the entire Amazon basin at different phase angles to show that the seasonality of SIF in the Amazon basin was consistent across sun sensor geometries. Such an analysis would not have been possible with gridded data because, after gridding, it is impossible to group the data by sounding-level attributes, such as phase angle or cloud fraction.

It is important to note that in areas where the SIF signal is near zero, the use of SIF at 757 nm would be more appropriate, as the SIF signal is stronger at this wavelength. In areas where vegetation is sparse or SIFyield is low due to vegetation responses to environmental conditions or canopy leaf physiology, SIF at 771 nm could be within the noise range due to its relatively far distance from the far-red peak at 740 nm. In these cases, we advise the use of SIF at 757 nm. Since SIF at 771 nm is used to compute SIF at 740 nm in the SIF Lite files, diligence should likewise be used when using SIF at 740 nm in analyses.

OCO-3 SIF has been shown to have a very high correlation (r>0.9) with OCO-2 (Taylor et al., 2020). Here, we present the first comparisons between GOSAT and OCO-2 Level 2 data. Currently, there are not enough coincident soundings for GOSAT and OCO-3 to provide a robust analysis, but given that OCO-2 and OCO-3 compare very well, we expect a comparison between GOSAT and OCO-3 to mimic the findings from our GOSAT and OCO-2 comparison.

Although the data records for GOSAT and OCO-2 overlap 6 years, only a small percentage of soundings flagged as best quality and cloud free from GOSAT and OCO-2 overlap on the same day (Fig. 5a). Despite this filter, the mean SIF values may differ widely on the same day due to differences in overpass time (and thus solar illumination angle and environmental conditions), viewing geometry, and the number of OCO-2 soundings comprising the mean. We progressively filtered the data, as illustrated in Fig. 5, to ensure the soundings were of a vegetated land surface and had similar sun sensor geometries, environmental and atmospheric conditions, and that the temperature was high enough for photosynthesis to occur, as indicated by the temperature_skin variable in the SIF Lite data.

We found that the correlation and slope improved with a more conservative filtering of the data, and that the comparison between GOSAT SIF and OCO-2 SIF was reasonable. However, it is important to note that any comparison between GOSAT and OCO data will inevitably be affected by spatial sampling bias, as the swath width for both OCO platforms is smaller than the diameter of the GOSAT sounding footprints (Fig. 6; left footprints). Also, it could be the case that only a small portion of the GOSAT footprint is sampled by OCO (Fig. 6; right footprints). Our filter of ≥10 OCO-2 soundings within a GOSAT footprint aimed to reduce this potential sampling bias in addition to reducing the uncertainty of the OCO-2 SIF retrievals. It must also be remembered that, in this comparison, we do not have the luxury to average several GOSAT soundings to reduce the uncertainty as we did with OCO-2, so the uncertainties of the GOSAT SIF is much higher than that for OCO-2.

Upon a more detailed comparison of GOSAT and OCO-2 SIF at 740, 757, and 771 nm, using the strictest filter we applied in Fig. 5f, we found SIF740 from the two platforms to have higher correlations than for SIF757 and SIF771 alone (Fig. 7). We also noticed that GOSAT and OCO-2 soundings most frequently overlap in the boreal winter, which corresponds to a period of little or no photosynthesis at mid- and high latitudes (Figs. S2 and S3). Thus, the direct comparison of GOSAT and OCO-2 SIF is severely restricted by the relatively infrequent overlap of the two platforms during the growing season.

In addition to the sounding level comparisons, we found that mean annual SIF757 for GOSAT and OCO-2 to compare reasonably well at the global scale during the boreal summer (Fig. 8). The relatively large differences in SIF illustrated at the grid cell level in Fig. 8c are due to differences in the spatial and temporal sampling of the two platforms. We presented the comparison here at 4.0∘ spatial resolution to improve the sampling by GOSAT (Fig. 1a).

Currently, the target and SAM soundings are not collated to the target to which they correspond, but variables will be added to upcoming versions (e.g., v11) of the SIF Lite files that will collocate the target and SAM soundings with their intended target site. For OCO-3, some of the target sites are in close proximity to each other, and thus, a target site may fall within the scan of another target. For these sites, users may also want to check scans that were intended for target sites adjacent to their target of interest. The OCO-3 targets, the dates of their scans, and scan maps are available at https://ocov3.jpl.nasa.gov/sams/index.php (last access: 7 March 2022). A list of target locations for OCO-2 and OCO-3 are available in Tables S1 and S2 in the Supplement, respectively.