1. Introduction
Many works have been devoted to reducing noise and artifacts in satellite data using spatial- and spectral-based approaches. Random and systematic errors in satellite earthshine radiance observations arise from quantization, shot noise, and dark currents. These errors can limit the usefulness of satellite-derived products for science and applications. Efforts to denoise hyperspectral imagery (HSI) have evolved from two-dimensional (2D) spatial approaches applied to individual bands to three-dimensional (3D) methods that consider the spectral dimension using techniques such as wavelet and low-rank modeling [see, e.g., the review of Rasti et al. (2018)]. Machine learning (ML) methods, including convolutional neural networks (NNs), have also been employed to denoise HSI (e.g., Wang et al. 2022). There has been less documented research with purely spectral methods that differ from spatial noise reduction techniques reviewed by Thakur et al. (2021).
Our work focuses on the reduction of noise and other artifacts in downstream satellite products that are retrieved using observed radiance spectra. Joiner et al. (2023) suggested the use of ML along with principal component analysis (PCA) of radiance spectra for noise reduction in retrieved satellite-based nitrogen dioxide (NO2) slant column densities (SCDs). Their hypothesis was that ML models can reproduce spatiotemporal characteristics of noisy retrievals with acceptable performance using coefficients of a limited number of leading principal components (PCs) of the observed spectral radiances as input features. By eliminating coefficients of trailing PCs that primarily contain random noise, ML models are unable to reproduce noise in the retrievals. Joiner et al. (2023) applied their approach directly to actual NO2 retrievals for which true values were unknown. Their method was evaluated using retrievals in an area where the target values were expected to have low variability.
Here, we apply a similar approach to satellite-derived solar-induced fluorescence (SIF) retrievals that suffer from relatively low signal levels as compared with retrieval noise. SIF from chlorophyll-a in terrestrial vegetation can be measured using satellite-, aircraft-, and ground-based spectrometers [see, e.g., the review of Mohammed et al. (2019)]. There are many potential applications of SIF measurements in the fields of ecology, precision agriculture, hydrology, climate feedback, and socioeconomics [see, e.g., the reviews of Porcar-Castell et al. (2021), Sun et al. (2023a,b), and references therein].
A conceptual framework of noise reduction discussed by Joiner et al. (2023) as applied to SIF retrievals is laid out in Fig. 1. The leading PCs (or eigenvectors) of near-infrared (NIR) radiances contain information about geophysical features in Earth’s surface and atmosphere, such as water vapor absorption. These features are present in every spectrum and explain a large fraction of spectral variability within any sample of spectra. For example, >99% of NIR spectral variability is captured by a handful of leading PCs (Joiner et al. 2013). Random noise in satellite measurements is typically spread out across a large number of trailing PCs, and these PCs explain only a small fraction of total spectral variability. Our hypothesis is that the SIF spectral signature is presumed to be present primarily in the leading PCs such that an NN can reconstruct the SIF retrieval with only the coefficients of the leading PCs and without propagating the noise contained in the trailing PCs. Our approach can be classified as a purely spectral noise reduction method. As explained in more detail below, our ML method is better able to reconstruct SIF retrievals and reduce noise as compared with retrievals performed with reconstructed radiances that use leading PCs to reduce noise. In short, the small SIF signal in real satellite spectra is spread out over a large number of PCs. This large spread leads to biases and marginal noise reduction in retrievals performed with reconstructed radiances.
Conceptual diagram of the NR technique described by Joiner et al. (2023).
Citation: Artificial Intelligence for the Earth Systems 3, 3; 10.1175/AIES-D-23-0085.1
Satellite SIF retrievals have relied upon the filling-in of solar Fraunhofer features in Earth’s backscattered reflectance spectra from SIF emission (Joiner et al. 2011; Frankenberg et al. 2011a; Joiner et al. 2013; Guanter et al. 2015; Köhler et al. 2015b, 2018a). Most retrievals have focused on SIF at the far-red emission feature that peaks near 740 nm. The retrieval spectral fitting window for SIF may contain atmospheric absorption features, such as from oxygen (O2) or water vapor (H2O). This absorption can be accounted for using a data-driven method that relies on an analysis of spectra over fluorescence-free scenes (Joiner et al. 2013, 2016). The retrieval in this case is nonlinear and requires an iterative solution. If a small spectral fitting window is used for which atmospheric absorption is small and can be neglected, the retrieval can be simplified (Joiner et al. 2011; Frankenberg et al. 2011a; Köhler et al. 2018a). For example, a linear regression can be used where instrumental artifacts and small absorption features can be fit simultaneously with spectral features associated with SIF and surface reflectance.
In this work, we consider only retrievals of far-red SIF emission near 740 nm that make up the vast majority of terrestrial SIF data products. Far-red SIF retrievals have been performed with spectrometers designed for greenhouse gas (GHG) retrievals including the thermal and near-infrared sensor for carbon observation Fourier transform spectrometer (TANSO-FTS) flying on the Greenhouse Gases Observing Satellite (GOSAT) (Joiner et al. 2011; Frankenberg et al. 2011b; Guanter et al. 2012; Köhler et al. 2015a), the Orbiting Carbon Observatory 2 (OCO-2) on a free-flying satellite, OCO-3 mounted on the Japanese Experiment Module-Exposed Facility International Space Station (ISS) (Sun et al. 2017, 2018; Doughty et al. 2022), and the atmospheric carbon dioxide grating spectrometer (ACGS) aboard the Chinese carbon dioxide observation satellite (TanSat) (Du et al. 2018; Yao et al. 2022). The high spectral resolution of the GHG sensors, with coverage in the O2 A band for characterization of cloud and aerosol impacts on photon pathlength, enables the simpler linear approaches to SIF retrievals. Moderate-spectral-resolution spectrometers, designed to measure trace gas pollutants, typically have a broader spectral range that can be used to retrieve SIF with nonlinear retrieval approaches. Such instruments include the Global Ozone Monitoring Experiment 2 (GOME-2) sensors that fly on the European Organisation for the Exploitation of Meteorological Satellites (EUMETSAT) MetOp series (Joiner et al. 2013; Sanders et al. 2016; van Schaik et al. 2020), the GOME instrument on the European Space Agency’s (ESA) European Remote-Sensing Satellite-2 (ERS-2) (Joiner et al. 2019), and the Scanning Imaging Absorption Spectrometer for Atmospheric Cartography (SCIAMACHY) on the ESA Environmental Satellite (Envisat) (Köhler et al. 2015b; Khosravi et al. 2015). The Tropospheric Monitoring Instrument (TROPOMI) on the Copernicus Sentinel-5 Precursor (S-5P) satellite has moderate spectral resolution with a higher sampling rate and achieves good performance with a linear SIF retrieval (Köhler et al. 2018a; Guanter et al. 2021). Several of these datasets are being harmonized to form long-term SIF records (Parazoo et al. 2019; Doughty et al. 2022).
Individual far-red SIF retrieval errors are typically of the same order of magnitude as the peak emission strength. SIF retrievals thus have much lower signal-to-noise ratios (SNRs) as compared with many other satellite-derived data products. Due to the relatively large retrieval errors, individual SIF retrievals are typically aggregated in time and/or space for scientific studies (e.g., Köhler et al. 2018b). For example, in order to examine the dependences of the sun–satellite geometry on SIF retrievals, substantial spatiotemporal averaging was applied (e.g., Joiner et al. 2020). It has not yet been possible to estimate parameters related to the spatiotemporal varying bidirectionality of SIF as has been done for reflectances (e.g., Lyapustin et al. 2011a,b, 2012; Schaaf et al. 2002).
Here, we test our hypothesis that an NN model can reproduce spatiotemporal characteristics of noisy SIF retrievals using a limited number of leading PCs of NIR radiance spectra, thus enabling noise reduction. We first apply our ML noise reduction approach to simulated spectra under a variety of scenarios for which true SIF values are available. In this controlled test bed, we can vary the SNR of the simulated observations. Results are assessed by comparing retrieved with true SIF values. We then apply our methodology to SIF retrievals from GOME-2 instruments aboard two MetOp satellites. The training is completely independent of the simulation data and can account for instrument-specific artifacts. Because the true SIF is not known with the GOME-2 satellite retrievals, we assess the results by comparing retrievals from two different satellites. We also compare SIF variability with the variability of a completely independent satellite-based gross primary production (GPP) dataset that has been calibrated using ground-based eddy covariance measurements. Variability of SIF and GPP over time scales of weeks to months is expected to be highly similar.
2. Data and methods
a. Machine learning framework
Figure 2 shows a flow diagram of our SIF noise reduction approach as applied to either simulated or GOME-2 satellite data. We start with reflectances (or radiances) from a sensor. If the observations are simulated, noise is added according to a particular SNR model. A retrieval is then performed with the noisy spectra and will be referred to as a noisy retrieval. With real satellite data, it is beneficial to perform a bias adjustment (Joiner et al. 2020) using data over the ocean, where far-red SIF values should be close to zero, to mitigate systematic errors such as those described by Köhler et al. (2015b). In contrast to terrestrial SIF, marine SIF that escapes the ocean originates almost exclusively at red wavelengths and not at the far-red wavelengths considered here (Köhler et al. 2020). Removal of poor-quality spectra is also needed for effective training. A collection of noisy, bias-adjusted, and quality-controlled retrievals comprises the target for NN training. The noisy spectra are used to compute the coefficients of the leading PCs. These coefficients and optional geometrical predictors, e.g., cosines of the solar zenith angle (SZA), view zenith angle (VZA), and the phase angle (the angle between the sun and satellite), form the set of NN predictors or features.
Flow diagram showing steps of our SIF NR algorithm. Optional bias adjustments and quality control measures are applied to the actual satellite data and are not needed with the simulated data.
Citation: Artificial Intelligence for the Earth Systems 3, 3; 10.1175/AIES-D-23-0085.1
The noise-reduced (NR) SIF retrievals may undergo a final bias adjustment that can be beneficial for real satellite data. The NR retrievals are compared with true SIF within the simulation framework. The NR retrievals can also be compared with the noisy retrievals to check for biases. This is important as there is no guarantee that the NR SIF retrievals will represent an unbiased version of the noisy retrievals. Finally, with real satellite-based SIF data, an NN may be trained using data where SIF is expected to be zero (e.g., over ocean or deserts) to predict the SIF retrieval noise for noisy or NR retrievals as a function of various parameters, including SZA, VZA, and/or continuum-level reflectance/radiances (Joiner et al. 2020).
The NN training is designed to reproduce the corresponding target SIF values from the original noisy retrievals. By presenting the NN during training with only the coefficients of the leading PCs as features, the NN will not be able to reconstruct the random noise in the retrievals. So long as sufficient information about SIF is contained in the leading PCs, our hypothesis is that the trained NN will reproduce the broad spatiotemporal variability of the retrievals.
The NR algorithm may benefit from utilizing a larger spectral fitting window than is used in noisy retrievals. For example, current satellite-based SIF retrievals have generally avoided wavelengths within the strong O2 A band. Nonlinearity and stray light may impact radiance measurements in the deep core of this band. Our ML-based method may be able to account for these effects to utilize the SIF information within this absorption band. Our method may also be able to incorporate more spectral range across the so-called red edge (wavelengths from about 717–734 nm for far-red SIF retrievals) that is associated with sharp increases in reflectance owing to decreasing chlorophyll absorption. Strong H2O features within this spectral range can be filled in by SIF emission and lead to systematic SIF retrieval errors (Parazoo et al. 2019).
For the NN architecture, we use a model with two hidden layers and a variable number of nodes in the hidden layers between 1 and 2 times the number of input features. Both inputs and outputs are scaled to give a mean of zero and a standard deviation of unity. The training uses an adaptive moment estimation optimizer with a learning rate of 0.1. The activation functions for each layer are soft sign, logistic (sigmoid), and bent identity in that order.
b. Simulated data
The simulated radiance and irradiance data are similar to those used previously (Joiner et al. 2013, 2016; Guanter et al. 2015). As in Joiner et al. (2013), radiance spectra are computed for a range of observing geometries and atmospheric and surface conditions:
-
VZAs of 0° and 16°; SZAs of 15°, 30°, 45°, and 70°; middle-latitude summer and winter temperature profiles; surface pressures Ps of 955, 980, 1005, and 1030 hPa; total water vapor column (TWVC) amounts of 0.5, 1.5, 2.5, and 4.0 cm; aerosol optical thicknesses (AOTs) at 550 nm of 0.05, 0.12, 0.2, 0.3, and 0.4; and aerosol layer heights (ALHs) of 700–900, 600–800, and 500–700 hPa with a typical continental aerosol model.
There are two separate simulated datasets: 1) simulated without fluorescence, consisting of 10 different models of surface spectral reflectance based on soil and snow, and 2) surface spectral reflectance and SIF produced with the FluorSAIL and FluorMODleaf codes (Miller et al. 2005; Jacquemoud et al. 2009; Pedrós et al. 2010). The latter has 60 different SIF values based on combinations of chlorophyll content (5, 10, 20, and 40 μg cm−2), leaf area index (LAI) (0.5, 1, 2, 3, and 4), and quantum efficiency (0.02, 0.05, and 0.08). The atmospheric radiative transfer uses the matrix operator model (MOMO) (Fell and Fischer 2001; Preusker and Lindstrot 2009) with gaseous absorption line parameters from the high-resolution atmospheric radiance and transmittance model code (HITRAN) 2008 dataset (Rothman et al. 2009).
Earthshine radiances are computed monochromatically and sampled at 0.005 nm. A solar irradiance spectrum is sampled in the same way. The spectra are convolved with instrument line shape functions of various full-width half-maximum (FWHM) values and resampled to produce data similar to those from a satellite instrument. We add noise according to a normal distribution based on an SNR model that is a function of the simulated radiance.
c. GOME-2 data
GOME-2 is a nadir-viewing spectrometer that measures earthshine radiance and solar irradiance at wavelengths from the ultraviolet to the NIR (Munro et al. 2016). It scans cross track providing a swath of width 1920 km in its nominal mode. GOME-2 flies aboard the EUMETSAT Polar System (EPS) MetOp mission series. So far, three instruments have flown. Here, we use data from GOME-2 on MetOp-A (GOME-2A), launched in October 2006, and on MetOp-B (GOME-2B), launched in September 2012. The drifting orbit of MetOp-A has resulted in a situation where GOME-2A lost its direct line of sight to the sun on 26 January 2018. As a consequence, the calibration quality deteriorated along with the measurement time, and this rendered SIF measurements less useful. We focus on the overlap period between the GOME-2A and GOME-2B datasets from April 2013 through January 2018. Both satellites are in a polar orbit with an equator crossing time near 0930 LT.
GOME-2 has four separate detector channels. We use earthshine radiance and solar irradiance data from revision R2 in channel 4 that covers wavelengths 590–790 nm. The spectral resolution (FWHM) in this band is ∼0.5 nm, and it is sampled at approximately 0.2 nm (Munro et al. 2016). GOME-2 has an SNR in band 4 > 1000.
3. Results
a. Results with simulated data
We use simulated data from a hypothetical instrument with an FWHM of 0.1 nm sampled every 0.03 nm. The spectral fitting window used for the noisy retrievals is 750–758 nm. This spectral range contains little atmospheric absorption, enabling a linear retrieval. Figure 3 shows the baseline SNR as a function of radiance used in the simulations.
Assumed nominal SNR used in simulations.
Citation: Artificial Intelligence for the Earth Systems 3, 3; 10.1175/AIES-D-23-0085.1
Statistical results for retrievals with no noise and with the baseline SNR are presented as experiments (Exps.) 1 and 2, respectively, in Table 1. We then apply the noise reduction (NR) approach using an expanded fitting window of 735.5–758 nm. The input features for this experiment (and all others in this section unless otherwise specified) are coefficients of the leading 12 PCs and cosine of the SZA. We are able to achieve a significant noise reduction; RMSE decreases by more than a factor of 3 as shown in Exp. 3 in Table 1.
Statistical comparison of SIF retrievals from simulations using 130 227 independent observations as compared with the truth (i.e., data not used in training) including the wavelength (nm) range used for the retrieval or NR algorithms and coefficient of determination r2 [units of the SIF bias (mean difference with respect to the truth) and RMSE are in mW m−2 nm−1 sr−1]. Unless otherwise noted, the NR uses 12 PC coefficients, includes the cosine of the SZA, and does not include VZA as input features.
The so-called red edge comprises spectral reflectances in the approximate range of 700–760 nm. Within the range, reflectances for vegetated scenes rise dramatically with wavelength as a result of decreasing chlorophyll absorption. Many popular vegetation indices such as the normalized difference vegetation index (NDVI) and near-infrared reflectance of vegetation (NIRv) make use of wavelengths on either side of the red edge. These indices are used as proxies for the total amount of green vegetation in a scene. To address the question of whether or not substantial information for the NR is coming from the red edge reflectances, we apply additional processing to the simulated reflectance spectra. First, we fit a third-order polynomial to each reflectance spectrum. We then use only the fitted polynomial for the PCA. This ensures that only the broad shape of the red edge is used for NR. The results are shown as Exp. 4. While there is a small decrease in the RMSE as compared with the noisy retrievals, the captured variance r2 is substantially lower than that of the original noisy retrievals. These statistics indicate that the red edge by itself is not able to reproduce the SIF variability contained in the simulated dataset.
In Exp. 5 in Table 1, we subtract the fitted polynomial from the spectra before applying the PCA. This ensures that red edge information is not utilized in the NR. In this scenario, practically all information about SIF is provided by the filling-in of solar and telluric features. Here, we achieve a substantial noise reduction although not as much as with the unaltered spectra. These results indicate that the NR relies primarily on information from the Fraunhofer filling-in with a small amount of additional information from the red edge.
Experiments 6–8 use different spectral fitting windows for NR with no alteration to the original reflectance spectra. Using increasingly smaller fitting windows, as may be expected, leads to less NR. However, even the relatively narrow retrieval fitting of 750–758 nm provides a reduction in RMSE of ∼20%.
Experiments 9–14 include the O2 A band spectral range in the NR. The 735.5–775 nm range (Exp. 9) provides NR similar to the 735.5–758.0 nm range. Increasing the number of PCs to 14 provides slightly more noise reduction (Exp. 10) than with 12 PCs. Adding the cosine of the VZA as a predictor (Exp. 11) did not improve the results. Removing both SZA and VZA as predictors in Exp. 12 results in a very slight degradation compared with results that use SZA as a predictor. Using 750–775 nm (Exps. 13–14) reduces RMSE by about 30% as compared with 750–758 nm (Exp. 8). The use of 14 PCs as compared with 12 PCs with this spectral range provides a small additional benefit.
Experiments 15–16 similarly show results of the NR where the SNR in Fig. 3 has been multiplied by a factor of 2 (the noise is halved). In this scenario, the noisy retrievals are much improved (RMSE about 50% lower) as compared with the nominal SNR. The amount of noise reduction is about a factor of 2, somewhat less than with the nominal SNR. Histograms of the SIF retrieval errors for these experiments are shown in Fig. 4 along with those from the simulations with nominal noise.
Histograms of SIF retrieval errors using simulated observations with (a) nominal SNR from Fig. 3 and (b) 2 × nominal SNR (noise halved).
Citation: Artificial Intelligence for the Earth Systems 3, 3; 10.1175/AIES-D-23-0085.1
In Exps. 17–19, we repeat some of the above experiments with the SNR reduced by a factor of 2. This increases the RMSE of the noisy retrievals (Exp. 17) by a factor of about 2. The NR with the 735.5–758 and 750–775 nm windows (Exps. 18 and 19) reduces the RMSE by about 75% and 50%, respectively, as compared with the noisy retrievals.
b. Results with GOME-2 satellite data
1) Training frequency
Here, we train an NN for each day of data collection separately for GOME-2A and GOME-2B. Separate training for each instrument allows for individual satellite instrumental artifacts to be accounted for. Frequent training may mitigate the effects of temporal variations in noise or artifacts. However, such frequent retraining adds computational cost and could lead to some unintended day-to-day variations.
2) Bias adjustment
Here, we use the bias adjustment scheme described in Joiner et al. (2020). The bias adjustment employs a similar NN model as for the NR and uses SIF retrievals over the ocean as the target that are assumed to have values of zero. The five predictors are cosines of the VZA, SZA, and relative azimuth angle along with latitude and 780-nm reflectance. Note that in the original noisy retrievals, the bias adjustment scheme did not include 780-nm reflectance as a predictor. We found that better results were obtained for the NR when 780-nm reflectance was used as a bias adjustment predictor.
3) Quality control filtering
For training, we use only retrievals where −1 < SIF < 5 mW m−2 nm−1 sr−1, solar zenith angle < 70°, and the retrieval quality value > 1, indicating a good retrieval with effective cloud fractions fc < 30%. It should be noted that the effective cloud fraction is not the same as the geometrical cloud fraction and is applied in the context of the independent pixel approximation in which a pixel radiance is modeled as the sum of clear and cloudy subpixels, where the cloudy subpixel is modeled as an opaque surface with a reflectance of 0.8. We estimate fc = (RTOA − Rs)/(0.8 − Rs), where Rs is the surface albedo and RTOA is the top-of-atmosphere reflectance. We remove data where the reflectance at 680 nm > 0.6 (assumed to be over snow or ice) and at latitudes 60°–90°S.
Positive SIF retrieved values in the southern part of the Arabian Peninsula are attributed to the luminescence of the surface rocks (Köhler et al. 2021). We therefore exclude data covering 10°–30°N and 45°–60°E from the training set. Removing retrievals in the area improved SIF retrieval fits to the independent data and reduced the false positive terrestrial SIF values in this region in the NR retrievals.
An additional two-step check eliminates spectra that are not well reconstructed using leading PCs. Problematic spectra can include those affected by saturation or nonlinearity in bright scenes, large amounts of stray light, severe noise in the South Atlantic anomaly (SAA), and eclipses (Gorkavyi et al. 2021). In step 2, we reconstruct each spectrum using 28 leading PCs from step 1. The reconstruction errors (differences between the observed and reconstructed spectra) are used to identify problematic spectra. For each spectrum i, we compute the maximum absolute value |ϵmax| and the standard deviation σi of the reconstruction error over all wavelengths. We retain only spectra for a second PCA for which 1) |ϵmax| < 5σmax, where σmax is the standard deviation of |ϵmax| computed over all spectra, and 2)
The final sample of data used for NN training includes all data over land passing the abovementioned quality control criteria as well as ocean data with an additional check that |SIF| < 0.75 mW m−2 nm−1 sr−1. We found frequent cases of clear sky pixels over the ocean adjacent to cloudy pixels where we hypothesize that 3D radiative cloud effects contribute light from outside the footprint that can fill in solar Fraunhofer lines, leading to false SIF signals. Maintaining this check on SIF over the ocean removes such cases from the training set. The cloud 3D effects likely have a somewhat different spectral signature than SIF so that during the NN training, if eliminated from the training set, the NN does not learn to attribute them to SIF. Thus, excluding the affected observations leads to noticeable improvements over the dark ocean and presumably to less obvious improvements over land.
4) Spectral window and input features
To optimize the spectral fitting window and number of PCs used in the noise reduction, we examine SIF values over the ocean that should tend toward zero as random and systematic errors are reduced. We report statistics of noisy and NR SIF retrievals over the ocean derived from GOME-2A on 30 July 2007 in Table 2 along with statistics of the model fit of the NR to the noisy retrievals. An important objective in assessing NR performance for real data is to check random and systematic SIF errors over the ocean, where the expected SIF values should be close to zero as explained above.
Mean and standard deviation σ of SIF over the ocean from GOME-2A on 30 Jul 2007 and statistics of the model fit to the noisy bias-adjusted retrievals: coefficient of determination r2, bias, and RMSE (units of the SIF mean, σ, bias, and RMSE are in mW m−2 nm−1 sr−1).
Using 710–758 nm for NR, the SIF standard deviation over the ocean is reduced by more than a factor of 4 with 12–20 PCs. We found that more PCs are needed with real data as compared with the simulated data, likely because the fitting window is larger and incorporates stronger water vapor absorption; additional PCs are needed to explain systematic instrumental artifacts, such as from wavelength shifts, that are absent in the simulated data. As expected, the mean SIF value also decreases in magnitude toward zero for NR retrievals, the SIF standard deviation over the ocean increases with the number of PCs used, and the r2 values for the fit to the noisy data also increase.
The use of a larger wavelength range for NR that encompasses the O2 A band with 22 PCs did not substantially improve results as compared with results that did not include the A band. This finding contrasts with the results using simulated data. It is possible that there are complex measurement errors within this deep band, such as effects of stray light or enhanced cloud 3D effects, that limit its usefulness for noise reduction.
For all subsequent GOME-2 SIF results, we use the 710–758 nm fitting window with 20 PCs. The use of 20 PCs provides both good noise reduction and a reasonable fit to the dynamic range of original noisy SIF data (both high and low values). Below this number, the SIF standard deviation over the ocean is smaller, but the small number of SIF retrievals with large values was not captured as well. We perform further analyses below at different time scales.
5) Comparisons at a daily time scale
Figures 5a and 5b show example daily maps of noisy SIF retrievals from GOME-2B without and with bias adjustments, respectively. Figures 5c and 5d show corresponding NR retrievals without and with additional bias adjustment, respectively. The NR retrievals are noticeably cleaner than the original noisy retrievals. The bias adjustment performed on the NR retrievals does not show clear signs of further improvement by eye but produced slightly better statistical results over the ocean and is therefore maintained in the processing chain.
Maps from GOME-2B on 28 Jul 2013 showing (a) noisy retrievals without bias adjustment; (b) as in (a), but with bias adjustment; (c) as in (a), but for NR retrievals; (d) as in (b), but for NR retrievals; (e) estimated uncertainties for noisy retrievals; and (f) as in (e), but for NR retrievals.
Citation: Artificial Intelligence for the Earth Systems 3, 3; 10.1175/AIES-D-23-0085.1
Very noisy SIF retrievals are seen over the SAA region (southern Brazil extending into the Atlantic Ocean) that are not present in the NR retrievals. Severely affected retrievals have been filtered out by the spectral reconstruction quality control check.
As described above, SIF retrievals over the ocean are used to train additional NNs to estimate SIF uncertainties for noisy and NR retrievals. Figures 5e and 5f show estimated uncertainties from an NN trained with data over the ocean. Single pixel uncertainties for the original noisy retrievals are mostly in the range of 0.6 to more than 1 mW m−2 nm−1 sr−1, with the largest uncertainties corresponding to bright scenes at high latitudes. In contrast, uncertainties are below about 0.2 mW m−2 nm−1 sr−1 for NR retrievals over most of the globe with only a small latitudinal band of larger errors up to about 0.5 mW m−2 nm−1 sr−1 at the region of highest radiance values in the tropics.
6) Comparisons at a monthly time scale
Visual improvements in the NR as compared to noisy retrievals are also seen in monthly averages. Figures 6a and 6c show monthly mean gridded SIF maps from GOME-2B without and with NR, respectively, for July 2014. The monthly mean SIF values are similar for noisy and noise-reduced retrievals although there are some subtle differences. High SIF values in the U.S. corn belt are a bit lower in the NR SIF as compared with noisy retrievals, while NR SIF values in the tropics are somewhat higher than in the noisy retrievals. We found similar differences in other months.
Monthly averaged gridded data from GOME-2B in July 2014 at 0.5° × 0.5° resolution (a) noisy SIF retrievals; (b) monthly gridded standard deviations from noisy retrievals; (c) as in (a), but for NR retrievals; (d) as in (b), but for NR retrievals; (e) as in (a), but for GPP from the FluxSat dataset; and (f) as in (b), but for FluxSat GPP.
Citation: Artificial Intelligence for the Earth Systems 3, 3; 10.1175/AIES-D-23-0085.1
Some artifacts over the ocean are removed in the NR retrievals. We found that positive SIF values in the noisy retrievals are sometimes found in clear or nearly clear sky pixels over the ocean that border cloudy regions. As discussed above, cloud 3D effects may contribute to such erroneous SIF retrievals. We found unexpected filling in of the Hα solar line at 656 nm for some of these cases. The NN appears able to distinguish these spectral effects from those of SIF.
In the southern part of the Arabian Peninsula, where luminescence from minerals in rocks occurs, positive SIF values are much reduced in NR retrievals. The spectral dependence of the luminescence was shown to be different from that of chlorophyll SIF (Köhler et al. 2021). The trained NN again appears to somewhat filter out the luminescence spectral signal.
Corresponding monthly SIF standard deviations, calculated at each grid box using all the retrievals over the given month, are shown in Figs. 6b and 6d. These gridbox standard deviations include both retrieval errors and natural variability. The monthly standard deviations are substantially smaller for the NR retrievals as compared with noisy retrievals. Extremely high standard deviations in the vicinity of the SAA are present in the noisy retrievals with values in excess of 1.5 mW m−2 nm−1 sr−1, whereas these high standard deviations are absent in the NR retrievals.
As an independent validation check as to whether the more defined spatial features in the NR monthly averages are realistic, we show monthly mean and gridbox standard deviations of estimated GPP from the FluxSat dataset in Figs. 6e and 6f, respectively, at the same spatiotemporal resolution as the SIF data. FluxSat GPP is derived from MODIS reflectances calibrated with eddy covariance estimates (Joiner et al. 2018; Joiner and Yoshida 2020). It is not expected that there is a constant scaling from SIF to GPP (Sun et al. 2023b); therefore, we should not expect a one-to-one correspondence between SIF and GPP. Nevertheless, spatiotemporal variations in the GOME-2 NR SIF and FluxSat GPP show similar characteristics, providing an independent verification. For example, many fine-scale spatial features, particularly at high northern latitudes, are present in both the GOME-2 SIF NR and FluxSat GPP datasets.
The contrast in standard deviations between noisy and NR SIF retrievals suggests that variability in the noisy retrievals is dominated by random errors, while variability in the NR retrievals may be more related to geophysical effects. The spatial features of variability are similar for the NR SIF retrievals and FluxSat GPP although there are some fine-scale differences. FluxSat GPP uses the clear-sky MCD43C nadir bidirectional reflectance distribution function (BRDF) adjusted reflectance (NBAR) dataset that has been adjusted to nadir view to remove variability due to BRDF effects. The variability in FluxSat GPP therefore is expected to be dominated by temporal variations in vegetation structure over the month. GOME-2 SIF should show similar variability due to changes in vegetation. However, SIF should show additional variability due to its dependence on the sun–satellite geometry as the scattering angles vary over a given location from orbit to orbit (Joiner et al. 2020). Gridded SIF data have also used less strict filters on cloud contamination as compared with NBAR data; effective cloud fractions of up to 30% are included in the SIF gridded dataset, and this contributes to gridbox variability. Therefore, we do not expect the standard deviations of SIF and GPP to be identical.
Figure 7 shows density maps of the NR versus the original noisy SIF gridbox monthly averages at 0.5° × 0.5° resolution over land from GOME-2B for June and December of 2014. We expect to see the scatter about the one-to-one line with little overall bias. This has been achieved to a large degree although again subtle differences are shown here including a slight underestimate of high SIF values in the Northern Hemisphere summer months and a slight overestimation of the low values. We also see about 75% less number of negative SIF gridbox mean values for the NR retrievals. The same data are plotted as probability distribution functions (PDFs) in Fig. A1 in the appendix.
Two-dimensional histograms (density maps) of gridded monthly mean NR SIF retrievals vs noisy (original) SIF retrievals at 0.5° × 0.5° resolution over land from GOME-2B for (a) June and (b) December in 2014, where colors represent the number of grid boxes, along with statistical metrics. N: number of grid boxes; R: correlation computed over all grid boxes; bias: mean SIF difference over all grid boxes; N_or (<0): number of gridbox averages less than zero in the original noisy data; and N_nr (<0): number of gridbox averages less than zero in the NR dataset (bias and SIF units are in mW m−2 nm−1 sr−1).
Citation: Artificial Intelligence for the Earth Systems 3, 3; 10.1175/AIES-D-23-0085.1
7) Interinstrument comparisons
Another means of assessing the quality of the NR retrievals is to compare results from two different instruments. Here, we compare results from GOME-2A and GOME-2B at a monthly time scale for which the effects of the different sun–satellite geometries for the two different satellites are mitigated by averaging over a range of angles. Figure 8 shows density maps of the monthly mean gridded SIF retrievals from GOME-2B versus GOME-2A in June and December of 2014 for original noisy and NR retrievals. There is improved agreement between the two instruments for NR SIF retrievals as compared with noisy retrievals. It should be noted that during this time, GOME-2A was operating in a reduced swath mode so that it experienced smaller view angles across its swath than GOME-2B that was operating in nominal swath mode. Therefore, GOME-2B observes scenes at its swath edge with higher phase angles than are experienced by GOME-2A (Joiner et al. 2020). Therefore, we may not expect perfect agreement between the two datasets.
As in Fig. 7, but showing gridded monthly mean SIF retrievals from GOME-2B vs those from GOME-2A in June and December 2014 for (a),(b) original (noisy) retrievals and (c),(d) NR retrievals (SIF units are in mW m−2 nm−1 sr−1).
Citation: Artificial Intelligence for the Earth Systems 3, 3; 10.1175/AIES-D-23-0085.1
Figure 9 shows the time series of SIF monthly means computed with the original noisy and NR retrievals from GOME-2A and GOME-2B for two 0.5° × 0.5° grid boxes. Gridbox monthly standard deviations are shown with vertical lines about the mean value. SIF values from the two instruments generally agree within their standard deviations but with much smaller standard deviations for the NR retrievals.
Time series of GOME-2A (black lines) and GOME-2B (pink lines) monthly mean SIF from original noisy and NR retrievals at selected 0.5° × 0.5° grid boxes along with their correlations R; gridbox standard deviations are shown with vertical lines above and below the monthly mean values. Shown with NR retrievals are time series of GPP (blue lines) from the independent FluxSat product based on MODIS reflectances aggregated to the same spatial and temporal resolution along with gridbox standard deviations. Correlations between GPP and NR GOME-2A or GOME-2B SIF monthly means are denoted R2A_GPP and R2B_GPP, respectively.
Citation: Artificial Intelligence for the Earth Systems 3, 3; 10.1175/AIES-D-23-0085.1
Seasonal cycles and interannual variability are much more apparent for the NR retrievals. There is better consistency between the two SIF datasets with the NR retrievals, lending further confidence in the method.
Time series of GPP from the independent FluxSat dataset are shown for comparison with the NR retrievals and display similar seasonal and interannual variability. Correlations are higher between GPP and SIF from each of the two GOME-2 instruments than between the SIF NR retrievals from the two GOME-2 instruments. The higher correlations likely reflect lower uncertainties in the FluxSat GPP dataset as compared with either SIF dataset. This analysis further supports that the temporal features brought out by the NR SIF retrievals are realistic.
To assess the time series globally, Fig. A2 displays maps of the gridbox SIF time series correlations between GOME-2A and GOME-2B noisy and NR retrievals for the overlap period of April 2013 through January 2018 with generally high correlations for NR retrievals over most moderately to highly vegetated areas. Figure A3 similarly shows maps of the gridbox correlations between NR SIF from the two GOME-2 instruments and GPP for the same time period as in Fig. A2 (April 2013–January 2018); these maps show similar spatial patterns as those of the correlations between the two GOME-2 SIF datasets in Fig. A2 (see the appendix).
c. Comparison with PCA spectral noise reduction
For comparison, we performed noise reduction on the observed radiances using PCA to reconstruct spectra with leading PCs followed by SIF retrievals with the noise-reduced radiances. We applied this approach to both simulated and real GOME-2 data. With simulated data, we obtained improved results for SIF retrievals with reconstructed spectra as compared with the original noisy retrievals. However, with real GOME-2 data, we found that in order to obtain acceptable results, we required the coefficients of at least half the total number of PCs (number of wavelengths). With less PCs, biases were present. The use of a larger number of PCs resulted in less bias in SIF retrievals but with limited noise reduction.
It appears that reconstructed spectra using PCA for noise reduction is ineffective for SIF retrievals because the very small SIF signal is spread out among a large number of PCs, including many that also contain a substantial amount of the random noise. For the same reason, applying our noise reduction methodology directly to the observed spectra rather than the SIF retrievals may not work well; we would not be able to capture the total SIF signal in the radiances without using a large number of PCs, and this would in turn limit noise reduction.
The good outcome with reconstructed spectra in simulated data presumably results from our inability to simulate all of the spectrally complex types of errors that can occur in real satellite data (e.g., stray light and other types of instrumental artifacts that lead to biases in the SIF retrievals). With only purely random noise in the simulation, the SIF signal is more cleanly packed into the leading PCs, and therefore, good results can be obtained using a smaller number of PCs.
The use of ML in our noise reduction approach allows for effective capture of the SIF signals in real satellite data with far fewer PC coefficients than is needed with reconstructed radiances. The NN effectively uses the information in fewer leading PCs to reconstruct the retrieved SIF. The use of fewer PCs leads to good reconstruction of the original signal with substantial noise reduction.
4. Conclusions
We provide a general framework for reducing the effects of random instrument noise in satellite SIF retrievals. While applied to the specific example of simulated and GOME-2 SIF datasets, the methodology should be applicable to other satellite SIF data products as well as other noisy remotely sensed data products. The amount of improvement depends on several factors including the SNR of the original product, the spatial and temporal scales involved, and whether or not additional spectral observations (outside of the fitting window used in the original retrieval) can be effectively utilized in the approach. Different optimizations of training datasets, neural network architectures, numbers of retained PC coefficients, etc., may be needed for application to other noisy data products.
The noise reduction on a daily time scale of more than a factor of 4 with GOME-2 data is equivalent to averaging approximately 16 observations. We therefore achieve equivalent data quality in 2 days with the noise-reduced data as we would have with approximately a month of the original noisy retrievals. With the substantial noise reduction, many applications of SIF data will be enabled or improved as less spatiotemporal averaging will be required for various analyses. For instance, with reduced noise, a much-improved characterization of the SIF variation with orbital geometry, similar to the BRDF, may be possible. Complex impacts of vegetation physiology and structure on the observed top-of-canopy or top-of-atmosphere SIF, such as characterization of the geometrical dependence of SIF, must be accounted for to effectively harness the information in SIF measurements such as extraction of information regarding physiological status (see, e.g., Porcar-Castell et al. 2021; Sun et al. 2023a).
Acknowledgments.
The authors thank the EUMETSAT science team who provided the level 1b GOME-2 data. This work was supported by NASA through the Making Earth System Data Records for Use in Research Environments (MEaSUREs) program. Coauthors at the Jet Propulsion Laboratory, California Institute of Technology, were supported under a contract with the National Aeronautics and Space Administration (80NM0018D0004). The authors thank Michael Yan for assistance with the archival of the noise reduced datasets at the Goddard Space Flight Center (GSFC) Aura Validation Data Center (AVDC).
Data availability statement.
The GOME-2A and GOME-2B raw data were obtained through the U.S. National Oceanic and Atmospheric Administration (NOAA) and are currently archived at the Comprehensive Large Array-Data Stewardship System (CLASS) at https://www.aev.class.noaa.gov/saa/products/welcome. The level 2 SIF products used in this work are archived at the Oak Ridge National Laboratory Distributed Active Archive Center (ORNL DAAC) (see references within the manuscript). The authors have published the noise-reduced GOME-2A and GOME-2B SIF daily and monthly mean retrievals at the GSFC AVDC at https://avdc.gsfc.nasa.gov through May 2021. Simulation data, noise reduction codes, and sample input and output datasets are available (Joiner 2024).
APPENDIX
Additional Analysis of GOME-2 SIF Data
Figure A1 shows PDFs of monthly mean gridded SIF from GOME-2B noisy and NR retrievals for June and December 2014. The tails of the distributions are smaller for the NR retrievals. The mode of the distributions is slightly higher (positive) for NR retrievals, whereas it is close to zero for the original noisy retrievals.
PDFs, fraction of grid boxes on the y axis vs monthly mean SIF values, derived from GOME-2B data for (a) June and (b) December of 2014 corresponding to the density plots shown in Fig. 7.
Citation: Artificial Intelligence for the Earth Systems 3, 3; 10.1175/AIES-D-23-0085.1
Figure A2 shows maps of gridded correlations of monthly mean SIF time series for GOME-2A and GOME-2B noisy and NR retrievals. Correlations are generally high (>0.85) for moderately to heavily vegetated regions, except for heavily clouded tropical regions such as Amazonia. Figure A3 shows similar spatial patterns of the correlation maps between SIF from both satellites and FluxSat GPP. SIF–GPP correlations are a bit lower than the intrasatellite correlations over cloudy Amazonia which could indicate cloud contamination of the MODIS reflectances used in FluxSat and southern Australia where FluxSat has known difficulties in reproducing the GPP seasonal cycle (Joiner and Yoshida 2020). Note that during the period of overlap between GOME-2A and GOME-2B used for all maps here (April 2013–January 2018), GOME-2A has less coverage over polar regions due to its reduced swath.
Gridded correlations of monthly mean SIF time series computed during overlap periods (April 2013–January 2018) between GOME-2A and GOME-2B for (a) original noisy retrievals and (b) NR retrievals.
Citation: Artificial Intelligence for the Earth Systems 3, 3; 10.1175/AIES-D-23-0085.1
As in Fig. A2, but showing correlations between SIF NR from (a) GOME-2A and FluxSat GPP and similarly from (b) GOME-2B and FluxSat GPP. Both maps use data from April 2013 to January 2018.
Citation: Artificial Intelligence for the Earth Systems 3, 3; 10.1175/AIES-D-23-0085.1
REFERENCES
Doughty, R., T. P. Kurosu, N. Parazoo, P. Köhler, Y. Wang, Y. Sun, and C. Frankenberg, 2022: Global GOSAT, OCO-2, and OCO-3 solar-induced chlorophyll fluorescence datasets. Earth Syst. Sci. Data, 14, 1513–1529, https://doi.org/10.5194/essd-14-1513-2022.
Du, S., L. Liu, X. Liu, X. Zhang, X. Zhang, Y. Bi, and L. Zhang, 2018: Retrieval of global terrestrial solar-induced chlorophyll fluorescence from TanSat satellite. Sci. Bull., 63, 1502–1512, https://doi.org/10.1016/j.scib.2018.10.003.
Fell, F., and J. Fischer, 2001: Numerical simulation of the light field in the atmosphere–ocean system using the matrix-operator method. J. Quant. Spectrosc. Radiat. Transfer, 69, 351–388, https://doi.org/10.1016/S0022-4073(00)00089-3.
Frankenberg, C., A. Butz, and G. C. Toon, 2011a: Disentangling chlorophyll fluorescence from atmospheric scattering effects in O2 A-band spectra of reflected sun-light. Geophys. Res. Lett., 38, L03801, https://doi.org/10.1029/2010GL045896.
Frankenberg, C., and Coauthors, 2011b: New global observations of the terrestrial carbon cycle from GOSAT: Patterns of plant fluorescence with gross primary productivity. Geophys. Res. Lett., 38, L17706, https://doi.org/10.1029/2011GL048738.
Gorkavyi, N., Z. Fasnacht, D. Haffner, S. Marchenko, J. Joiner, and A. Vasilkov, 2021: Detection of anomalies in the UV–vis reflectances from the Ozone Monitoring Instrument. Atmos. Meas. Tech., 14, 961–974, https://doi.org/10.5194/amt-14-961-2021.
Guanter, L., C. Frankenberg, A. Dudhia, P. E. Lewis, J. Gómez-Dans, A. Kuze, H. Suto, and R. G. Grainger, 2012: Retrieval and global assessment of terrestrial chlorophyll fluorescence from GOSAT space measurements. Remote Sens. Environ., 121, 236–251, https://doi.org/10.1016/j.rse.2012.02.006.
Guanter, L., and Coauthors, 2015: Potential of the TROPOspheric Monitoring Instrument (TROPOMI) onboard the Sentinel-5 Precursor for the monitoring of terrestrial chlorophyll fluorescence. Atmos. Meas. Tech., 8, 1337–1352, https://doi.org/10.5194/amt-8-1337-2015.
Guanter, L., and Coauthors, 2021: The TROPOSIF global sun-induced fluorescence dataset from the Sentinel-5P TROPOMI mission. Earth Syst. Sci. Data, 13, 5423–5440, https://doi.org/10.5194/essd-13-5423-2021.
Jacquemoud, S., W. Verhoef, F. Baret, C. Bacour, P. J. Zarco-Tejada, G. P. Asner, C. François, and S. L. Ustin, 2009: PROSPECT + SAIL models: A review of use for vegetation characterization. Remote Sens. Environ., 113 (Suppl. 1), S56–S66, https://doi.org/10.1016/j.rse.2008.01.026.
Joiner, J., 2024: Solar-induced fluorescence noise reduction using simulated and real data (v1.1). Machine Learning Based Noise Reduction for Satellite Products: Application to Solar-Induced Fluorescence Retrievals using Simulated and Real Data, San Francisco, CA, Zenodo, B51H-1870, https://doi.org/10.5281/zenodo.10724722.
Joiner, J., and Y. Yoshida, 2020: Satellite-based reflectances capture large fraction of variability in global Gross Primary Production (GPP) at weekly time scales. Agric. For. Meteor., 291, 108092, https://doi.org/10.1016/j.agrformet.2020.108092.
Joiner, J., Y. Yoshida, A. P. Vasilkov, Y. Yoshida, L. A. Corp, and E. M. Middleton, 2011: First observations of global and seasonal terrestrial chlorophyll fluorescence from space. Biogeosciences, 8, 637–651, https://doi.org/10.5194/bg-8-637-2011.
Joiner, J., and Coauthors, 2013: Global monitoring of terrestrial chlorophyll fluorescence from moderate-spectral-resolution near-infrared satellite measurements: Methodology, simulations, and application to GOME-2. Atmos. Meas. Tech., 6, 2803–2823, https://doi.org/10.5194/amt-6-2803-2013.
Joiner, J., Y. Yoshida, L. Guanter, and E. M. Middleton, 2016: New methods for the retrieval of chlorophyll red fluorescence from hyperspectral satellite instruments: Simulations and application to GOME-2 and SCIAMACHY. Atmos. Meas. Tech., 9, 3939–3967, https://doi.org/10.5194/amt-9-3939-2016.
Joiner, J., Y. Yoshida, Y. Zhang, G. Duveiller, M. Jung, A. Lyapustin, Y. Wang, and C. J. Tucker, 2018: Estimation of terrestrial global gross primary production (GPP) with satellite data-driven models and eddy covariance flux data. Remote Sens., 10, 1346, https://doi.org/10.3390/rs10091346.
Joiner, J., Y. Yoshida, P. Köhler, C. Frankenberg, and N. Parazoo, 2019: L2 Daily Solar-Induced Fluorescence (SIF) from ERS-2 GOME, 1995–2003. ORNL DAAC, https://doi.org/10.3334/ORNLDAAC/1758.
Joiner, J., and Coauthors, 2020: Systematic orbital geometry-dependent variations in satellite Solar-Induced Fluorescence (SIF) retrievals. Remote Sens., 12, 2346, https://doi.org/10.3390/rs12152346.
Joiner, J., S. Marchenko, Z. Fasnacht, L. Lamsal, C. Li, A. Vasilkov, and N. Krotkov, 2023: Use of machine learning and principal component analysis to retrieve nitrogen dioxide (NO2) with hyperspectral imagers and reduce noise in spectral fitting. Atmos. Meas. Tech., 16, 481–500, https://doi.org/10.5194/amt-16-481-2023.
Khosravi, N., M. Vountas, V. V. Rozanov, A. Bracher, A. Wolanin, and J. P. Burrows, 2015: Retrieval of terrestrial plant fluorescence based on the in-filling of far-red Fraunhofer lines using SCIAMACHY observations. Front. Environ. Sci., 3, 78, https://doi.org/10.3389/fenvs.2015.00078.
Köhler, P., L. Guanter, and C. Frankenberg, 2015a: Simplified physically based retrieval of sun-induced chlorophyll fluorescence from GOSAT data. IEEE Geosci. Remote Sens. Lett., 12, 1446–1450, https://doi.org/10.1109/LGRS.2015.2407051.
Köhler, P., L. Guanter, and J. Joiner, 2015b: A linear method for the retrieval of sun-induced chlorophyll fluorescence from GOME-2 and SCIAMACHY data. Atmos. Meas. Tech., 8, 2589–2608, https://doi.org/10.5194/amt-8-2589-2015.
Köhler, P., C. Frankenberg, T. S. Magney, L. Guanter, J. Joiner, and J. Landgraf, 2018a: Global retrievals of solar-induced chlorophyll fluorescence with TROPOMI: First results and intersensor comparison to OCO-2. Geophys. Res. Lett., 45, 10 456–10 463, https://doi.org/10.1029/2018GL079031.
Köhler, P., L. Guanter, H. Kobayashi, S. Walther, and W. Yang, 2018b: Assessing the potential of sun-induced fluorescence and the canopy scattering coefficient to track large-scale vegetation dynamics in amazon forests. Remote Sens. Environ., 204, 769–785, https://doi.org/10.1016/j.rse.2017.09.025.
Köhler, P., M. J. Behrenfeld, J. Landgraf, J. Joiner, T. S. Magney, and C. Frankenberg, 2020: Global retrievals of solar-induced chlorophyll fluorescence at red wavelengths with TROPOMI. Geophys. Res. Lett., 47, e2020GL087541, https://doi.org/10.1029/2020GL087541.
Köhler, P., W. W. Fischer, G. R. Rossman, J. P. Grotzinger, R. Doughty, Y. Wang, Y. Yin, and C. Frankenberg, 2021: Mineral luminescence observed from space. Geophys. Res. Lett., 48, e2021GL095227, https://doi.org/10.1029/2021GL095227.
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. J. Geophys. Res., 116, D03210, https://doi.org/10.1029/2010JD014985.
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. J. Geophys. Res., 116, D03211, https://doi.org/10.1029/2010JD014986.
Lyapustin, A. I., Y. Wang, I. Laszlo, T. Hilker, F. G. Hall, P. J. Sellers, C. J. Tucker, and S. V. Korkin, 2012: Multi-Angle Implementation of Atmospheric Correction for MODIS (MAIAC): 3. Atmospheric correction. Remote Sens. Environ., 127, 385–393, https://doi.org/10.1016/j.rse.2012.09.002.
Miller, J. R., and Coauthors, 2005: Development of a vegetation fluorescence canopy model. ESTEC Final Rep. 16365/02/NL/FF, 138 pp., https://www.ipgp.fr/∼jacquemoud/publications/miller2005.pdf.
Mohammed, G. H., and Coauthors, 2019: Remote sensing of Solar-Induced chlorophyll Fluorescence (SIF) in vegetation: 50 years of progress. Remote Sens. Environ., 231, 111177, https://doi.org/10.1016/j.rse.2019.04.030.
Munro, R., and Coauthors, 2016: The GOME-2 instrument on the Metop series of satellites: Instrument design, calibration, and level 1 data processing—An overview. Atmos. Meas. Tech., 9, 1279–1301, https://doi.org/10.5194/amt-9-1279-2016.
Parazoo, N. C., C. Frankenberg, P. Köhler, J. Joiner, Y. Yoshida, T. Magney, Y. Sun, and V. Yadav, 2019: Towards a harmonized long-term spaceborne record of far-red solar-induced fluorescence. J. Geophys. Res. Biogeosci., 124, 2518–2539, https://doi.org/10.1029/2019JG005289.
Pedrós, R., Y. Goulas, S. Jacquemoud, J. Louis, and I. Moya, 2010: FluorMODleaf: A new leaf fluorescence emission model based on the PROSPECT model. Remote Sens. Environ., 114, 155–167, https://doi.org/10.1016/j.rse.2009.08.019.
Porcar-Castell, A., and Coauthors, 2021: Chlorophyll a fluorescence illuminates a path connecting plant molecular biology to Earth-system science. Nat. Plants, 7, 998–1009, https://doi.org/10.1038/s41477-021-00980-4.
Preusker, R., and R. Lindstrot, 2009: Remote sensing of cloud-top pressure using moderately resolved measurements within the oxygen A band—A sensitivity study. J. Appl. Meteor. Climatol., 48, 1562–1574, https://doi.org/10.1175/2009JAMC2074.1.
Rasti, B., P. Scheunders, P. Ghamisi, G. Licciardi, and J. Chanussot, 2018: Noise reduction in hyperspectral imagery: Overview and application. Remote Sens., 10, 482, https://doi.org/10.3390/rs10030482.
Rothman, L. S., and Coauthors, 2009: The HITRAN 2008 molecular spectroscopic database. J. Quant. Spectrosc. Radiat. Transfer, 110, 533–572, https://doi.org/10.1016/j.jqsrt.2009.02.013.
Sanders, A. F. J., W. W. Verstraeten, M. L. Kooreman, T. C. Van Leth, J. Beringer, and J. Joiner, 2016: Spaceborne sun-induced vegetation fluorescence time series from 2007 to 2015 evaluated with Australian flux tower measurements. Remote Sens., 8, 895, https://doi.org/10.3390/rs8110895.
Schaaf, C. B., and Coauthors, 2002: First operational BRDF, albedo Nadir reflectance products from MODIS. Remote Sens. Environ., 83, 135–148, https://doi.org/10.1016/S0034-4257(02)00091-3.
Sun, Y., and Coauthors, 2017: OCO-2 advances photosynthesis observation from space via solar-induced chlorophyll fluorescence. Science, 358, eaam5747, https://doi.org/10.1126/science.aam5747.
Sun, Y., C. Frankenberg, M. Jung, J. Joiner, L. Guanter, P. Köhler, and T. Magney, 2018: Overview of Solar-Induced chlorophyll Fluorescence (SIF) from the Orbiting Carbon Observatory-2: Retrieval, cross-mission comparison, and global monitoring for GPP. Remote Sens. Environ., 209, 808–823, https://doi.org/10.1016/j.rse.2018.02.016.
Sun, Y., and Coauthors, 2023a: From remotely sensed solar-induced chlorophyll fluorescence to ecosystem structure, function, and service: Part I—Harnessing theory. Global Change Biol., 29, 2926–2952, https://doi.org/10.1111/gcb.16634.
Sun, Y., and Coauthors, 2023b: From remotely-sensed solar-induced chlorophyll fluorescence to ecosystem structure, function, and service: Part II—Harnessing data. Global Change Biol., 29, 2893–2925, https://doi.org/10.1111/gcb.16646.
Thakur, R. S., S. Chatterjee, R. N. Yadav, and L. Gupta, 2021: Image de-noising with machine learning: A review. IEEE Access, 9, 93 338–93 363, https://doi.org/10.1109/ACCESS.2021.3092425.
van Schaik, E., and Coauthors, 2020: Improved SIFTER v2 algorithm for long-term GOME-2A satellite retrievals of fluorescence with a correction for instrument degradation. Atmos. Meas. Tech., 13, 4295–4315, https://doi.org/10.5194/amt-13-4295-2020.
Wang, X., Z. Luo, W. Li, X. Hu, L. Zhang, and Y. Zhong, 2022: A self-supervised denoising network for satellite-airborne-ground hyperspectral imagery. IEEE Trans. Geosci. Remote Sens., 60, 5503716, https://doi.org/10.1109/TGRS.2021.3064429.
Yao, L., and Coauthors, 2022: Retrieval of Solar-Induced chlorophyll Fluorescence (SIF) from satellite measurements: Comparison of SIF between TanSat and OCO-2. Atmos. Meas. Tech., 15, 2125–2137, https://doi.org/10.5194/amt-15-2125-2022.