The time, wavelength, and scan-angle dependent degradation correction is derived from analysing daily global reflectance trends over time. First, GOME-2B reflectance data for the spectral range of 712–785 nm are collected between 60° S and 60° N, and with solar zenith angles below 85°. Scenes are not filtered on cloud conditions or sun glint, but data corresponding to static or narrow swath observations are excluded. Daily averages are then obtained for scan-index s (n=24) and detector pixel at wavelength λ.

As an illustration, Fig.  shows the daily global reflectances at λ= 747.2 nm and s=1 and s=24 in green. Different long-term temporal reflectance patterns are observed over the easternmost (s=1) and westernmost (s=24) pixel. The reflectance over the easternmost pixel shows a clear decreasing pattern over time, whereas it remains more stable over the westernmost pixel. The stronger eastward degradation is consistent with patterns observed in GOME-2A and predecessors GOME and SCIAMACHY . The degree of degradation, or signal attenuation, depends on the properties of the contamination layer that develops on the scan mirrors over time and its interaction with polarised light . Eastward light is likely more affected due to the higher degree of polarisation.

Next, we model the temporal variation in global mean reflectance (Rλ,s*(t)). Global mean reflectances are expected to vary seasonally due to changing geometry and scene observation. Although no major long-term trends are expected, substantial long-term trends are noted (Fig. ). Therefore, Rλ,s*(t) is represented by a combination of a polynomial Pλ,sp(t) and a finite Fourier series Fλ,sq(t): 1Rλ,s*(t)=Pλ,sp(t)[1+Fλ,sq(t)] Fλ,sq(t) and Pλ,sp(t) capture the seasonal variation and long-term trends, respectively. For GOME-2B, the Fourier order q is set to 6, and the polynomial degree p to 5. A high polynomial degree was necessary for the fit of GOME-2B reflectances to accurately capture the abrupt drop in signal from around 2020, particularly on the eastern side, as visible in Fig. . While geophysical changes, such as variations in cloud fraction, aerosols, and global greening, can affect long-term trends in global mean reflectance, their impact is expected to be minor compared to the observed trends and attributed impact of throughput loss following instrument degradation . Moreover, the pronounced scan-angle dependence strongly indicates an instrumental origin.

To obtain the correction factors (cλ,s(t)), Eq. () is fitted to the observed reflectances using least-squares regression. The obtained coefficients of the polynomial Pλ,sp are then used to correct the long-term trends. We scale the reflectance value at day t to the value at a reference day t0, as 2cλ,s(t)=Pλ,sp(t0)/Pλ,sp(t) With cλ,s(t) as correction factor at day t, wavelength λ and scan-index s. The reference day t0 represents a day early on in the mission, assuming no spectral degradation at that time. For GOME-2B, t0 is selected as 1 November 2012. The resulting correction factors are applied to each observed reflectance value before the SIF retrieval to counteract the identified degradation patterns.

Figure  shows the relative change in reflectance at ∼ 740.1 nm across the swath for multiple years, for both GOME-2A and GOME-2B. The reflectance degradation pattern represents an inconsistency in the relative loss of throughput between the observed radiance and solar irradiance. As reflectance is defined as the ratio of radiance to solar irradiance signal, a decrease in reflectance, for instance, indicates a stronger degradation of the radiance signal. In both sensors, reflectance degradation is more pronounced on the eastern side of the swath, and an increasing East-West bias develops over time. For GOME-2A, reflectances at all scanning positions increase over the first six years, then decline and eventually drop below the reference value. For GOME-2B, an overall decreasing trend in reflectances is noted.

These results confirm that the reflectance degradation of GOME-2B, similar to GOME-2A , is strongly scan-angle dependent and of the same order of magnitude as the temporal dependency. Importantly, the degradation trends substantially differ between sensors. Furthermore, when GOME-2B launched in 2013, GOME-2A had already drifted significantly from its original reflectance values, amplifying intersensor divergence.

The SIFTER retrieval algorithm obtains far-red SIF by making use of the relative infilling by fluorescence of solar Fraunhofer absorption lines near the 740 nm peak. A narrow retrieval window of 734 to 758 nm is used. This window represents the trade-off of minimising interference from water vapour and O₂ absorption features, while still capturing sufficient Fraunhofer lines and spectral points to ensure reliable SIF retrieval . However, some residual sensitivity due to inclusion of water vapour absorption features may remain under humid conditions. In this work, this sensitivity is reduced by using the SIFTER v3 algorithm, which includes an improved representation of atmospheric variability in the principal component construction . We use the latest reprocessed level-1b dataset, Release-3 (R3), as input. The R3 dataset ensures consistent processing and auxiliary data for GOME-2A and GOME-2B up to July 2020 . From July 2020 onwards, a different processor version is used, but this is not expected to introduce significant inconsistencies in the level-1b data.

The retrieval isolates the additional vegetation fluorescence signal from atmospheric features by matching a modeled reflectance spectrum to the observed spectrum. We model the reflectance (R) using a Lambertian surface reflectance model, as described by: 3R(λ,μ,μ0)≈as(λ)T↓(λ,μ0)T↑(λ,μ)+πISIF(λ)E0(λ)μ0T↑(λ,μ) where μ and μ0 are the cosines of the viewing and solar zenith angles, respectively. The surface albedo is denoted by as(λ), the SIF emissions from the vegetated surface by ISIF, and the atmospheric transmission –both downwards and upwards – by T↓ and T↑. The modeled reflectance contains 16 unknowns, one of which is the SIF signal to be extracted. These unknowns are solved by minimising the difference between the modeled and observed R using a Levenberg-Marquardt least-squares regression . Among the other 15 unknowns, five coefficients come from the surface albedo, which is estimated using a fourth-order polynomial, and ten are related to the atmospheric transmittance, which is characterised by 10 principal component (PC) functions.

The atmospheric transmittance varies with each scene and atmospheric conditions. To capture its variability, we apply principal component analysis (PCA) to a large selection of spectra across the Sahara region (16–30° N, 8° W–29° E). These observations are filtered for barren areas using land classification data to ensure the absence of vegetation and, therefore, SIF emission. We select GOME-2B spectra from five complete years (2013–2018), matching the period length used for the PC calculations of GOME-2A (2007–2012). Note that degradation corrected spectra are used (Sect. ) to avoid biases. Furthermore, as introduced in SIFTER v3 , the spectra are mean-centered and scaled by standard deviation before the PCA.

Figure  presents the first principal component (PC) for both GOME-2A and GOME-2B, and the cumulative explained variance for each PC (10 in total) for both datasets. For both sensors, PC #1 shows a similar overall structure, but the GOME-2A pattern is more sharply defined, with more pronounced features, such as deeper troughs and higher peaks. The less defined pattern in GOME-2B may reflect its generally higher reflectance uncertainty in the NIR and/or its slightly courser spectral resolution (0.5 nm vs 0.48 nm, Table ) compared to GOME-2A. Nevertheless, the PCs capture 99.95 % of the total variance, indicating that the main spectral structures are consistently represented in both sensors.

The SIFTER retrieval algorithm detects the SIF signal as variations in the relative depth of Fraunhofer lines; however, instrumental artifacts can also cause false “in-filling” or “deepening” of these lines and can therefore mimic fluorescence signals e.g.. These biases are latitude-dependent and may stem from temperature-driven changes in slit function width throughout the orbit, affecting the observed depth of Fraunhofer lines. Moreover, recent work by demonstrated that rotational Raman scattering can induce seemingly large SIF values. Such effects are particularly evident over oceans and deserts, where fluorescence should be near zero, as a zero-level offset.

Figure  shows the zero-level offset observed in GOME-2A and GOME-2B SIF over the Pacific Ocean (130–150° W), plotted by latitude. In GOME-2A SIF, a clear annual pattern of negative SIF values is noticeable, which shifts from north to south across latitudes. Additionally, the negative offsets intensify over time, with larger negative values observed in the later years. In GOME-2B, a different pattern is noticeable with strong positive offsets appearing early in the years ranging between 20° N–20° S. Additionally, GOME-2B shows strong variation between positive and negative offsets across latitude. Due to the differing patterns observed in both instruments, SIF values from GOME-2A and GOME-2B are not directly comparable. Consequently, it is necessary to apply a correction to account for these discrepancies, reduce intersensor biases, and ensure spatial consistency within and across the datasets.

To address the latitude bias effects and create higher consistency between GOME-2A and GOME-2B, we apply the post-hoc correction method from SIFTER v3 . This additive correction adjusts the retrieved SIF retrieval based on daily- and latitude-specific biases observed across reference areas over the Pacific and Atlantic Ocean. The correction is characterised as a function of reflectance (at 744 nm) per latitude band and on a daily basis, thereby accounting for brightness-dependent effects. The latitude bias correction brings SIF values over the Pacific Ocean region, within the expected near-zero range for both GOME-2A and GOME-2B (shown in Fig. S3). This confirms that latitude-dependent offsets were effectively corrected for, reducing the correlated intersensor divergence.

Here, we compare the spatial distribution of SIF as retrieved from GOME-2A and GOME-2B over the period when both sensors operated in tandem. For both datasets, valid SIF data were selected and seasonally averaged over the 2013–2017 period at a 0.5° × 0.5° spatial resolution. To avoid potential biases, we excluded the GOME-2B commissioning phase and only used complete seasons.

Figure  shows the mean SIF retrieved from GOME-2A and GOME-2B for the December–February (DJF) and June–August (JJA) seasons. Both datasets exhibit similar spatial distributions in SIF and align strongly with each other. However, slight differences are noted. GOME-2B SIF tends to be more negative over barren areas, such as Western China. Moreover, on average, GOME-2B SIF values exceed those of GOME-2A SIF by approximately 5 %–6 %. This positive bias appears to be most pronounced in regions with high vegetation activity, such as the Corn Belt region in JJA and the Amazon in DJF.

Aside from the differences in SIF values, GOME-2B also shows consistently higher uncertainty in SIF than GOME-2A, on the order of 17 %–18 %. This likely reflects a combination of larger uncertainties in the input reflectance, lower spectral resolution (Table ), and the less sharply defined principal components (Fig. ), all of which can propagate through the retrieval and increase the final uncertainty.

Overall, GOME-2A and GOME-2B SIF values agree well with consistent spatial patterns and strong correlations (r=0.96 for JJA, and r=0.97 for DJF). These results indicate coherence between both datasets. Nonetheless, GOME-2B SIF values are slightly but systematically biased against GOME-2A SIF values. In the following subsections, we examine to what extent this divergence reflects true biases or results from sampling differences.

During the operational tandem phase, GOME-2A and GOME-2B operated under different swath configurations. The swath reduction of GOME-2A limited the viewing zenith angle (VZA) range of the observations from ∈ [-52, +52°] to [-35, +35°] (Table ). As a result, the GOME-2B SIF averages shown in Fig.  include observations made under larger VZA angles (|VZA| > 35°) than GOME-2A. To understand how these discrepancies in viewing geometry might induce biases between GOME-2A and GOME-2B SIF, we first discuss the principle behind the angular dependence of SIF observations.

The SIF signal detected by the satellite sensor reflects the fraction of total emitted chlorophyll fluorescence that escaped the canopy and reached the sensor. This fraction depends on the photon scattering, leaf properties, and canopy architecture, influencing the propagation of SIF photons through the canopy e.g.,. As a result, the SIF signal is anisotropic and therefore depends on the viewing geometry between the Sun, vegetation, and the satellite sensor e.g..

Figure  illustrates the scan and illumination dependencies of the observed SIF by GOME-2. In the following, we describe each individual mechanism contributing to these dependencies on SIF. The relative importance of these mechanisms depend on the conditions, such as canopy structure and illumination geometry. At nadir, the sensor mainly detects photons emitted from the top of the canopy. Toward the swath edges, at larger viewing zenith angles (VZA), the sensor observes the canopy from a slanted angle, thereby enhancing the probability of detecting photons originating deeper within the canopy. Moreover, the viewing perspective determines whether the sunlit or shaded side of the canopy is viewed, leading to higher or lower observed values, respectively. For GOME-2A observations, the western ground pixels are typically sunlit, whereas eastern ground pixels tend to be shaded.

Wide-swath instruments, such as GOME-2 and TROPOMI, also introduce across-track variations in incoming solar irradiance . The wide swaths cover an extensive longitudinal range that spans different local solar times and thus solar zenith angles (SZA). As a result, GOME-2 observes eastern pixels later in the morning when solar illumination is typically higher (lower SZA), leading to potentially higher SIF values. Together, these viewing geometry factors can lead to systematic asymmetry in the observed SIF values across-track . The geometry effects on observed SIF are well known, and SIF is generally averaged over a sufficient number of observations to mitigate these effects e.g..

To assess the effect of discrepancy in viewing geometry between GOME-2A and GOME-2B, we average both datasets across scan positions. Figure  shows seasonal SIF as a function of VZA for both sensors across various geographical regions. A consistent across-track asymmetry is observed in all regions during the peak seasons, specifically JJA in the Northern Hemisphere and DJF in the Southern Hemisphere. During these high activity seasons, large SIF values near the eastern and, more prominently, the western edges of GOME-2B's 1920 km swath contribute to a higher overall average (x‾B), compared to GOME-2A SIF (x‾A). For instance, GOME-2B SIF has a bias of 13.9 % compared to GOME-2A SIF over Zambia. SIF values from both sensors align well when restricted to the shared VZA range of [-35, +35°]. The largest divergence across-track emerges when the |VZA| exceeds ∼ 30°. Within GOME-2B's wide swath, westward pixels can exhibit up to 35 % higher SIF values than eastward pixels during peak seasons (see Sect. S2.2 for more detail).

Figure  shows zonal averages of SIF from GOME-2A (in solid blue), GOME-2B using observations from the full VZA range (in solid green), and VZA ranges matched with GOME-2A (in dashed green). Overall, GOME-2A and GOME-2B SIF values converge when their VZA ranges match. For instance, the difference in DJF SIF between GOME-2A and GOME-2B over 0 to 40° S decreased from 13.8 % to 4.9 %. These results indicate that a significant portion of the inter-sensor differences in SIF can be attributed to variations in VZA sampling. Nonetheless, some divergence between GOME-2A and GOME-2B SIF remains. The remaining biases may result from differences in spatial and temporal sampling, as well as from differences in sensor and orbit characteristics.

To further evaluate the divergence between GOME-2A and GOME-2B SIF, we perform two co-sampling experiments (see Sect. S2.3). In Experiment 1, GOME-2A and GOME-2B pixels are spatially and temporally collocated. Matching pixel pairs had a maximum 50 km distance between their centers, at least 60 % spatial overlap of the smaller GOME-2A pixel with the GOME-2B pixel, and observation times less than 50 min apart. Experiment 2 applies the same constraints, with an additional requirement that both observations are within the same viewing geometry range of [-35, +35°]. These constraints help to isolate instrument-specific biases from those arising from sampling mismatches. Due to the limited spatial overlap of GOME-2A and GOME-2B ground pixels, sampling in Experiment 2 is practically limited to the Northern latitudes.

Figure  shows GOME-2A and GOME-2B SIF data, sampled to both experiments accordingly, and averaged over JJA 2014–2017. GOME-2A and GOME-2B SIF are better aligned for Experiment 2 (1.9 % bias) than for Experiment 1 (5.6 % bias) – as expected. This is confirmed by the empirical cumulative distribution functions (ECDFs) (bottom plots in Fig. ). ECDFs visualise the cumulative distribution of each dataset, enabling clear detection of systematic shifts between datasets. When only spatially and temporally co-sampled, Exp. 1, GOME-2B SIF values at the 90th percentile are 7.1 % higher than the corresponding GOME-2A value. Including co-sampling of viewing angles, Exp. 2, reduces the bias of GOME-2B SIF to GOME-2A SIF to 2.1 %.

Despite the reduced structural bias in Experiment 2, the correlation between GOME-2A and GOME-2B SIF is lower (r=0.92) than in Experiment 1 (r=0.97) (shown in Fig. S13). However, the SIF averages corresponding to Experiment 1 are based on nearly three times as many observations per grid cell as those in Experiment 2. This discrepancy in data density between the two experiments might reduce the comparability of their results. To enable a more balanced comparison, we constructed a reduced version of Experiment 1. The additional test involved averaging Experiment 1 data over a randomly selected subset of days from the JJA 2014–2017 period to match the data density of Experiment 2 better. The comparison between this reduced Experiment 1 and Experiment 2 was performed using only grid cells with valid and sufficient observations in both experiments. The divergence between GOME-2A and GOME-2B SIF remained similar to that of Fig. , but the correlation between GOME-2A and GOME-2B SIF for Experiment 2 (r=0.96) now exceeds that of Experiment 1 (r=0.95) (shown in Fig. S16). These results confirm that the co-sampling requirements as set in Experiment 2 result in the best agreement among GOME-2A and GOME-2B SIF. The details on this additional test are in Sect. S2.3.1.

GOME-2A and GOME-2B SIF show strong coherence in terms of spatial distribution. However, our results indicate the importance of similar sampling to reduce systematic biases between the two datasets. Particularly, discrepancies in viewing geometry sampling have been shown to distort the coherence between inter-sensor SIF datasets. When both datasets are sampled similarly, GOME-2A and GOME-2B SIF agree to within 2 %. A slight discrepancy is expected, as they are separate sensors with minor differences in instrumental characteristics (Table ). These results provide confidence in the consistency of GOME-2A SIF (pre-July 2013) and GOME-2B SIF (from July 2013 onwards), when both sensors operated with the same 1920 km swath and viewing zenith angle ranges (VZA ∈ [-52, +52°]).

Figure  shows the monthly regional averages of SIF from January 2007 to December 2023. For most regions, the transition from GOME-2A SIF (in blue) to GOME-2B SIF (in green) appears seamless. However, in the Amazon, a slight downward shift in GOME-2B SIF, relative to GOME-2A SIF, is noted. To test whether this shift is related to the instrument transition, we proceed with statistical and analytical tests.

To statistically investigate whether a structural break occurs at the transition month T0, July 2013, from GOME-2A to GOME-2B SIF data, we fit a simple model to the monthly SIF time series, which includes a linear trend (αt), seasonal component (St), and a mean level shift term (δUt): 4yt=μ+αt+St+δUt+εt where yt represents the regional SIF value at month t, μ the monthly mean and εt the residual of the observed and fitted SIF. This equation is adapted from and commonly used to model trends in environmental variables that include a sudden level shift e.g.. The seasonal component St represents the annual cycle of SIF as a first-order harmonic: 5St=β1sin⁡(2π(t-ϕ)12)+β2cos⁡(2π(t-ϕ)12) where β1 and β2 control the amplitude, ϕ is the phase shift, and the angular frequency corresponds to annual periodicity.

Before fitting the full model, the phase shift ϕ is fixed through a pre-optimalisation step. We estimated its value by evaluating the model with 100 evenly spaced values of ϕ, ranging from 0 to 1, and selecting the value under which the model performs best. This ensures good alignment of the modelled SIF with the real seasonal timing and avoids non-linearity in the model.

The possible offset in GOME-2 SIF related to the moment when the time series shifts from GOME-2A to GOME-2B data in July 2013, or T0, is accounted for by the step-change indicator Ut as: 6Ut=0,if t<T01,if t≥T0 In this equation, Ut is set to 0 during the GOME-2A period (January 2007–June 2013) and switches to 1 from July 2013 onwards during the GOME-2B period. Finally, Ordinary Least Squares (OLS) regression is used to fit the five remaining unknowns: μ, α, β1, β2, and δ.

We use the fitted coefficient δ and its p-value to evaluate the existence and magnitude of any step change at transition T0. Additionally, we apply both the Chow and the Likelihood Ratio (LR) test to evaluate the statistical significance of the potential break. The Chow test evaluates whether regression parameters differ before and after the breakpoint. It tests if the data structure is best described with one or two regressions by fitting the same model to (i) the full GOME-2 dataset, (ii) the data before T0 (GOME-2A), and (iii) the data after T0 (GOME-2B). The LR test complements this by comparing the regression results from the full model (including the δUt term) with a reduced model that excludes the step-change term. A significant p-value from the LR test indicates that including the step-change term significantly improved the fit.

Table  shows the fit and statistical test results for all cases. The modelled and observed SIF are strongly correlated (r≥ 0.91), implying that the model in Eq. () effectively captures the seasonal patterns and long-term trends necessary to detect structural breaks. In most regions, the step change coefficient δ is insignificant, indicating no jump in SIF from T0. However, significant step changes (p<0.05) are detected within the records across Eastern China and the Amazon. This bias is 0.13 ± 0.04 mW m⁻² sr⁻¹ nm⁻¹ over Eastern China and -0.10 ± 0.03 mW m⁻² sr⁻¹ nm⁻¹ over the Amazon region, respectively reflecting 16.6 % and 8.9 % of monthly averaged SIF over 2007–2023.

To interpret the detected offset bias δ, we evaluate its magnitude against the underlying uncertainty in SIF. Uncertainty in SIF retrieval can be distinguished into random errors and systematic errors. The first vary stochastically, arise from, for example, fit residual noise or sampling divergence, and decrease through averaging. Due to the spatial and temporal averaging applied here, random errors are small relative to the detected offsets. Systematic errors originate from algorithmic settings and persist despite averaging . To obtain a first-order estimate of these systematic errors, independently of regional offsets, we performed sensitivity tests over the Congo Basin, a region previously used as a sensitivity testbed by . We perturbed the settings of four distinct retrieval steps: (i) the degradation correction, (ii) the PC's, (iii) the interpolation across the slit function (discussed in more detail in ), and (iv) the latitude bias correction to assess their sensitivity. The largest sensitivity arises from the PCs used within the fitting model, particularly the number of PCs chosen to represent the atmospheric transmission. Varying the number of PC's from 10 to 6 led to a divergence of 0.47 mW m⁻² sr⁻¹ nm⁻¹ for GOME-2B SIF at 14 January 2017 over the Congo Basin. Combining the uncertainties from the four perturbed retrieval settings gives an illustrative value of 0.55 mW m⁻² sr⁻¹ nm⁻¹. Retrieval sensitivities of comparable order are expected in other regions, including Eastern China and the Amazon. Details on these tests are provided in Sect. S5.

The offsets δ detected in Eastern China (+0.13) and the Amazon (-0.10) are smaller than the estimated systematic error of 0.55 mW m⁻² sr⁻¹ nm⁻¹. This indicates that the offsets lie within the range of retrieval sensitivities, but they nevertheless represent a persistent inter-sensor difference. Additionally, the LR test confirms that the addition of an offset term enhances the model fit for these two regions (Table ). On the other hand, the Chow test results suggest the regression parameters remain consistent over the full GOME-2 record. These results imply that, while there is confidence in a mean-level bias at T0, the seasonal pattern and structure did not differ significantly before and after July 2013. Therefore, we add the found δ magnitudes as a correction term to the monthly regional SIF values during the GOME-2B period (from July 2013 onward) to align the records of GOME-2A and GOME-2B and support the use of both records as one consistent record. The corrected time series for Eastern China and the Amazon are shown and evaluated in Sect. . Both regions are characterised by higher atmospheric humidity. The enhanced retrieval sensitivity under such conditions may contribute to the observed intersensor difference. A full attribution of these remaining differences, however, is beyond the scope of this study and requires further investigation.

Finally, we verify the impact of the correction on the record's coherence by using an independent dataset as a reference. Specifically, we use FluxSat GPP data, which overlaps with the analysed period from January 2007 to December 2020 . FluxSat GPP is a satellite-derived global product that uses geometry-adjusted, daily-scaled MODIS MCD43D reflectance data and a machine learning approach to upscale eddy-covariance flux measurements from FLUXNET 2015. Assuming that FluxSat GPP correlates with SIF similarly across both sensors, it serves as a common reference to reveal intersensor biases. Each dataset – GOME-2 SIF with/without intersensor offset correction, and FluxSat GPP – is standardised over January 2007–December 2020 to enable cross-evaluation of their temporal variability. The standardisation relates the temporal variability to the data's standard deviation. We then compare the average difference between GOME-2 SIF and FluxSat GPP before and from T0 (July 2013), using both uncorrected and GOME-2 SIF corrected for the detected intersensor offset. The period before T0 covers January 2007 to June 2013, and the period from T0 covers July 2013 to December 2020.

Figure  shows the improvement in temporal consistency between GOME-2 SIF and FluxSat GPP over Eastern China and the Amazon after applying the intersensor offset correction. Before the correction, or the alignment of GOME-2A and GOME-2B SIF, the mean difference in SIF and GPP shifts substantially around T0. Since the datasets are standardised, this shift in SIF-GPP directly reflects the divergence between GOME-2A (<T0, January 2007–June 2013) and GOME-2B SIF (≥T₀, July 2013–December 2020) as a fraction of the total variability over time. For example, the alignment of SIF from both sensors reduced their divergence from 27.7 % to 4.6 % of the total variability over the Amazon region – a reduction of 23.1 %. This confirms that the application of the intersensor offset correction enhanced the temporal consistency within the combined GOME-2 record. Note that remaining differences around T0 (July 2013) may also reflect changes in the SIF–GPP relationships due to changed environmental conditions over time.

Overall, the presented approach to detect and correct for intersensor biases seems effective in enhancing the record's coherence. The methodology shown in this section can be used as a framework to detect, assess, and correct for potential biases between GOME-2A and GOME-2B SIF time series.

To evaluate whether the combined GOME-2 SIF records consistently track vegetation activity, we compared them to independent satellite-based proxies for photosynthesis: FluxSat GPP and TROPOMI SIF data obtained by . Since TROPOMI SIF data is available from early 2018, it is explicitly used for cross-evaluation against the GOME-2B SIF period. Both datasets are widely used to track inter-annual vegetation dynamics. True validation of SIF observation is restricted due to the lack of ground truth. Direct in situ validation is limited due to the mismatch in spatial resolution and the dependence of SIF values on observation time, viewing geometry, and the instrument's spectral characteristics . Therefore, cross-comparison with established independent datasets provides the common practice e.g.,. The correlations between GOME-2 SIF and independent data are shown in Sect. S4.

Figure  shows the regional monthly time series of GOME-2 SIF, TROPOMI SIF, and FluxSat GPP, with GOME-2 SIF corrected for intersensor offsets in Eastern China and the Amazon (Sect. ). All datasets were standardised to enable comparison. Across all regions, GOME-2 SIF consistently follows the seasonal cycle of FluxSat GPP with high correlations of r≥ 0.98 outside the Amazon and r=0.92 over the Amazon. The application of the intersensor offset correction showed a positive impact on these correlations with an increase from r= 0.97 to r=0.98 over Eastern China and from r= 0.91 to r= 0.92 over the Amazon (shown in Fig. S20). Although these improvements are modest, they nonetheless suggest that applying the bias correction enhances the temporal consistency.

GOME-2 SIF also agreed strongly with TROPOMI SIF (January 2018 to December 2022), with r=0.86 over the Amazon and r≥ 0.95 elsewhere. In Eastern China, GOME-2 SIF often reveals an early-season peak, which FluxSat GPP does not capture. This feature is also present in TROPOMI SIF (e.g., in 2022), suggesting that SIF is sensitive to subtle phenological features not reflected in reflectance-based GPP.

When evaluated separately, e.g. in Zambia, GOME-2A SIF generally demonstrates slightly higher correlations with FluxSat GPP (r=0.98) than GOME-2B SIF (r=0.94), resulting in an overall lower correlation for the combined record (r=0.96). While these analyses are based on different periods and should therefore be interpreted with caution, they suggest enhanced robustness within the GOME-2A SIF record. This is consistent with larger uncertainty within GOME-2B SIF as compared to GOME-2A SIF (Fig. ). Nevertheless, GOME-2B SIF maintains strong correlations with both FluxSat GPP and TROPOMI SIF, indicating that the underlying signals remain sufficiently robust.

Overall, these results indicate that the combined GOME-2 SIF record does not exhibit temporal biases or inconsistencies, including false trends induced by instrumental artifacts. Together with the findings by , this confirms the effectiveness of the advanced degradation correction applied by SIFTER v3 to enable robust and temporally consistent SIF records from both GOME-2A and GOME-2B SIF. When small intersensor offsets are present, our framework in Sect.  has proven effective in resolving these biases and enabling GOME-2A and GOME-2B to be treated as one long-term dataset for monitoring vegetation activity.