Tropospheric ozone is a regulated air pollutant and a greenhouse gas with a central role in chemical, dynamical, and radiative processes in the atmosphere. Changes in its concentrations affect the climate through radiative and biogenic feedbacks and pose risks to human and ecosystem health (). Although it has no direct emission sources, it is produced through photochemical reactions involving sunlight, volatile organic compounds (VOCs), and nitrogen oxides (NO_x) emitted by natural and primarly anthropogenic activities (). Its short lifetime, spatial variability of sources and sinks, and dependence on sunlight result in significant spatial and temporal variability ().

A key natural source of tropospheric ozone is atmospheric circulation (), particularly through Stratospheric-Tropospheric Exchange (STE) processes, which facilitate the bidirectional transfer of air masses, gases, and aerosols between the stratosphere and troposphere. STE dynamics, driven by mechanisms like Brewer-Dobson circulation, tropopause folding, and atmospheric waves, significantly influence tropospheric ozone concentrations (). Stratospheric contributions to tropospheric ozone account for 25 %–50 % in the extratropics, with the strongest influence in the Southern Hemisphere (). Low stratospheric ozone explains up to 70 % of the variability in mid-tropospheric ozone at mid-latitudes (). A global overall statistically significant increase in tropospheric ozone has been reported (), and mainly attributed to changes in atmospheric circulation and the strengthening of the Brewer-Dobson Circulation (BDC) ().

The Himalayan region is particularly affected by stratospheric intrusions due to its high altitude, complex topography, and dynamic meteorological conditions (). Processes such as tropopause folding, monsoonal convection, and interactions with the subtropical jet stream facilitate the downward transport of ozone-rich air masses. Studies have reported significant stratospheric contributions to surface-level ozone in this region, with a 25 % contribution detected during winter (). The rise in tropospheric ozone levels near the southern flank of the Himalayas has also been associated with increased stratospheric influence (). Despite these observations, substantial uncertainty remains in explaining surface ozone variations due to discrepancies between models and observations.

Our current understanding of the spatial and temporal climatology of tropospheric ozone is closely related to our ability to measure and simulate it. In this context, determining STE contribution to tropospheric ozone concentrations is crucial. Ozone observations are currently performed using ozonesondes and other in-situ and remote sensing instruments. These measurements are often combined with statistical models based on ozone tracers (e.g.beryllium (Be), lead (Pb), and radon (Rn)) (). While in-situ measurements are limited in spatial and temporal coverage, remote sensing observations offer a broader perspective, despite the constraints introduced by the inversion techniques required for their retrieval. Dynamical and chemical models are also widely used to simulate STE events and are often compared with observations; however, systematic biases between models and observations still exist (). Understanding STE processes, particularly in regions such as the UTLS, requires high vertical resolution to resolve strong gradients. Horizontal sampling also plays a role, as the events often occur over spatially confined regions. Satellite remote sensing instruments offer complementary strengths: limb-sounding sensors provide high vertical resolution in the stratosphere with limited sensitivity in the troposphere and limited horizontal coverage, while nadir-viewing sensors offer dense horizontal sampling and coverage of the tropospheric column, but with coarser vertical resolution. Therefore, combining coincident measurements from both observation geometries can improve the overall representation of ozone profiles.

In this context, we combine limb observations from the Michelson Interferometer for Passive Atmospheric Sounding (MIPAS), operating aboard ESA's Envisat between 2002 and 2012 (), with nadir observations from the Infrared Atmospheric Sounding Interferometer (IASI), active since 2006 aboard EUMETSAT's MetOp satellite series (). Both instruments operate in the mid-infrared (MIR) spectral range and provide valuable, complementary information on atmospheric composition. To optimally exploit their synergy, we apply the a posteriori Complete Data Fusion (CDF) methodology (). The choice of this combination is motivated by two main factors: (i) the authors have privileged access to MIPAS data allowing reprocessing with Optimal Estimation (OE) and access to detailed metadata, while IASI retrievals are based on OE and provide all elements required by the CDF; (ii) previous work () demonstrated the potential of the Measurement Space Solution (MSS; ) method – equivalent to CDF () – in improving vertical resolution and reducing uncertainties when applied to simulated measurements from these instruments and confirmed this with an example of a real profile. A more recent study () has extended this approach to simulated data from possible successors of MIPAS and IASI, comparing 2-dimensional vs. 1-dimensional retrieval and demonstrating how the 2-dimensional combination of new generation limb and nadir observations allow to improve performances in the UTLS and in the middle and lower troposphere. Building on this foundation, we apply the latest version of the CDF method () extensively to real measurements for the first time, thereby assessing its practical feasibility and scientific value.

In this paper, we present the development and validation of a novel 2008–2011 ozone dataset, created using the CDF methodology applied to MIPAS and IASI observations, and assess its quality using ozonesonde data. The Himalayan region is used as a reference and it is the focus area in part of the analysis due to its relevance for STE processes, which motivate this study. The main objective of this work is to exploit the propagation of vertical information from high-quality limb measurements in the stratosphere down to the troposphere, using the CDF approach. This allows us to extend the vertical resolution and sensitivity of limb data into regions where only nadir observations are available. In this lower region, MIPAS sensitivity decreases, retrieval uncertainties increase, or MIPAS data may be missing, whereas IASI provides more consistent coverage. This approach leverages the ability of MIPAS to resolve the UTLS region and the broader horizontal coverage and near-surface sensitivity of IASI.

The paper is structured as follows: Sect. 2 outlines the data and methods, including descriptions of the input datasets, a summary of the fundamental principles behind the CDF approach, algorithm choices, and the validation strategy. Section 3 presents the resulting dataset and discusses its validation and inter-comparison. Finally, Sect. 4 summarizes the main conclusions of this work.

As previously mentioned, the tuning phase was performed using data from 2008 over the 5–55° N latitude band. This phase focused on the initial validation of the different versions of the fusion method against ozonesonde profiles, on the analysis of DOFs, and on the retrieval errors of the fused and L2 products.

Figure  shows the validation results for the nine proposed versions of the fusion method. All versions exhibit a bias peak around 10 km, which is mainly inherited from the IASI dataset due to its vertical sensitivity structure and retrieval characteristics, as explained at the end of this section. However, this peak is reduced (by less than 25 %) in version v05. Bias alone is not sufficient to determine the best-performing version. Due to the different constraints applied in each version, some profiles exhibit oscillatory behavior or even negative values. These negative values do not have physical meaning; they arise from the retrieval process, particularly in regions of low sensitivity or very low ozone abundance, and reflect the limitations of the inversion algorithm combined with measurement noise. Although such values are occasionally kept to mantain statistical integrity of the dataset, their occurrence rate was adopted as a further criterion for assessing the quality of each version. Consequently, each version may have a different number of valid profiles. The quantity n in Fig.  shows, for each version, the number of profiles contributing to the mean. More strongly constrained versions generally exhibit fewer negative values. However, a stronger constraint typically reduces the information content of the profiles, and thus the number of DOFs. To improve the statistical robustness of the analysis, this aspect together with DOFs was also assessed using a larger dataset, selecting six random days in 2008 and including all fused profiles within the region, rather than only those coincident with ozonesondes.

Table  presents, for each version, the total number of profiles considered, the percentage of fused profiles containing at least one negative value, and the mean number of DOFs, both over the full vertical range and below 20 km. These metrics are reported for the fused product as well as for the original Level 2 datasets. Since the original Level 2 products are defined on different vertical grids and may rely on different a priori information, their DOFs are not directly comparable. To enable a consistent comparison, MIPAS profiles were interpolated onto the common fusion grid and a shared a priori using the CDF algorithm. This step harmonizes vertical resolution and alignment, reducing discrepancies due to differing retrieval configurations and enabling a meaningful comparison of the DOF values across datasets.

Among all versions, v00, followed by v05 and v06, shows the highest percentages of profiles with at least one negative value. Nevertheless, these versions, together with v01, perform best in terms of DOFs and agreement with MIPAS and IASI. This is especially evident for v05, which shows the highest DOFs across the full vertical range and a notable gain in DOFs (+0.67) over MIPAS, while also offering a good compromise in the lower atmosphere (+0.68 below 20 km). For MIPAS, values below 5 km are reconstructed using the CDF algorithm, since the instrument does not provide measurements below this altitude. Therefore, the actual improvement in vertical information content should be primarily assessed with respect to IASI, which provides 3.29 DOFs over the full profile, of which 1.85 are below 20 km.

These results clearly demonstrate how limb measurements can significantly improve nadir observations by increasing the vertical information content. Notably, the increase in DOFs is observed not only in the altitude range directly sampled by MIPAS, but also in regions where MIPAS does not provide measurements. This is due to the vertical correlations introduced by the fusion process, which propagate information from the limb instrument across the vertical profile. This outcome is fully consistent with one of the key objectives of our work, namely to exploit the complementary sensitivities of limb and nadir sensors to enhance the vertical information content of the resulting product, especially in the troposphere and in the UTLS. Furthermore, by fusing individual MIPAS and IASI retrievals, we can isolate and assess the specific contribution of each instrument to the enhanced information content, providing a detailed understanding of how the synergy between the two geometries improves the retrieval.

Finally, Fig.  presents the total retrieval error – propagated from both random and systematic components – for all CDF versions. A consistent error peak is observed at approximately 10 km. This peak arises from the changing influence of the IASI observation versus its a priori. Above 10 km, the measurement provides significant information, reducing errors. Below, the retrieval is governed by the a priori, resulting in higher errors. The 10 km altitude represents the transition between these two regimes, and the associated shift in information dominance induces a discontinuity in the derivative of the error, manifesting as the peak.

Based on all the considerations discussed above, version v05 was selected for further analysis, as it offers the most balanced trade-off, especially with regard to bias, DOFs, and retrieval errors. Version v05 was subsequently applied to the remaining years. Figure  shows the comparison between fused and L2 for 2008–2011.

As can be seen, the fused product shows an improvement over the L2 products across all years, starting from the fact that it provides information from the surface to high altitudes. Based on a range-dependent analysis, the following considerations can be drawn.

Above 30 km, the difference with ozonesondes increases, reaching a positive bias exceeding 100 %. However, at these altitudes, the accuracy of ozonesonde measurements decreases due to the very low pressure, and this bias should not be considered fully reliable (). Between 15 and 30 km, the fused profile shows smaller differences with ozonesondes than MIPAS, and even more so compared to IASI. In this altitude range, the fusion is mostly driven by MIPAS, leading to improved performance relative to IASI. Notably, the fused DOFs are equal to those of MIPAS, indicating that the fusion does not increase the retrieval sensitivity in this range. However, the influence of IASI is still evident in the comparison with ozonesondes, indicating a potential impact on the profile shape or bias characteristics. Between 20 and 30 km, MIPAS contributes the largest number of measurements. Below this range, it decreases, as shown in the right panels of Fig. .

Below 15 km, the contribution of IASI to the fused product increases, and the fused profile tends to follow the IASI profile. Below 5 km, since MIPAS provides no measurements, the information comes almost entirely from IASI. Nonetheless, even below 15 km, the information propagated from higher layers, where MIPAS data is available, improves the IASI-based retrieval.

All these aspects become clearer and more quantifiable when examining the DOFs distributed over multiple altitude ranges for the fused and L2 products over years (Table ).

Finally, the comparison between L2 and fused mean total errors (Fig. ) shows that the fused product exhibits lower errors than each L2 product in the UTLS region, and more generally below 18 km in all years. Above 18 km, the fused error is much lower than that of IASI and comparable to that of MIPAS, where MIPAS performs best. The slightly higher fused error compared to MIPAS is likely due to additional uncertainties introduced during CDF processing, such as co-location and interpolation errors. The error profile is fundamentally shaped by the limitations and synergies of the individual sensors. The prominent peak at 10 km is a feature of the IASI retrieval, arising from its reliance on a priori information at lower altitudes. Although this peak is imposed on the regridded MIPAS data during comparison, the fused product successfully overcomes this limitation. The synergy of IASI and MIPAS not only diminishes the peak but also achieves lower errors above 10 km, where MIPAS measurements provide a substantial improvement.

Validation over the 5–55° N latitudinal band for the remaining years 2009–2011 confirms the key features of the fused dataset already highlighted previously (Fig. ). The fused product exhibits a significant bias across most altitude levels. Between 10 and 30 km, the bias is positive and always below 25 %, peaking around 10 km. Below 8 km, the fused product shows a significant negative bias, also remaining within 25 %.

The validation across three latitudinal bands for the same years (Fig. ) allows assessment of the spatial variability of the bias in the fused product, although this analysis is limited by the number of coincident observations available in each region.

In the latitude band −75 to −25° S, which covers most of the Southern Hemisphere including Antarctica and sub-Antarctic regions, the statistics based on 28 profiles show a positive bias of less than 25 % throughout the vertical range, although it is not significant below 8 km and between 20–30 km. A significant peak of approximately 35 % is observed around 10 km.

In the −25 to 25° latitude band, which includes tropical and subtropical regions, the statistics rely on 10 coincidences, and the bias shows much stronger variability. It is negative below 12 km and becomes significant around 10 km, reaching a minimum of approximately −35 %. Between 12 and 18 km, the bias exceeds 100 %, becoming very large. Above 18 km, the bias is no longer significant. However, due to the limited number of coincidences, no definitive conclusions can be drawn.

In the 25 to 75° N band, representing the Northern Hemisphere (temperate and sub-Arctic regions), the number of coincidences increases to 71. The statistics indicate that the bias is not significant across most of the vertical profile: it is negative and below 25 % below 8 km, and positive above 8 km, with a peak of about 25 % near 10 km.

These results suggest that the fused product improves upon the individual products in terms of bias over the entire vertical profile. However, a consistent feature across all regions is the presence of a bias peak around 10 km, which corresponds to the altitude range typically associated with stratosphere-to-troposphere intrusion events. This feature, as explained above, appears to originate from IASI, as can be observed in Fig.  and is also consistent with the findings of , who reported a similar bias in the FORLI IASI dataset.

This observation motivated a bias correction of the IASI input profiles prior to applying the CDF algorithm. The fusion algorithm is mathematically robust to such modifications. Therefore, we applied a correction to each IASI profile for the year 2008 by subtracting the percentage bias identified in the IASI validation over the 5–55° N region.

The new validation for 2008 (Fig. ) shows that, compared to the results presented in Fig. , correcting the IASI bias leads to an improvement in the fused product.

The new validation for 2008 (Fig. ) shows that correcting the IASI bias significantly improves the fused product: the peak near 10 km is greatly reduced and becomes statistically non-significant. Applying the same IASI bias correction derived for 2008 to the remaining years yields the results shown in Fig. . It can be seen that the 2008 IASI bias also improves the fused product for the other years, although to varying degrees. This suggests that the IASI bias changes from year to year.

Nevertheless, this correction preserves the key characteristics of the fused product in terms of DOFs and retrieval errors, while considerably reducing the bias – especially in the UTLS region, where it becomes nonsignificant.

A new ozone dataset has been developed through the CDF algorithm, combining limb observations from MIPAS with nadir observations from IASI. The primary objective is to support the study of stratospheric intrusions (SIs), which require high-resolution measurement both vertically and horizontally, especially in the UTLS. The fusion targets the overlapping operational period of the two instruments (2008–2011), during which ozone profiles retrieved with OE are available, with a particular focus on the Himalayas, a hotspot for SI events. In this context, the aim of using CDF applied to MIPAS and IASI is to evaluate and enhance the propagation of information from high-quality limb measurements in the stratosphere down to the troposphere, where the MIPAS sensitivity decreases and retrieval uncertainties increase, or MIPAS data are unavailable – yet IASI provides coverage.

After tuning the algorithm on data from 2008, the dataset was validated over the remaining years using ozonesondes from WOUDC stations. The validation was performed both within the 5° S–55° N band and across three latitudinal bands: -75 to -25° S, -25° S to +25° N, and +25 to +75° N. The final fused product is a L2-like dataset composed of ozone profiles in partial columns, including associated CMs, AK, and a priori information (profile and CM), all mapped onto the IASI grid. The IASI a priori information was also used as a constraint in the fusion process.

The best-performing CDF configuration was the one that models the interpolation error using the a priori CM of each input and the coincidence error using the fused a priori CM. We evaluated the performance of this configuration in terms of bias against ozonesondes, DOFs, and errors in 2008, and extended the comparison to the entire 2008–2011 period in terms of DOFs and L2 profile differences. In general, the fused product shows increased information content compared to the individual input profiles, with improved DOFs and reduced uncertainty. Results show how MIPAS information is transferred to the lower layers through the application of CDF, also where MIPAS does not measure. This turns out to be in line with the goals of this work.

Validation results over 2009–2011 in the 5° S–55° N band show a statistically significant bias at most altitude levels. A positive bias (up to 20 %) is observed above 8 km, peaking around 10 km, while a negative bias (reaching -25 %) is found below 8 km. A similar behavior is seen in the other latitudinal bands, with notable variations: (1) in 75–25° S, the bias is positive at all altitudes, with a peak at 35 % around 10 km; (2) in 25° S–25° N, the bias exhibits strong oscillations but results are not considered reliable due to a low number of coincidences; (3) in 25–75° N, a significant positive bias (25 %) is observed above 8 km, while a significant negative bias (-20 %) is found below.

In most cases, the fused product shows a consistent positive peak (25 %) around 10 km. Comparison with the L2 IASI product suggests that this bias is inherited from IASI, consistent with findings from . Assessing the robustness of the fusion algorithm to input bias, we corrected IASI profiles based on the 2008 bias and reran the fusion. This correction reduced the fused bias to 10 %, making it statistically insignificant across most of the profile. Applying the same correction to the other years showed improvements, though analysis indicates a temporal variability in the IASI bias. This variability – both temporal and latitudinal – is also supported by , and warrants further investigation. However, it highlights the potential to further refine the dataset, particularly in reducing the bias in the UTLS.

The ozone dataset developed in this study provides a new tool for improving the representation of atmospheric ozone. It is worth emphasizing that the potential for assimilating fused data products has been thoroughly investigated within the AURORA project (Advanced Ultraviolet Radiation and Ozone Retrieval for Applications), where numerical experiments using the CDF algorithm demonstrated the technical feasibility of assimilating such products ().

This paves the way for further exploring the potential of the above-mentioned dataset of fused ozone profiles for data assimilation in reanalyses such as ERA5 and CAMS, which are known to exhibit biases in both tropospheric and stratospheric ozone due to uncertainties in assimilated observations and limitations in capturing ozone dynamics and chemistry accurately . Additionally, this dataset can serve as a valuable reference for detecting discrepancies in the physico-chemical modeling of ozone both in reanalyses and in coupled chemistry–climate models.

Finally, this dataset lays the foundation for future improvements in atmospheric reanalyses, modeling, and observational strategies. It also supports air quality studies and contributes to addressing key challenges, such as quantifying the stratospheric contribution to tropospheric ozone, particularly in hotspot regions like the Himalayas.