Random Forest Classifier for Cloud Clearing of the Operational TROPOMI XCH4 Product

1. Introduction

Methane (CH₄) is the second most important anthropogenic greenhouse gas after carbon dioxide (CO₂), surpassing CO₂ in its heat-trapping effectiveness [1]. Human activities, such as livestock digestion, rice cultivation, and the fossil fuel industry, are the primary sources of anthropogenic CH₄ emissions [2]. Satellite-based CH₄ measurements play an important role in identifying emission sources, quantifying sinks, and devising strategies to mitigate climate change [3,4].

The Tropospheric Monitoring Instrument (TROPOMI) on ESA’s Sentinel 5 Precursor satellite has global coverage within a day and provides the dry air mole fraction averaged methane columns XCH₄ for clear-sky scenes with a spatial resolution of up to 5.5×7 km². TROPOMI is a grating spectrometer with a wide spectral coverage including ultraviolet (UV), visible (VIS), near infrared (NIR) and short-wave infrared (SWIR) [5]. The TROPOMI XCH₄ data product is retrieved from the instrument’s NIR and SWIR measurements by deploying the RemoTeC algorithm [6]. It’s worth noting that due to algorithmic limitations on the viewing zenith angle, the swath is narrower for TROPOMI XCH4, resulting in a lack of full daily global coverage.

During recent years, the TROPOMI XCH₄ product was continuously improved; e.g., a new bias correction scheme was introduced to correct the dependence of XCH₄ on the brightness of the scene [7]. The data set was extended to include observations over oceans under glint geometry [8] and the identified XCH₄ anomalies over carbonated rock formations were solved by improving the spectral fit of the surface reflectivity [9]. The TROPOMI XCH₄ dataset is widely used, for example, for the detection of anthropogenic CH₄ emissions and atmospheric modeling [3,4,10].

The TROPOMI XCH₄ data product meets the stringent mission requirements on precision (<1.5%) and accuracy (<1%). Extensive validation efforts using ground-based FTIR measurements from the Total Carbon Column Observing Network (TCCON), have confirmed that the errors of the TROPOMI XCH₄ data are in line with these demands [11]. Achieving such precision necessitates rigorous cloud-clearing of the TROPOMI measurements, as clouds in the satellite’s observation path can introduce significant retrieval errors. Although the S5-P satellite itself lacks a dedicated cloud imager, it flies in loose formation with the SUOMI-NPP (SUOMI-National Polar-orbiting Partnership) satellite. This mission synergy allows us to use the VIIRS (Visible Infrared Imaging Radiometer Suite) cloud data for cloud clearing of TROPOMI measurements, since both instruments observe the same ground scene within < 3 minutes [12].

Here, the VIIRS data serve a dual role in ESA’s data processing workflow. Initially, they are employed for pre-filtering TROPOMI data, reducing the volume of data to be processed—a step representing weak data filtering. The more demanding second step involves cloud clearing, where VIIRS data contribute to a posteriori data filtering of the retrieved XCH₄ data, ensuring the high quality of the data product. This study primarily focuses on the second step of cloud filtering because it is more demanding and interesting from the scientific point of view. However, we also applied our method on the pre-filtering step and it is implemented in ESA’s processing framework.

In the event of missing VIIRS data, a backup filter was established for the TROPOMI XCH₄ retrieval, relying solely on TROPOMI measurements. This alternative filter incorporates information from both weak and strong CH₄ absorption lines in the shortwave infrared, which are included in the output of the TROPOMI CO product [12]. A similar approach, for instance, was applied by [13] for the OCO-2 instrument. However, our investigation revealed that the performance of the backup filter is found to be insufficient. It is crucial to highlight that these filtered thresholds were defined preflight and have not undergone testing on real data.

SUOMI-NPP is approaching its end of life and will be succeeded by the NOAA-20 satellite [14]. Although NOAA-20 carries a cloud imager equivalent to VIIRS, the different orbit positions cause a time gap of more than 20 minutes with TROPOMI, which presents a challenge for critical cloud clearing of TROPOMI data. A malfunction of the VIIRS instrument of a full month 2022 (August) and 7 days in November 2023 clearly demonstrated the dependence of the TROPOMI XCH₄ data quality on the VIIRS data product. The daily distribution of XCH₄ values in regions such as North America, Siberia, and Australia showed a significant low bias, due to the inclusion of cloud-contaminated scenes. Apparently, in August, the global XCH₄ distribution was skewed towards lower values, impacting both the mean by 3.9% and the standard deviation by 303% for the three regions while introducing numerous outliers. Hence, this clearly showed that the current backup filter is not sufficient and an alternative cloud-clearing of TROPOMI XCH₄ data is needed that is complementary to the SUOMI-NPP measurements.

This study introduces an innovative machine learning approach for cloud-clearing of the TROPOMI XCH₄ data product using a random forest classifier (RFC). Trained on five years of collocated measurements from both TROPOMI and SUOMI-NPP data (about 20000 orbits), the RFC can replace the VIIRS cloud-clearing process relying solely on TROPOMI data. Therefore, it represents an alternative cloud-clearing in the absence of VIIRS data and so can solve data processing issues due to the expected end-of-life of SUOMI-NPP. Moreover, the RFC approach is an essential step toward a near real-time TROPOMI XCH₄ data product, as it eliminates the need to await the availability of VIIRS data. A near real-time XCH₄ data product is requested for chemical forecasting of the atmosphere, as done by the Copernicus Atmosphere Monitoring Service (CAMS) and the Integrated Forecasting System (IFS) developed by European Centre for Medium-Range Weather Forecasts (ECMWF). It focuses on monitoring and forecasting atmospheric composition, including greenhouse gases, aerosols, and reactive gases and already assimilates TROPOMI CO in near-realtime [15].

Our study is structured as follows. In Section 2, we discuss the datasets utilized in our research and Section 3 explains our machine learning approach. In Section 4, we apply the RFC to address the one-month absence of VIIRS data in August 2022. Furthermore, a validation is presented for measurements at 12 TCCON stations. Finally, Section 5, summarizes our findings and draws conclusions based on our research.

2. Data

2.1. SUOMI-NPP VIIRS

The SUOMI-NPP VIIRS cloud data is our reference standard for cloud clearing. SUOMI-NPP is a collaborative satellite mission led by NASA (National Aeronautics and Space Administration) and NOAA (National Oceanic and Atmospheric Administration), and was successfully launched in 2011 [16]. Among its suite of instruments, the VIIRS (Visible Infrared Imaging Radiometer Suite) provides essential data, including cloud-related parameters. VIIRS cloud information is sampled on the TROPOMI footprints [17], providing valuable information on cloud coverage of TROPOMI observations with a minimal time delay of <3 minutes [12]. Resampling of VIIRS data on the SWIR footprints of TROPOMI introduces a processing delay of approximately 1.5 days. The study used the resampled VIIRS data product version 1.0.3. All threshold values for the VIIRS data filtering are sourced from [18].

2.2. TROPOMI CO

To train a RFC algorithm on the relationship between TROPOMI and resampled VIIRS data, we must select TROPOMI data with sufficient cloud sensitivity sampled on the SWIR pixel mesh. First, we use the CH₄ column retrieved from both weak and strong CH₄ absorption in the SWIR band using a non-scattering radiative transfer model [12]. The differences in columns indicate the presence of atmospheric scattering and were suggested by [13] for cloud filtering. Furthermore, we employ the TROPOMI CO product, which is also retrieved from TROPOMI SWIR observations. Here, the SICOR algorithm [19] infers the total column density together with effective cloud parameters. Here, prior knowledge of atmospheric CH₄ concentration is used to derive cloud information. This improves CO coverage with data quality that is well within mission objectives [20,21,22]. The data product includes the column averaging kernel [23] that describes the vertical sensitivity of CO recovery and thus indicates the presence of clouds. Here, the ground value of the total column averaging kernel is particularly sensitive to cloud contamination. Under clear-sky conditions, its value approaches about 0.9, gradually decreasing to 0 in the presence of clouds. Note that the cloud sensitivity could be biased by the retrieval approach and errors in the a priori CH₄ information. However, these biases can be corrected by the training of the RFC, and so do not necessarily hamper the accuracy of the trained model. Moreover, it is worth noting that from an implementation perspective, the TROPOMI CO product is well suited, as this retrieval precedes the TROPOMI processing of XCH₄.

In summary, we use the following key TROPOMI parameters: (1) the CH₄ column retrieved from a strong and weak CH₄ absorption feature in the SWIR spectrum (2) the CO averaging kernel at the ground level, (3) the latitude of the observed ground scene, (4) the viewing zenith angle of the observation, (5) the retrieved surface albedo at 2334 nm and (6) ECMWF surface pressure. All data are within the The TROPOMI CO product with version number 2.4.0 and later.

To demonstrate the validity of our RFC cloud clearing approach, we use the operational TROPOMI XCH₄ dataset derived from the NIR and SWIR measurements of the instrument. The operational algorithm is RemoTeC [12,24]. First, we consider the TROPOMI data for August 2022. During this period, the VIIRS instrument malfunctioned and, therefore, the data quality of the TROPOMI XCH₄ product was significantly compromised. To illustrate the potential of the RFC approach, we apply the new cloud clearing method to filter the data using the operational TROPOMI XCH₄ data from the latest data reprocessing and corresponding forward stream (version number 2.4.0).

Second, we validated the TROPOMI XCH₄ product against TCCON measurements for the full mission period by comparing cloud clearing based on the resampled VIIRS data and the new RFC machine learning approach. To this end, we reprocessed the TROPOMI XCH₄ data using the same configuration employed in the operational retrieval, with the exception of excluding any pre-filtering. This poses a significant processing challenge, but it is crucial to demonstrate the validity of our cloud clearing approach across all potential cloud contamination cases. Furthermore, this validation ensures the independence of our method from VIIRS data, which is also utilized in the pre-filtering stage. To streamline computational efforts, we limited the processing to a region of 300 km around the designated TCCON stations.

2.3. TCCON XCH₄

We used XCH₄ data from the Total Carbon Column Observing Network (TCCON) to show the impact of the RFC cloud clearing model on the validity of the TROPOMI XCH₄ dataset. TCCON comprises a globally distributed network of ground-based stations designed to provide accurate and precise retrievals of the vertically integrated total column of various trace gases. These measurements have gained recognition as a standard for satellite validation and have been utilized to validate satellite data products such as TROPOMI XCH₄ and CO (see, e.g., [11]). In this study, we used data from 12 TCCON stations, and their details are summarized in Table 1. Although a more recent TCCON dataset is available, namely GGG2020, we opt to use the version GGG2014 to maintain comparability with previous studies assessing the quality of the TROPOMI XCH₄ dataset [7,8,25]. Our validation approach is the same as that used in the previously mentioned publications, encompassing the coordination of TCCON XCH₄ measurements with TROPOMI XCH₄ measurements within a radius of 300 km around each station, while allowing a time discrepancy of $\pm 2$ hours. The collocation process yields daily means from both the TROPOMI and TCCON data, and from the corresponding time series, we compute the mean bias b, defined as the difference between the daily XCH₄ means (TROPOMI - TCCON), along with its corresponding standard deviation $\sigma$. Furthermore, we calculate the mean station-to-station bias $\overline{b}$, which is the average bias between all stations, and the standard deviation of the station-to-station bias $\overline{\sigma }$.

3. Methods

3.1. Random Forest Classifier

The concept of random forest classifiers (RFC) is well established in the field of machine learning. It comprises a set of decision trees n, where each individual tree is trained on a randomly selected subset of training data. This subset includes m input vectors $[{\mathbf{x}}_1,{\mathbf{x}}_2,\dots ,{\mathbf{x}}_m]$ paired with their corresponding correct classifications $[y_1,y_2,\dots ,y_m]$, where $y_i$ assumes values of 0 or 1. For this study, $\mathbf{x}$ consists of the parameters taken from the CO data product and y is the cloud-clearing deploying the SUOMI-NPP data as defined above. When classifying the input data, the Random Forest aggregates the predictions of all n trees and makes a decision based on the majority vote. RFC are known to be more robust against overfitting compared to other machine learning approaches, can handle outliers in the input data, and provide valuable insight into the feature importance. For this study we deployed the implementation of random forest classifiers provided by the sklearn python library version 1.2.2 [40].

We opted for a RandomForest configuration with specific parameters (n_estimators=150, max_depth=50, max_features=sqrt, min_samples_leaf=1, min_samples_split=2). In this configuration, the choice of 150 estimators determines the number of trees in the forest. A maximum depth of 50 is set for each tree, allowing for intricate pattern capturing. The use of max_features=sqrt introduces randomness by considering the square root of the total features for node splitting, fostering diversity among trees and potentially enhancing model generalization. A minimum leaf size of 1 ensures that each leaf in the tree contains at least one sample, and a minimum of 2 samples is required to split an internal node. These parameter choices were made based on the hyper parameter tuning that is implemented in the sklearn python library, aiming to achieve a well-balanced and effective RandomForest model.

For this study, we implemented an iterative training approach of the Random Forest Classifier (RFC) to optimize its performance. To ensure proper balance in the training set, we carefully matched the number of clear-sky and cloud-contaminated observations using the collocated VIIRS data. To this end, we first select all clear-sky retrievals and then randomly choose an equivalent number of cloudy data from the same orbit. Note that the large number of cloudy data allowed us to maintain this approach (e.g., typical orbit may yield approximately 20% clear-sky retrievals). Following the initial training, we evaluated the RFC’s performance on a test set comprising 1000 randomly selected orbits from a pool of approximately 20,000 available orbits. We identified the orbit with the highest number of false clear-sky predictions by the RFC, which we used subsequently to update the training set. Here, we ensure that the orbits in the training set are excluded in the test set. False clear-sky classifications are the most undesirable, as they diminish the accuracy of our data product. Here, accuracy is defined as the relative number of true clear-sky and false cloudy classification. On the contrary, false cloudy classifications primarily reduce data coverage, as true clear-sky measurements are omitted.

Our approach to selecting orbits with the most significant false clear-sky classifications aims to mitigate this specific error class. Figure 1 demonstrates the efficiency of this learning methodology. Initial training using only one orbit proved insufficient, yielding an accuracy below 60 percent. Moreover, the false cloudy classification rate exceeds 40 %, leading to a substantial reduction in data coverage compared to the original cloud filtering based on the resampled VIIRS data.

When the training set is expanded to include more orbits, there is a rapid increase in accuracy as the number of false cloudy classifications decreases. However, this improvement comes at the expense of a slight increase in false clear-sky measurements. The slow increase in false clear-sky classifications is primarily due to our selection criteria to update the learning set. A similar dependency is evident when examining the scatter of predictions in Figure 1. The scatter of accuracy is predominantly influenced by the high variability in false-cloudy classifications, a variability that diminishes with the inclusion of more orbits in the learning set.

Therefore, the final number of orbits for training is a crucial performance parameter. On the one hand, it should effectively mitigate the initially high number of false-cloudy classifications and their associated scatter. On the other hand, the objective is to minimize the occurrence of false clear-sky classifications. As illustrated in Figure 1, this optimal number of orbits tends to fall within the range of 100 to 150 orbits. Throughout this iterative process, we conducted a validation of the corresponding XCH₄ data product with TCCON observations for various number of orbits (see Figure A1 in the Appendix A). The RFC classifier for 110 orbits emerged as our baseline for this study (accuracy=82%, false clear-sky=11%, false cloudy=7%) with its ability to exhibit the lowest bias with TCCON within the specified orbit range between 100 and 150 of -7ppb.

3.2. Destriping Approach

The new cloud clearing approach suggested in this study is based on parameters retrieved from the SWIR spectral range by deploying the TROPOMI CO retrieval as discussed in Section 2.2. The data retrieved from the SWIR measurements of TROPOMI show a significant artificial stripping pattern in the flight direction of the satellite. This pattern is e.g noticeable in the TROPOMI CO, H₂O/HDO, and XCH₄ products [41]. Although the exact cause of this artifact has yet to be determined, it is likely associated with calibration of the TROPOMI level 0 data.

Hence, all the parameters sourced from the TROPOMI CO retrieval, essential inputs for our cloud-clearing approach, are susceptible to stripe artifacts. This susceptibility necessitates a preliminary destriping step to ensure that these artifacts do not inadvertently influence the cloud mask for the TROPOMI XCH₄ data. The destriping process is applied to several key parameters, including the XCH₄ column retrieved from both strong and weak CH₄ absorption features in the SWIR spectrum, the CO averaging kernel, and the retrieved surface albedo at 2334 nm. An illustrative example is depicted in Figure 2, showcasing stripes in the retrieved XCH₄, a side parameter derived from the TROPOMI CO retrieval and indicative of strong methane absorption in the shortwave infrared (SWIR).

We apply the following approach, which is robust against outliers and computationally fast. Assume $v(i,j)$ to be the parameter that is retrieved from the TROPOMI level 1 product. Figure 2 illustrates an example for the TROPOMI XCH₄ product. Here, the index i represents the number of swaths (in flight direction), and j corresponds to the number of a ground pixel within that swath (across the flight direction). In the initial step, we estimate a smoothed background $b(i,j)$ (see Figure 2B). This is achieved by applying a moving window median smoothing along the swatch direction. The width of the window is $w_1=7$ pixels. The residual component $r(i,j)=v(i,j)-b(i,j)$ contains valuable information on the presence of stripes in the data (shown in Figure 2C). We proceed with a second moving window median smoothing of the residual $r(i,j)$, this time in the flight direction. The width of this second smoothing operation is defined as $w_2=20$ pixels. This second smoothing step limits the stripe sensitivity of our method along the track direction and results in the stripe pattern $s(i,j)$ that we are going to remove. Finally, the destriped data product is obtained by calculating $d(i,j)=v(i,j)-s(i,j)$ as shown in Figure 2 C. The destriping approach is very efficient, as it operates on subsets of an orbit, which distinguishes it from earlier approaches using entire orbit data (see, e.g., [21]).

This novel destriping method holds significance not only for enhancing the destriping of input parameters crucial to our new cloud-clearing method but also for its potential application directly on the TROPOMI XCH₄ data in the future. To demonstrate this potential, we applied the destriping technique to a 5-year dataset of TROPOMI XCH₄ data and validated it against the 12 TCCON stations featured in this study. The results revealed that while the bias between TROPOMI and TCCON remained unchanged, there was a slight reduction in the standard deviation of the bias by 0.2 ppb. This outcome is particularly promising, as the removal of stripe noise in the data is expected to primarily reduce scatter rather than bias. Future plans include the implementation of destriping for ESA’s TROPOMI XCH₄ and CO products, as well as for our scientific water vapor isotope product [42].

4. Results

The RFC cloud-clearing for TROPOMI XCH₄ as described in Section 3.1 depends on the TROPOMI CO data and its availability. This dependency does not mean any restriction, as the TROPOMI CO data product has significantly larger coverage than TROPOMI XCH₄. CO retrieval is more resilient with respect to cloud contamination and level 1 data quality, so data that are rejected by the CO data processing are also not usable for XCH₄ processing. To evaluate the performance of our new machine learning-based cloud clearing approach, we reprocessed the TROPOMI XCH₄ data for August 2022 over North America [37.0°±17.0°N, 101.5°±34.5°W], Siberia [62.5°±6.5°N, 110.0°±14.0°E], and Australia [27.2°±16.5°S, 133.45°±20.12°E], a period when VIIRS data was unavailable. As a backup option when VIIRS data are not available, the operational TROPOMI XCH₄ retrieval applied a simple threshold filter to remove cloud-contaminated measurements. This filtering was used as a backup cloud clearing in case VIIRS data were not available [12]. However, as depicted in Figure 3A, the threshold filtering introduces large errors to the data set and allowed many cloud-contaminated measurements to remain in all regions considered. This discrepancy becomes evident when comparing the data distribution for August with that for July and September, when VIIRS data were available.

For August, the global XCH₄ distribution was skewed towards lower values, affecting both the mean value by 3.9% and the standard deviation by 303% for the three regions while introducing numerous outliers. This is mainly due to the shielding of air masses below the cloud. Thus, for this month, the data no longer align with the mission requirements and are flagged to be of low quality. Figure 3B shows the filtered data using our RFC learning approach. For July and September, this data set maintained the cloud screening based on the resampled VIIRS data. For August, the distribution of XCH₄ for the three regions is in good agreement with the other months. The mean and standard deviations are no longer skewed towards lower values, and even the number of outliers, indicated by the red data points, is reduced when compared to July and September.

The RFC cloud clearing approach appears to be more restrictive compared to the VIIRS-based method. This is illustrated in Figure 4 for TROPOMI overpasses over the US. Here, the RFC cloud clearing identifies similar clear-sky regions as the VIIRS approach but filters out more lower XCH₄ columns. As previously discussed, lower XCH₄ values are indicative of cloud contamination, suggesting that the RFC method may perform even better than VIIRS in this case. A explanation for this difference might be that the RFC approach is less sensitive to challenges in the VIIRS data because it has learned a general correlation between the VIIRS and the TROPOMI CO data. The persistence of cloud-contaminated scenes in the VIIRS-filtered data might be attributed to the 2-3 minute time difference between TROPOMI and VIIRS. Depending on the meteorological situation, the time delay can introduce a wrong clear-sky classification of the TROPOMI data using VIIRS.

The quality of the operational TROPOMI XCH₄ data is subjected to validation and continuous monitoring using XCH₄ reference data from TCCON network. To assess the effectiveness of the RFC cloud clearing approach, we processed five years of TROPOMI data around 12 TCCON sites (within a radius of 300 km), where we intentionally refrained from applying any cloud screening. This approach allows us to make a direct comparison between the different cloud clearing techniques using collocated VIIRS data and the RFC approach. For each station, we derived XCH₄ time series of daily means for both TCCON and TROPOMI. This is illustrated in Figure 5 showing data for Sodankyla, Finland, and Wollongong, Australia. The time series of all stations are shown in the Appendix A (Figure A2 and Figure A3). When comparing the TROPOMI XCH₄ data filtered with the RFC approach (A, C) to the current VIIRS filtering, we observe highly comparable results. Both approaches exhibit a similar data density, but when applying the new approach, we notice slightly fewer outliers with lower XCH₄ values. This agrees with the discussion above.

Figure 6 shows the validation statistics for all 12 stations. Overall, we see very good agreement in the bias (-7.4 ppb for RFC vs -5.6 ppb for VIIRS) and in the standard deviation of the bias (11.6 ppb for RFC vs 12 ppb for VIIRS). The amount of data is about 7 % higher for RFC compared to the VIIRS cloud clearing. In Figure 7, we present the correlation between the daily TCCON and TROPOMI XCH₄ means for all stations. The error bars in the figure represent the error of the daily means. The TROPOMI-TCCON correlation is highly comparable using the two cloud clearing approaches. Overall, from the results presented in Figure 6 and Figure 7, we conclude that the data quality around the TCCON sites is very comparable using both cloud clearing approaches.

5. Discussions

The quality of the TROPOMI XCH₄ data product depends on accurate cloud clearing of spectrometer data, a process facilitated by using the SUOMI-NPP VIIRS cloud product. During the last six years of mission operation, this approach proved to be very effective because both satellites operate in a formation, ensuring that they capture the same ground scene with a minimal delay of 2-3 minutes. However, the upcoming decommissioning of SUOMI-NPP necessitates the development of an alternative cloud-clearing method for the TROPOMI XCH₄ data product. This need arises due to the significant time delay of about 20 min between TROPOMI and the NOAA-20 satellite, which is the successor of SUOMI-NPP, and the malfunctioning backup cloud filter of the current TROPOMI XCH₄ processor. To maintain the accuracy and reliability of the TROPOMI XCH₄ data, a new cloud clearing approach had to be established in response to this new situation.

In this study, we introduced a new machine learning approach based on the Random Forest Classifier technique, which replicates VIIRS cloud clearing using only TROPOMI data. The classifier is trained on a subset of 5 years of collocated TROPOMI and SUOMI-NPP VIIRS data (about 20000 orbits). To this end, we used parameters derived from the TROPOMI CO retrieval, which is inherently sensitive to the presence of clouds and is processed prior to the retrieval of TROPOMI XCH₄ in the mission operational pipeline. This strategic choice simplifies the integration of our approach into the existing processing framework. In addition, we presented an efficient and robust method for despriping TROPOMI data, relying on median smoothing techniques. This approach was validated with ground-based TCCON measurements to show that it is not changing the bias, but slightly improves the standard deviation of the bias. This is a positive results since removing the stripe noise from the data should only improve the scatter and not introduce a bias change. The destriping will be suggested as an update for the operational TROPOMI XCH₄ and CO retrieval in future.

We demonstrated the performance of the new cloud-clearing approach by filtering three months of TROPOMI data in summer 2022 over North America, Siberia, and Australia. During one of the months, SUOMI-NPP experienced a temporary outage, which resulted in the lack of availability of VIIRS data. Remarkably, our machine learning approach demonstrated comparable performance during this data gap. Additionally, we conducted a TCCON validation covering the entire 5-year duration of the TROPOMI mission for 12 stations. The results highlighted the effectiveness of our RFC cloud clearing approach, showcasing its ability to match the performance of VIIRS filtering with a very similar bias and about 7 % more data.

Currently, our cloud-clearing approach is tailored for land-only scenes and our future plans include expanding its application to glint geometries observed over the oceans. Moreover, we will also apply RFC for improved quality filtering of TROPOMI XCH₄ data in future. When the dependence on SUOMI-NPP data is eliminated, faster processing becomes possible. This is in particular interesting for the scientific data application monitoring CH₄ point sources or assimilating TROPOMI data in near-real-time as done by CAMS-IFS for TROPOMI CO.