We use Level 2 total vertical column density (VCD) SO₂ data retrieved from TROPOMI, a high-resolution UV–Vis–NIR–SWIR spectrometer, onboard the Sentinel-5P satellite, January 2019–December 2024, inclusively. TROPOMI measures the solar radiation backscatter in the UV range (around 310–330 nm) where SO₂ has a distinct absorption feature. This analysis uses the TROPOMI Differential Optical Absorption Spectroscopy (DOAS) retrieval from version 3 of the offline analysis product. Sentinel-5P was launched in October 2017 into a sun-synchronous orbit with a local equatorial overpass time of 13:30. TROPOMI has a swath width of 2600 km divided into 450 across-track pixels, with a pixel resolution of 3.5 × 5.5 km (across × along track) at nadir for SO₂ (increased from a resolution of 3.5 × 7 km from 6 August 2019). This sampling strategy results in near-daily global coverage , subject to cloud-free scenes. In this study, we only use pixels with a quality flag > 0.5, as recommended by the TROPOMI Level 2 Product User Manual . No further filtering was applied to the data. Because this study uses the full VCD rather than the 1, 7 or 15 km layer products, we acknowledge that some potential SO₂ plumes located above cloud tops may be missed when applying a quality-flag threshold of 0.5. The broader implications of using the full VCD instead of layer-specific products are not assessed here, but this could be incorporated into future model developments and retrieval-selection strategies. A further development to explore in future work is the use of the Covariance-Based Retrieval Algorithm (CORBA) developed by . The COBRA retrieval method has been shown to reduce noise and biases in the SO₂ column when compared with the DOAS method and would therefore likely impact the plume detection results. The comparison between DOAS and COBRA is not presented here due to data availability at the time of analysis.

An existing SO₂ detection flag is provided in TROPOMI files as described in the Sentinel-5P/TROPOMI Algorithm Theoretical Basis Document built on a detection algorithm . This detection algorithm combines the SO₂ observation, solar zenith angle, VCD error and proximity to other detections to assign a flag to each pixel. This flag uses five categories: (0) no enhancement, (1) general SO₂ detection, (2) near a known volcano, (3) near a known anthropogenic source, and (4) a potential false positive due to a high solar zenith angle. This study does not include this flagging data in training the model as we want the model to learn SO₂ enhancement and plume shapes without relying on existing methods. The flags indicating if the detection are near a known source (flags 2 and 3) also rely on proximity to a known source, which could adversely affect the ability of model to detect plumes from new sources. Our plume detection model is designed to be distinct from the detection flag by not relying on proximity to known sources and looking at plumes as a whole structure, rather than a pixel-by-pixel basis. We compare the results from our plume detection model with SO₂ flag provided in Sect. . The method presented in this study does not assign a threshold to a potential plume and tests whether a model can learn to correctly identify plumes from the TROPOMI data without pre-defined limits.

To automatically detect plumes of SO₂ from TROPOMI data, we use a U-Net style fully convolutional network model to perform image segmentation . A U-Net model is designed to produce a pixel-by-pixel classification of an image and has been widely used in medical sciences (e.g. tumour detection) as well as in land cover classification in satellite imagery e.g.. It follows the basic principle of convolving an image over successive layers to reduce the spatial dimensions and extract feature information and then using transposed convolutions to rebuild the image to the original shape, the schematic of the model architecture creates a “U” shape, as shown in Fig. . This figure shows a schematic of the model architecture used for the plume detection algorithm. We add Gaussian noise to the training images to improve the robustness of the model before passing the image through four downsampling blocks. Each block consists of a double convolutional layer, a max-pooling layer and a dropout layer set at 20 %. These blocks then feed to one more double convolutional layer to extract patterns in the image before four upsampling blocks create a mask of the plume. Each upsampling block contains a double transpose convolutional layer, a concatenation layer, another dropout layer set at 20 % and a double convolutional layer.

This architecture was based upon the original U-Net designs described in then manually adapted through iterative tuning to fit this study. We found training the model for 20 epochs sufficient for our relatively simple image problem. We found no further model improvement when it was trained over more epochs. As our training dataset was relatively small (∼ 1000 images), we chose a batch size of 64 to balance over-fitting with training efficiency. The model uses a sigmoid activation as we wanted the model to make a binary pixel-wise classification of within a plume or not.

Image segmentation models offer significant advantages over traditional image classification e.g. because they parse essential information from the background rather than simply assigning a single label to an entire image. For detecting SO₂ plumes, this capability is crucial: it allows for more precise geolocating of the plume and the ability to handle multiple distinct plumes within a single satellite scene. Furthermore, segmentation excels over other feature-parsing methods, such as activation maps , because it is trained directly on ground-truth image masks. This direct comparison during training provides a clear, quantifiable performance metric, ensuring reliable results.

Our segmentation model was trained using a custom database of over 1000 TROPOMI images showing SO₂ plumes, each paired with a precise plume mask manually created by the lead author. These images were sampled randomly from all available TROPOMI files across the full study period and the global domain to minimize the risk of introducing geographical biases. While it would be possible to generate modelled plumes and corresponding masks for training, providing labels and complete background removal, this approach may introduce errors by misrepresenting observed plume characteristics. Consequently, we restrict training to SO₂ plume data derived from TROPOMI data.

The model is designed to detect single plumes and is not currently able to distinguish between multiple plumes that have merged (e.g. from nearby sources). Training the model to identify and separate merged plumes would require hundreds to thousands of additional examples, with merged plumes subdivided at the labelling stage, thereby introducing further potential for subjective error. More detail on the creation of the training dataset is provided in Appendix . To maximize the effectiveness of training, we augmented this dataset through rotations and flips, yielding a final training pool of over 4000 images and corresponding masks. This dataset size is sufficient to demonstrate the capabilities of this method. Unlike classification models (e.g., ), including images without plumes is not necessary for the training of segmentation models. As we are training the model to be interested in part of an image, the model learns the background of images (i.e. outside the area/object of interest) as the ‘no plume’ case.

We chose an image size of 32 × 32 pixels (roughly 112 × 176 km² at nadir) as it successfully captures most SO₂ plumes. The computational cost for increases slightly more than linearly with image size. Doubling both dimensions typically increases the cost by about four times. Images this size correspond to plumes approximately 90 km in length. We find this is more than adequate for encompassing the majority of plumes For validation, we trained the model on 80 % of the data and tested it on the remaining 20 %. The model's performance was measured using precision (correctness of positive predictions) and recall (completeness of positive detection), yielding scores of 65.7 % and 74 %, respectively, giving an F1 score of 0.69. Crucially, the small 32 × 32 image size disproportionately penalizes minor errors, meaning an offset of just a few pixels drastically lowers the score. Although these scores indicate the performance of the model, the precision and recall percentages are not directly interpretable as plumes missed or not, as segmentation models work on a pixel per pixel basis. For example, the model may correctly detect a plume in the test dataset but draw a smaller or larger mask than the test data and would therefore be penalised. It is for this reason that we cannot directly compare these performance metrics to plume classification models e.g.. It is also important to note that because the training dataset was created manually and relied on the authors’ judgement of what constitutes a plume, the model cannot outperform the quality of this dataset. Any biases or recurring misclassifications present in the training examples may be learned by the model and subsequently propagated into the final results. An alternative approach would be to augment the training dataset with model-simulated SO₂ plumes, but for this study we chose to rely exclusively on real data to maximise the diversity of scenes and plume morphologies that may not be captured in model simulations.

To ensure we capture SO₂ plumes that may be straddling multiple image boundaries, we employ a 32 × 32 pixel rolling window that moves four pixels at a time across and along the satellite swath. This systematic sampling generates approximately 100 000 images per swath for input into the segmentation model. The model returns each image with a pixel-by-pixel probability of plume presence. We then reconstruct the original swath into an amalgamated mask by taking the median probability of all overlapping pixels. Using the median not only combines the individual predictions but also boosts detection confidence, as an accurately identified plume will appear consistently across multiple overlapping images. We found that a four-pixel step provides adequate coverage to resolve straddling issues while maintaining reasonable computing speed and costs. Crucially, this overlaying method means the final predicted plume shape is not limited by the 32 × 32 pixel input size, allowing plumes and their corresponding masks to be accurately mapped across the full scale of the swath (which is typically 4172 × 450 pixels, spanning 2600 km from pole to pole).

To extract the details of individual plumes from the reconstructed satellite swath, we apply connected component analysis (using Open-CV; ) to the pixel probability array. This analysis effectively identifies the unique plume masks within the swath and provides the bounding boxes for each one, allowing us to precisely delineate the plume outline. Figure  illustrates this capability, showing twenty randomly selected TROPOMI-observed plumes alongside their corresponding predicted plume outlines. The images in this figure show the varying sizes of the plume images, and that they are not limited to the 32 × 32 pixel training image size.

Figure  demonstrates that the model generally performs well, although some inaccuracies are present. Detecting atmospheric features like SO₂ plumes inherently involves subjectivity, as there is no clear, objective physical boundary for the feature of interest. This human subjectivity is inevitably encoded in the model's training dataset and subsequently reflected in the trained model itself. Continuously refining the model or updating the training dataset is an endless task, so for the practical scope of this paper we have chosen to present results based on a rigorous process involving three training iterations (i.e., checking model output, expanding the dataset with new examples, and retraining).

We have created a comprehensive database documenting each detected plume, which includes its location, date of detection and plume outline. We also estimate the length of the plume, based on taking the length of the primary axis of an ellipse fitted to the plume outline. We find the detected plumes have a median plume length of 37.5 km, much shorter than the length of the model input image. A bounding box (with a three-pixel buffer) is also recorded, with its limits specifically used to determine the background SO₂ concentration outside the plume. Computationally, processing a single swath is highly efficient, taking only about 15 s using a GPU or 60 s using CPUs.

We processed approximately 31 000 Level 2 files from TROPOMI observations spanning January 2019 to December 2024. To ensure the reliability of our output, we implemented a filter requiring each detected plume to contain a minimum of six pixels, as the accuracy of smaller predicted plumes is difficult to validate.

In total, 53 993 plumes were identified over the study period. The year 2019 saw the highest number of detections, exceeding 16 000, while the period from 2020 through 2024 showed more consistent annual counts, ranging from ∼ 6800 to ∼ 8000 plumes. To locate the emission source, we employ the coordinates of the maximum SO₂ concentration within the plume boundary as a practical proxy for the origin, noting that this might not perfectly match the true source location. Figure  shows the annual global distribution of these estimated plume origins, 2019–2024.

Initial inspection confirms plumes cluster predictably around known volcanoes and industrial hotspots. However, all years also show regions with a wide, “noisy” spread of detections (e.g., Alaska, Canada, and Eastern Russia annually, plus specific areas like the mid-Atlantic in 2021 or Central Africa in 2024). Figure  illustrates the daily plume counts, where each spike corresponds to these noisy geographical spreads (Fig. ). Critically, all but one spike (Peak I in 2023) align with major volcanic eruptions in the region. Table  details the specific volcanic emissions believed responsible for this widespread geographical dispersal. Plume transport away from the source is evident, as detections further from the origin often occur in the days immediately following an eruption event (see Appendix , Fig. ).

We attribute the major spike in 2019 plume detections mostly to the coinciding eruptions of Raikoke, Russia, and Ulawun, Papua New Guinea. The Raikoke eruption on 21 June is particularly notable, injecting one of the largest amounts of sulfur dioxide (SO₂) into the stratosphere since the 1991 Mount Pinatubo eruption and references therein.

While Peak I in Fig.  corresponds to the Shiveluch eruption in Russia on 11 April 2023, it also coincides with a surge in detections concentrated over Norilsk in Northern Russia (88.1° W, 69.3° N), as shown in Fig. . This region is a known major source of SO₂ emissions due to its large metal smelting operations. The precise cause of this specific increase in plume detections, however, remains unknown.

Using the DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm , we grouped the detected plumes based on the coordinates of their maximum SO₂ observation (our proxy for origin). We set the clustering parameters to a maximum distance of 50 km between points and a minimum of 20 samples per cluster. This technique effectively identifies global areas with a high concentration of plumes. Figure  shows the centers of the 90 detected clusters, coloured by the number of plumes they contain, which further validates that our model successfully identifies the major SO₂ sources worldwide.

We observed regions on the global maps, most notably China (a location with numerous known SO₂ sources), where the model failed to detect expected plumes. Assuming the quality and volume of TROPOMI data are consistent globally, this deficiency likely stems from the plume detection model itself. We hypothesize that high background SO₂ concentrations in such regions may “hide” point sources, making plumes difficult to isolate. Although emissions of SO₂ have declined in China over the past couple of decades, ambient SO₂ levels are still relatively high compared to the rest of the globe , potentially making this effect more prominent in this region. Since the model was trained on global data, it may struggle with these outlier scenarios. Future iterations should address this by specifically training on data that includes plumes from high-background regions like China.

Table  details the ten largest clusters of SO₂ plumes detected over the 6-year study, alongside their probable sources. Volcanic activity dominates this list, accounting for seven out of the ten largest clusters. The remaining three clusters are associated with industrial activities: metal smelting, coal mining, and oil and gas operations. It is crucial to note that these clusters are often complex; some may contain multiple source types. For example, the Iztaccíhuatl volcano cluster is located just south of Mexico City, meaning the cluster likely includes industrial SO₂ sources alongside the volcanic emissions.

There is no notable impact on the plume detection results due to the increase in the across-track spatial resolution of the TROPOMI SO₂ product on 6 August 2019 from 7.0–5.5 km. There is larger variation in pixel size between the nadir and edge of swath observations than the increase in resolution and the model copes well with all these variations.

Using the Smithsonian Institution's Global Volcanism Program (GVP) database of known eruptions since 1960 , we identified 227 active global eruptions during our study period. Comparing this list with our plume database revealed that 7943 plumes were detected within a 50 km radius of an eruption during its active date. This confirms plume detection for 110 (49 %) of these eruptions. The 117 missing detections are likely due to poor TROPOMI retrievals caused by high cloud cover or heavy aerosol loading. Figure  maps these GVP-reported eruptions, coloured by whether plumes were detected nearby (red) or not (green).

The plume detection algorithm struggles with very large plumes (> 1000 km) that often occur far downwind of major eruptions. Figure  illustrates this, showing a large SO₂ plume over the United Kingdom on 24 August 2024, originating from the Svartsengi volcanic system in Iceland (63.8° N, 22.38° W) six days earlier. The red outlines show the model splits this large plume into multiple smaller segments. This error likely results from the method used to break the TROPOMI swath into smaller images, which prevents the model from seeing the larger context of a plume that can span the entire swath.

The TROPOMI Level 2 data assigns a flag to SO₂ pixels based on five categories: (0) no enhancement, (1) general SO₂ detection, (2) near a known volcano, (3) near a known anthropogenic source, and (4) a potential false positive due to a high solar zenith angle. Since a single plume often covers many pixels, we assign the final plume label only if over 80 % of its pixels share the same TROPOMI flag (likely determined by spatial proximity to known sources); otherwise, the plume is labeled as a combined source. We found that the dominant category is a general SO₂ detection with no source attribution (43.5 % of plumes), followed by volcanic plumes (30.8 %) and known anthropogenic sources (14.7 %). Figure  illustrates the geographical distribution of these categorized plumes.

The high number of unlabelled SO₂ detections is likely caused by plumes that have traveled downwind from their origin and are thus no longer close to the known volcanic or anthropogenic sources used for TROPOMI flagging. This explanation is strongly supported by the evidence of extensive volcanic outflow visible in both the geographical distribution of plumes (Fig. ) and the daily detection spikes (Fig. ).

The TROPOMI “No enhanced SO₂ detected” category accounts for 10.0 % of all detections, but we believe many of these should be attributed to an actual source rather than being false positives. We demonstrate this by applying the clustering algorithm (as described in Sect. ) to these “no enhanced detection” plumes. This allows us to extract meaningful groupings and calculate their proximity to clusters with known sources. Figure  displays all plumes in this category, highlighting 34 dense clusters (minimum 20 plumes within 150 km). Crucially, 27 of these 34 clusters overlap with clusters that have a known emission source, and the remaining seven are within 300 km. While these clusters account for 24 % of the “no enhanced detection” plumes, many of the remaining plumes form a large, dispersed swath extending from 100° E to 180 ° W around 10° S. This pattern aligns strongly with the downwind transport of SO₂ from the Taal Volcano eruption in the Philippines (120.99° E, 14.01° N) in April 2024 (detailed in Sect. ).

This analysis demonstrates that while the source labelling provided in the TROPOMI files is informative, there is potential for our plume detection algorithm to complement the current flagging through masking and grouping plumes, facilitating analysis into repeated new detections. Although our plume detection algorithm does not attribute sources itself, it successfully highlights geographical and temporal patterns that are crucial for identifying important, and potentially unlabelled, SO₂ emission locations.