As a case study, the Etna eruption activity between 13 December 2020 and 21 February 2022 was selected. During this period, Etna volcano produced 66 lava fountain episodes from the new South-East crater. Despite their short duration (an average time of about two hours; Calvari and Nunnari, 2022), they produced a strong impact on human life, environment, and air traffic.

These paroxysmal events produced large Volcanic Clouds Top Heights (VCTH) ranging from 4 to 13 kma.s.l. Based on VCTH, the entire period can be approximately divided into three main time ranges characterized by average VCTH values of 9 km (from 13 December 2020 to 19 March 2021), 6 km (from 24 March 2021 to 17 June 2021) and 10 km (from 19 June 2021 to 21 February 2022). Most of these volcanic clouds were composed of ash, ice, and SO2, with the ice content being greater than the ash. The ice generation was related not only to the volcanic cloud height but also to the season: during the summer, with almost the same plume height, the amount of ice was lower than that observed in winter (Guerrieri et al., 2023). The presence of ice in most volcanic clouds during this period makes it an interesting case study for exploring the challenges associated with volcanic cloud detection.

Data from the Spinning Enhanced Visible and InfraRed Imager (SEVIRI) instrument aboard the Meteosat Second Generation (MSG) geostationary satellite, specifically from the Meteosat-10 series orbiting at a 0° longitude, was considered in this work. The SEVIRI instrument can produce an image of the Earth's full disk every 15 min in 12 different spectral channels ranging from visible to infrared, with a nominal spatial resolution of 3 km × 3 km (1 km2 for High-Resolution Visible channel) at the sub-satellite point (EUMETSAT, 2017). The SEVIRI's channel description and their nominal centre wavelengths are shown in Table 1. All the SEVIRI data used in this work were acquired in near real-time using the Multimission Acquisition SysTem (MAST), which was developed at Istituto Nazionale di Geofisica e Vulcanologia - Osservatorio Nazionale Terremoti (INGV-ONT). The MAST system uses EUMETCast dissemination service for the near real-time delivery of satellite data and products (Stelitano et al., 2023).

An example of a volcanic cloud captured by SEVIRI from 19 February 2021 at 10:15 UTC is displayed in Fig. 1. This volcanic cloud is composed of a mixture of ash, ice, and SO2. The natural color composite, created by assigning the IR1.6, VIS0.8, and VIS0.6 channels to the red, green, and blue color beams, respectively, is shown in Fig. 1a, where the volcanic cloud appears as a bright white cloud, similar to the surrounding meteorological clouds. By contrast, the volcanic cloud shown in Fig. 1b is well distinguished in the Ash RGB composite (Ash RGB Quick Guide | EUMETSAT – User Portal, 2025), which uses the thermal infrared channel. In this latter figure, generated by visualizing the channel differences IR12.0 - IR10.8 and IR10.8 - IR8.7 in the red and green color beams, respectively, and the IR10.8 channel in the blue beam, the constituents of the volcanic cloud are clearly identifiable: ash appears in shades of brown, ice in dark blue, and SO2 in green color.

The dataset was generated from 1259 SEVIRI images covering 49 eruptive Etna events between December 2020 and February 2022 (see Fig. 2a). Overall, the dataset represents a comprehensive compilation of eruptive events, covering a wide range of volcanic cloud conditions. These events include SO2 plumes, SO2 dominated plumes containing minor ash components, and volcanic clouds composed of ash and SO2. The dataset also includes highly challenging cases in which the volcanic clouds were composed of ash, SO2, and ice, sometimes occurring under clear-sky conditions surrounding the volcanic cloud, and in other cases in the presence of meteorological clouds surrounding or mixing with the volcanic cloud. Table A1 provides a detailed description of the SEVIRI images employed to generate the dataset, including the number of images, the time range for each event, and a brief comment describing the composition of the volcanic cloud during each eruptive event.

All images were manually analysed in order to create a plume mask for each one; this step, generally known as the labelling process (see Fig. 2b), was also used in Guerrieri et al. (2023) to estimate the total masses of ash, ice, and SO2. Figure 1 also presents an example of a plume mask. The manual plume mask generation generally provides the most accurate volcanic cloud detection, but it is a time-consuming process that requires expert interpretation. Despite the expertise of a human operator, some challenging events remain difficult to discriminate. This is particularly true when volcanic clouds are surrounded by meteorological clouds, making the identification of plume mask boundaries more uncertain. In such cases, the human operator tends to adopt a conservative approach, generating a smaller plume mask that includes only pixels with strong evidence of volcanic cloud presence. Consequently, the most ambiguous pixels were intentionally excluded from the manual plume mask, with the aim of allowing the NN model to learn how to discriminate these challenging events during the prediction phase.

In this work, the detection of volcanic cloud using SEVIRI data is analysed as a pixel-scale classification (see Fig. 2b(iii)). Thus, the dataset includes information from more than 2 200 000 pixels, organized in two balanced classes: 50 % belonging to the Volcanic Cloud (VC) class and the other 50 % to the Non-Volcanic Cloud (NVC) class, as indicated in Fig. 2c. The original dataset was highly imbalanced, containing 0.3 % VC pixels and 99.7 % NVC pixels. All VC pixels were retained, while NVC pixels were randomly sampled to obtain the final balanced dataset. The VC class refers to volcanic cloud pixels containing ash, SO2, ice, or any combination of these components, whereas the NVC class refers to pixels associated with meteorological clouds, land surfaces, sea surfaces, and other non-volcanic features.

Among the pixel features that comprise the dataset are the thermal radiances from seven spectral channels, corresponding to the thermal-infrared region (6–14 µm) as detailed in Table 1. Selecting these spectral channels enhances model flexibility, allowing application both during the day and at night with a unique NN model. In the dataset, the thermal radiance from each channel is expressed as Brightness Temperature (BT) in Kelvin (K). In addition, several channel combinations were applied to derive six new features, known in the literature as Brightness Temperature Difference (BTD) including the traditional BTD between the channels centered around 11 and 12 µm. These new features are BTD[10.8–12.0], BTD[10.8–8.7], BTD[10.8–13.4], BTD[10.8–6.2], BTD[10.8–7.3] and BTD[10.8–9.7]. Thus, a total of 13 features are included in the dataset (as indicated in Fig. 2a), all serving as input variables for the neural network model. The balanced dataset is accessible in (Naranjo et al., 2026), where additional details regarding its spatial and temporal coverage are provided.

To determine the most suitable configuration for the NN model, a hyperparameter tuning phase was carried out. During this step, the hyperparameters listed in Table 2 were optimised using the search space specified in the table's right column. This process combined exhaustive search with a 5-fold cross-validation strategy (k=5) (Ojala and Garriga, 2010), evaluating all possible combinations within the defined optimisation space. Moreover, the balanced dataset (50 % VC and 50 % NVC, totalling 2.2 million pixels) was split into 80 % for the training set and 20 % for the testing set, in accordance with the proportions commonly employed in the literature (Sun et al., 2022). The split was performed using random sampling with a stratified strategy. This stratified approach ensures that the training and testing sets contain approximately the same proportion of samples from each class (VC and NVC) as the balanced dataset.

During the hyperparameter tuning phase, both the classification accuracy and the average time required for the NN model to perform a single classification were considered. A trade-off was therefore established between achieving a training accuracy of 96.0 % and maintaining an average inference time of approximately 1.5 s per classification. The complete hyperparameter tuning results are provided in the Supplement, while the final selected hyperparameter combination is reported in Table 3 (Sect. 3.2, Training phase). Using these hyperparameters, the evaluation metrics obtained for the testing set were 95.0 % accuracy, 98.6 % precision, and 91.3 % recall (see Sect. 3.5 for the definition of these metrics).

During the training phase, the dataset was also split into 80 % for the training set and 20 % for the testing set, as in the hyperparameter tuning phase. Additionally, the data for each feature was independently standardized by removing the mean and scaling to unit variance, the process is also known as Z-score normalization. In this standardization step, data from both classes (VC and NVC) were included to calculate the mean and standard deviation for each feature. Standardization is an important requirement for NN models, as it generally improves their performance (Pinheiro et al., 2025).

The NN model was trained using three hidden layers, each containing 60 neurons, as presented in Table 3. The learning rate and L2 regularization hyperparameters were both set to 1 × 10⁻³, based on the results obtained during the hyperparameter tuning phase. The activation function used was the Rectified Linear Unit (ReLU), which is the standard activation function for NN models based on MLPs (Braga-Neto, 2020; Yang, 2019). For the optimisation algorithm, the Adaptive Moment Estimation (Adam) method was applied, due to its computational efficiency and robustness in the training of models using large datasets (Kingma and Ba, 2014). Finally, the maximum number of training iterations was set to 300 and an early stopping strategy was implemented, whereby the training process was terminated when the model performance no longer improved between successive iterations.

In addition, the NN model was calibrated. When the model performs a classification, the output corresponds to the probability associated with the respective class, with values close to zero indicating the NVC class and values close to one indicating the VC class. However, the raw output probabilities are not inherently calibrated and may provide unreliable estimates of the true class probabilities. For this reason, a calibration procedure was incorporated into the training pipeline, enabling the predicted probabilities to be interpreted as the model's confidence that a given pixel belongs to a specific class (Niculescu-Mizil and Caruana, 2005).

The model's confidence specifically depends on the dataset on which it was trained, which in our case corresponds to the balanced dataset. For the calibration, the testing set (20 % of the balanced dataset) was used to train an additional calibrator model. This calibrator learned to map the raw output probabilities (praw) of the NN model to a calibrated probability (Pcal). In practice, the calibrator predicts the conditional event probability P(prediction=VC|praw). Therefore, the calibrated probabilities can be interpreted as confidence level.

The Pcal (depicted in Fig. 3a) is estimated by fitting an isotonic regression model using the NN raw output probabilities and the corresponding true labels from a testing set. The model learns the relationship between the raw probabilities and the observed probabilities (Guo et al., 2017; Niculescu-Mizil and Caruana, 2005). The calibrated probability for a new prediction is then obtained by applying the learned isotonic mapping to the NN model's raw output probabilities.

The calibration curve for the trained NN model is shown in Fig. 3a, where the y axis represents the observed proportion of VC class pixels in the testing set, whereas the x axis represents the predicted probabilities for these pixels. The calibration curve was generated by binning predicted probabilities and then plotting the mean predicted probability in each bin against the observed proportion of VC class pixels (Bröcker and Smith, 2007).

As shown in Fig. 3a, the NN model exhibits a slight tendency to overestimate low probabilities and underestimate high probabilities. For example, according to the Fig. 3a, among all pixels in the testing set predicted by the NN model with probability 0.80, more than 80 % actually belong to the VC class. However, the overall calibration performance is good, as can be seen in Fig. 3b. This figure presents the distribution of predicted probabilities, showing that for the VC class, most probabilities are concentrated at low (less than 0.2) and high (greater than 0.8) values. This symmetric and bimodal distribution indicates that only a small number of ambiguous probability values are present, which is a desirable characteristic of a well-calibrated classifier. Based on these results, a pixel will be classified as VC when the calibrated probability is higher than 0.8.

To validate the performance of the NN model, three sequences of SEVIRI images were selected, each corresponding to an eruption the model had not previously encountered during the training phase. The events are listed in Table 4, comprising two from 2021 and one from 2024. The SEVIRI images for the 2024 event were captured by the Meteosat-11 satellite. In contrast, the images for the 2021 events were captured by the Meteosat-10 satellite, the same satellite used to acquire the data for the training phase. All three events represent challenging scenarios in which the volcanic clouds are composed of ash, ice, and SO2, with the last event additionally characterised by the mixing of the volcanic cloud with meteorological clouds.

Each SEVIRI image sequence was also manually analysed to create a plume mask for each image, following the same labelling process carried out for the training dataset (as illustrated in Fig. 2a). The plume masks are used as the observed data to compare the predictions of the NN model and assess its performance through conventional evaluation metrics such as accuracy, precision, recall, F1-score and confusion matrix (Fawcett, 2006; Rainio et al., 2024).

The raw output of the NN model is subsequently post-processed using a plume tracking method and a non-local (NL) mean filter. This two-step post-processing procedure is designed to reduce false positive detections (see Fig. 2f). A similar methodology was previously presented by Pavolonis et al. (2018), who introduced the Cloud Growth Anomaly (CGA) technique.

The tracking algorithm developed in this work is initialized (t0) within an area corresponding to a 25 km radius around the Etna volcano, as depicted in Fig. 4a. Pixels falling outside this region are not considered in the analysis. It is important to note that the processing is performed in image coordinates (x,y pixel coordinates). Therefore, this area corresponds to selecting the pixels located within a circular region of approximately 8 pixels in radius, given that SEVIRI data has a spatial resolution of 3 km.

At first, the algorithm evaluates every new prediction from the NN model, searching for pixels classified as volcanic cloud (VC), which serve as the triggering event. The raw NN plume mask is treated as a potential VC object. When such a VC object appears within the area of interest, the tracking algorithm initiates tracking at time ti=1 and proceeds to analyse the subsequent images. Thus, in the subsequent images, two processes occur:

The radius of the circular region increases by approximately 9–15 km (3–5 pixels) for each new image (ti+1), depending on the wind speed for that day. These values are based on the SEVIRI spatial resolution and on typical wind speeds of 10–16 ms-1 for volcanic clouds at altitudes of 10–12 km, corresponding to the range of values observed for the validation events considered in this study. This parameter can be adjusted if required.

This operation continues until no pixels from the raw NN plume mask remain within the tracking area. Figure 4 shows an example of the plume tracking algorithm in action.

At each iteration (i) of the tracking algorithm, scattered “noisy” pixels are removed using a non-local (NL) mean filter (Buades et al., 2005; Pavolonis et al., 2006), followed by the application of a binary dilation filter. The dilation filter is employed to restore and expand the edges of the plume mask, which may become eroded during the application of the NL means filter.

Unfortunately, in the presented version of the post-processing algorithm, not all “noisy” pixels can be effectively removed through the application of the NL means filter. In addition, the tracking algorithm does not include a control to prevent excessive growth of the circular region, which may become problematic during near-real-time operations when tracking volcanic clouds over long time periods. Under such circumstances, a more sophisticated post-processing algorithm could be developed. An example of a real image before and after the application of this processing step is presented in Fig. 2f. The algorithm is also robust in tracking fragmented volcanic clouds, as shown in Fig. 5e in the Results section.

Finally, the results and evaluation metrics presented in the Results section correspond to the filtered plume masks obtained after the application of this post-processing step.

The performance of the binary classification between the VC and NVC classes was evaluated using a set of standard metrics, which are briefly defined in this section (Fawcett, 2006; Rainio et al., 2024). In this study, the convention adopted is as follows: the outcome of the manual labelling process is defined as the observed data, whereas the outcome of the classification process produced by the NN model is defined as the predicted data. Both observed and predicted data are considered positive when they belong to the VC class, and negative when they belong to the NVC class. Thus, considering this convention, each pixel prediction can be classified into one of four possible categories:

True Positive (TP): The observed class is VC, and the predicted class is also VC. Therefore, the VC class is correctly predicted.

The total number of predictions in each category can be represented in a 2 × 2 matrix, known as the confusion matrix (C), which provides a visual representation of the classification results. In this matrix, TP is located at C0,0, FP at C0,1, FN at C1,0, and TN at C1,1.

Based on the confusion matrix, several standard metrics for binary classification can be calculated. These metrics include accuracy, balanced accuracy, precision, recall, and F1-score, all of which range from 0 (poor performance) to 1 (optimal performance).

Accuracy: The proportion of correctly predicted pixels divided by the total number of pixels.1Accuracy=TP+TNTP+FP+FN+TN

It is worth noting that the accuracy metric was used for the training and testing datasets, whereas balanced accuracy was used for the validation dataset. This choice was made because the validation dataset is highly imbalanced, making balanced accuracy metric more appropriate for this type of dataset.

An additional analysis was conducted to estimate the volcanic cloud mass loading from both the manually observed data and the NN predicted data. The mass loading was retrieved using the Volcanic Plume Removal (VPR) algorithm (Pugnaghi et al., 2013, 2016), where the pixels classified by the NN model as VC class were used as input to the retrieval procedure.

Since the NN model output generates a general plume mask containing ash, ice, and SO2 without discriminating among these components, the procedure adopted to apply the VPR algorithm separately to each component was the same as that presented in Guerrieri et al. (2023). In this procedure, Radiative Transfer Model (RTM) computations were performed to derive a BTD[10.8–12.0] threshold. Accordingly, a positive BTD[10.8–12.0] threshold was used to discriminate ash from ice within the NN predicted plume mask, whereas the SO2 retrieval was applied to the entire plume mask. The thresholds used for each event were 1.15 for Event 1, 1.12 for Event 2, and 1.25 for Event 3. For further details regarding the calculation of the BTD[10.8–12.0] thresholds, see Sects. 2.2 and 2.3 in Guerrieri et al. (2023).

The retrievals derived from the NN predicted plume masks were subsequently compared with those obtained from the manually generated plume masks (using the same BTD[10.8–12.0] thresholds).

As outlined in the Sect. 3.2, the NN model consists of three hidden layers with 60 neurons each one and uses 13 input features, resulting in a complex model. This level of complexity limits the understandability of the model's classification process (Flora et al., 2024). In order to better understand the classification process, an explainability method called Shapley Additive Explanations (SHAP) was used. SHAP values is a game theoretic framework that quantifies the contribution of each feature to individual predictions, thereby providing insight into the model's overall internal mechanisms (Lundberg et al., 2017).

The discussion section presents a beeswarm plot illustrating feature relevance for the NN model based on the testing set. The beeswarm plot displays the features along the y axis, ordered according to their influence on the model prediction, from the highest to the lowest. The impact of each feature on the prediction is represented along the x axis through the SHAP values, while each dot corresponds to a pixel from the testing dataset. Positive SHAP values indicate a positive contribution to the prediction of the VC class, whereas negative SHAP values indicate a contribution toward the prediction of the NVC class. The color of the dots indicates whether the corresponding pixel exhibits a high or low value for the respective feature.

This event was a paroxysmal episode at the Southeast Crater, which lasted 10 h. It was primarily characterized by a lava fountain with a duration of 50 min and a volcanic cloud that reached more than 11 km in height (Guerrieri et al., 2023; INGV, 2021b).

At the onset of this event, the volcanic cloud was dispersed toward the northwest, and it is observed with a thick core displaying brown shades, indicating a high concentration of particles, presumably ash. In contrast, the cloud's edges appear in dark blue tones, suggesting thinner regions primarily composed of ice (see Fig. 5b). As the eruption progresses, the denser portion of the cloud disperses and eventually appears entirely dark blue, highlighting a strong presence of ice. Additionally, a sulphur dioxide (SO2) signal is visible in the northern part of the cloud, represented in green tones (see Fig. 5d). By the end of the sequence, the primary volcanic cloud has dissipated, and a new eruptive pulse produces a low-level volcanic cloud, visible in red shades in Fig. 5f.

Notably, in Fig. 5a, c, and e, the NN model effectively detects most components and the overall structure of the volcanic cloud throughout its evolution, even when the cloud becomes fragmented as is revealed in Fig. 5e. The balanced accuracy for the complete image sequence of this event was 82.9 %, while the precision, recall, and F1-score were 92.3 %, 66.0 %, and 77.0 %, respectively, as reported in Table 5.

The time series of the precision and recall metrics for each image in the sequence using the filtered NN plume mask are presented in Fig. 8a. It can be observed that both precision and recall are high at the beginning of the eruption, with values between 80 % and 90 %, but decrease to below 80 % during the middle stage of the eruption.

This second event was produced by a strombolian activity at the Southeast Crater, which started at 07:50 UTC. The activity then evolved into a lava fountain that generated a volcanic cloud rising to more than 11 km (Guerrieri et al., 2023; INGV, 2021a).

For this event, the composition of the volcanic cloud is quite similar to the previous case. In this instance, the cloud moves toward the northeast. At the onset, ash particles are observed in the inner part of the cloud, shown in brown shades (see Fig. 6b), which are later masked by ice (see Fig. 6d). Additionally, the SO2 signal, visible in green, becomes more prominent toward the end of the sequence.

Unlike the previous case, this volcanic cloud remains compact, and the NN model successfully detects most parts of the cloud throughout the sequence (see Fig. 6a, c, and e). The cloud was continuously tracked for 12 h and over a distance exceeding 550 km from the volcano. Toward the end of the sequence, some portions of the volcanic cloud containing SO2 are no longer detected, as illustrated in Fig. 6f. The reduced detection performance is likely attributable to cloud dilution in those areas, resulting in an overall balanced accuracy for the complete image sequence of 74.2 %, a precision of 96.0 %, but lower recall and F1-score values of 48.5 % and 64.4 %, respectively, as reported in Table 5.

Figure 8b shows the evolution of the precision and recall metrics for each image in the sequence. It is worth noting that, at the beginning of the eruption, this event presented the highest recall values among the three validation events, with values close to 100 %. In other words, during the initial stage of the eruption, the filtered NN plume mask detected nearly all pixels observed as VC. However, the recall metric decreased around 11:00 UTC, corresponding to the moment when the volcanic cloud passed over the continental surface, before increasing again and subsequently decreasing to below 80 % during the middle stage of the sequence.

This last validation event was also produced by a lava fountain, this time from the Voragine Crater, starting at 03:20 UTC. The fountain generated a volcanic cloud that rose to over 10 km (INGV, 2024). The volcanic cloud exhibited wind-driven behaviour, dispersing toward the east and southeast. This case is particularly challenging, as the volcanic cloud is composed of ash, ice, and SO2, and is both surrounded and mixed with meteorological clouds.

At the onset, the volcanic cloud appears in red shades, with some SO2 signal in green (see Fig. 7b), before encountering the mid-level meteorological clouds. Note how the volcanic and meteorological clouds blend in Fig. 7d and f. Detecting volcanic clouds under these conditions is generally very difficult as reported in previous studies (Prata et al., 2022; Taylor et al., 2023). Nevertheless, the filtered NN plume mask successfully identified the volcanic cloud, and the tracking algorithm subsequently followed its evolution even when it was blended with meteorological clouds (see Fig. 7a, c, and e). For this event, the overall balanced accuracy obtained was 92.0 %, while the other performance metrics were 78.1 % precision, 84.2 % recall, and 81.0 % F1-score, as reported in Table 5. The evolution of the precision and recall metrics across the image sequence is shown in Fig. 8c. It can be observed that these metrics remain above 80 % for recall and above 60 % for precision during most of the sequence.

As can be seen in Table 5 and the Fig. 8a–c, the metrics indicate strong performance across all validation events using the filtered NN plume mask. Looking at the precision and recall metrics, we can gain deeper insight into the performance of the filtered NN plume mask. For instance, the precision values for Event 1 and Event 2 are 92.3 % and 96.0 %, respectively, indicating that 92.3 % and 96.0 % of the pixels detected as VC in each event were correctly classified. These high precision values indicate that the detection was highly reliable, with only 7.7 % and 4.0 % of the detected pixels corresponding to false alarms, respectively.

Whereas the recall values for Event 1 and Event 2 are 66.0 % and 48.5 %, respectively, indicating that the model was able to detect 66.0 % and 48.5 % of the VC pixels observed in the reference plume mask. In other words, the filtered NN plume mask failed to detect 34.0 % and 51.5 % of the VC pixels observed in the reference plume mask for Event 1 and Event 2, respectively.

These relatively low recall values are likely due to dilution in the outer portions of the volcanic cloud, which fall below the detection limit of our NN model, although they remain visible to a human operator in the images. This hypothesis was also proposed by Theys et al. (2013). The limitation in detecting the more dilute portions of the volcanic cloud was previously discussed in Sect. 4.1 and 4.2, where the effect became evident toward the end of the sequence. This is expected, as dilution increases over time as the cloud disperses, resulting in more false negatives for the NN model.

The low overall recall values for the complete image sequence in Event 1 and Event 2 contrast with the recall values obtained for the individual images within the sequence, as shown in Fig. 8a–c, where high recall values are observed during the early stages of the volcanic clouds.

On the other hand, for Event 3, the behaviour of the precision and recall metrics is opposite to that observed in the previous events. In this case, the precision is lower at 78.1 %, while the recall is significantly higher at 84.2 %. In other words, the recall indicates that in Event 3, the filtered NN plume mask missed only 15.8 % of the VC pixels observed in the reference plume mask. Furthermore, the precision reveals that 78.1 % of the pixels detected as VC were correctly identified, indicating that 21.9 % of the detected pixels were false alarms.

These results may be attributed to the challenging nature of Event 3, in which the volcanic cloud is mixed with and surrounded by meteorological clouds. Under these circumstances, the human operator responsible for generating the manual plume mask may have faced difficulty in identifying the volcanic cloud and distinguishing it from meteorological clouds. Despite this complexity, the NN model successfully detected the volcanic cloud throughout the sequence (see the Video 3 in the Supplement).

The issue of the false negatives, likely caused by the dilution of the volcanic cloud is evident when examining the confusion matrices for the complete images sequence for each event, shown in Fig. 8d–f. These confusion matrices are constructed using the predictions (detections) of the filtered NN plume mask and the observed (reference) plume mask (Fawcett, 2006). In Fig. 8d–f the predictions are on the y axis and the observations on the x axis.

As previously discussed, the effect is more pronounced in the Event 1 and Event 2, which reach overall false negatives of 16 000 and 38 000 pixels, respectively, for the complete sequence. Meanwhile, for Event 3 the false negatives are 4500 pixels. It is important to note that Events 1 and 3 have similar durations, lasting approximately 6 h, whereas Event 2 lasted approximately 12 h. As a result, the total number of SEVIRI images analysed varied among the three events.

The question arises as to how the most challenging event achieved lower false negatives. We suggest that the answer lies in the generation of the manual plume mask. When the volcanic cloud is difficult to distinguish from meteorological clouds, the reference plume mask tends to be smaller and may not extend to the more dilute portions of the cloud. This situation reveals not only the potential of the NN model but also the limitations of the human operator under challenging conditions.

Furthermore, as illustrated in Fig. 8g–i, the number of false negatives progressively increases over time, as expected under the hypothesis that the dilution of the volcanic cloud leads to an increase in false negatives predictions. Consequently, lower recall values are obtained. This behaviour can be observed comparing Fig. 8a–c and g–i, where an increase in the number of false negatives corresponds to a decrease in the recall metrics.

Full animations illustrating the model's performance over the complete sequence are provided in the Supplement, including one animation for each validation event. These animations allow spatial visualization of true positives, false negatives, and false positives.