Developed in the context of TOAR phase II, the TOAR-II database stands as one of the world's largest collections of near-surface ozone measurements and related information. The database can be accessed through web services which provide a comprehensive suite of ozone-related data products including standard statistics, health and vegetation impact metrics, and trend information (https://toar-data.fz-juelich.de/, last access: 24 June 2026). The TOAR-II database includes extensive information describing the locations of air quality measurement stations based on pollution-relevant properties. These properties are extracted from EO data and stored as station metadata in the database. This metadata offers contextual information about the measurement site, enabling station location characterization. Table  below summarizes all metadata used in this study including references to the data origin.

Some metadata, such as station coordinates, are provided by many air quality agencies and scientific institutions that contribute to the TOAR database. The other metadata elements listed in Table  stem from EO datasets, which were downloaded from the respective provider sites. A special web service called Geospatial point extraction and aggregation service (GeoPEAS) has been developed to compute the aggregate information from the original gridded products. More information about the EO datasets used in GeoPEAS can be found on https://toar-data.fz-juelich.de/api/v2/#stationmeta (last access: 24 June 2026). After the metadata extraction with GeoPEAS, all metadata are available as lists of key-value pairs with the keys corresponding to the variable names in Table . For further processing, this data was collected into one table with the keys as data columns and the individual stations as rows.

The first step in our data preprocessing pipeline consists of cleaning the dataset. This involves removing all duplicate data points, replacing values of -999.0 with NaN to denote missing values, and eliminating rows where all metadata information is missing. To ensure consistency, we filtered out rows with inconsistent values, such as negative population density or negative maximum stable lights. In total, this step eliminates 211 stations out of 24 348 stations. For handling missing altitude data, we fill these with mean_topography_srtm_alt_90m_year1994. Other numerical missing values are estimated using the regression iterative imputer , and categorical missing values fill with most frequent instance. Categorical variables are encoded using OneHotEncoder from scikit-learn .

For K-means clustering, numerical features were scaled using standard scaling, and outliers were handled with IQR-based clipping within the range [Q1 − 1.5 ⋅ IQR, Q3 + 1.5 ⋅ IQR], where IQR (Interquartile Range) is the difference between the third quartile (Q3) and the first quartile (Q1). These preprocessing steps are essential to mitigate the algorithm's sensitivity to feature scale and outliers. For supervised learning algorithms, we applied a robust scaler to the entire dataset, which proved more effective for this task compared to alternative scaling methods such as standard or min-max scaling. This scaling approach helps mitigate the impact of outliers and ensures consistent feature ranges. In the feature selection process, we prioritized variables containing the most recent available information, ensuring our model utilizes the most up-to-date data for classification. Notably, geographical coordinates (longitude-latitude) and station codes are excluded from the machine learning models to prevent overfitting to geolocation. Following preprocessing steps, the final dataset comprises 24 125 stations. Of these, 13 225 are labeled as urban, suburban, or rural, while the remaining 10 900 are unlabeled. The distribution of the processed data is presented in Fig. .

For the K-means clustering analysis, we allocated 22 111 samples for model training and reserved 15 % of the labeled data (1979 samples) for testing. The supervised models were trained on a smaller subset because only 13 225 stations were explicitly classified as urban, suburban, or rural by the TOAR data providers. The remaining 10 900 stations lacked this specific classification, being reported as “unknown” or without a designated category. We used 85 % of the labeled data (11 211 samples) for training while the remaining 15 % is allocated for testing. The distribution of training and testing datasets is presented in Fig. . As shown, the training dataset is imbalanced, with a higher number of urban stations compared to suburban and rural ones. However, the class imbalance is not severe. We address it by applying the Synthetic Minority Oversampling Technique (SMOTE) to the training data before fitting the supervised classifiers. It is important to note that SMOTE was applied exclusively to the training set and not to the test data. The trained classifier is then used to predict the characteristics of unlabeled stations as illustrated in Fig. .

To ensure the reliability of our machine learning approach, we manually selected 35 stations with clear decision boundaries – that is, stations that are easy to classify as urban, suburban or rural, and excluded them from the training dataset. These stations were explicitly reported by data providers as urban, suburban, or rural, and we used Google Maps to manually verify and label them. During this process, we observed discrepancies between the labels provided by the data providers and those derived from our manual Google Maps analysis. Our models will first be evaluated on these 35 stations, with accuracy calculated both against the labels reported by the data providers and against the manually verified labels from Google Maps (referred to as hand-labeled data).

This section is devoted to a concise overview of the machine learning techniques employed in this study. Additionally, we describe the evaluation metrics used to quantify the effectiveness and accuracy of our models.

Figure  shows the elbow plot, a heuristic technique used to determine the optimal number of clusters for K-means clustering. As the gradient of classification accuracy flattens at 3 to 4 clusters, these values for K represent the optimal choices. This result is very encouraging since we want to distinguish 3 different types of stations. As a first analysis of the K-means clustering, Fig. b shows the K-means predictions evaluated on 35 manually selected and labeled stations, with clear decision boundary from different categories for sanity check of our method. Table presents the accuracy of K-means predictions on these manually labeled stations from the test set, comparing them with the characterization report from the TOAR database.

Figure a presents the confusion matrix computed from the K-means prediction and the classes from the TOAR database as ground truth, based on 1979 unseen test stations (see Sect. ). This confusion matrix highlights the high accuracy of K-means in classifying urban and rural stations (87.67 % and 71.88 %, respectively, see Table , Recall's column). However, many instances exist where stations reported as urban or rural are classified as suburban, and vice versa. This misclassification is primarily due to the subjective nature of defining suburban areas, which often lie at the interface between rural and urban regions or exhibit a mix of urban-suburban or rural-suburban characteristics. Additionally, Fig. b visualizes the clusters defined by K-means for the 1979 unseen test stations. This visual representation clearly illustrates the fuzzy boundaries between the clusters and the noticeable spacing among the three centroids (depicted as red cross).

Here, we evaluate the results from the three supervised machine learning classifiers, Random Forest, LightGBM, and CatBoost. Furthermore, the results from the three models were subjected to a robust voting classifier to maximize the classification accuracy. We observed that the results of all algorithms are quite similar. The models demonstrated exceptional accuracy (> 83 %), high precision (> 85 %), and high F1-score (> 81 %) in predicting urban and rural areas. However, all models struggled with the suburban class, yielding accuracies slightly above 60 % (Table ). As discussed above, the main reason for this lower accuracy can be attributed to the inherent subjective nature of defining this category. To address this issue, we implemented a strategy that capitalizes on model uncertainty. By adjusting the prediction probability threshold, as detailed in Sect. 2, we significantly enhanced the accuracy of suburban area classifications as shown in Table . Figure a presents the confusion matrix for the Random Forest classifier and Fig. b visualizes the feature importance. The global evaluation, i.e accuracy, balance accuracy, macro-F1 score, and weighted-F1 score of different classifiers are reported in Tables  and , which show the results before and after the probability threshold adjustment, respectively. While the overall accuracy remains relatively similar before and after the adjustment, the probability threshold adjustment significantly enhances the prediction accuracy for suburban stations, increasing it from a range of 60.85 %–65.72 % to 66.22 %–71.95 %, and also slightly increase the F1-score for the suburban, these can be seen in details in Tables  and presenting the per-class evaluation before and after applying the probability threshold adjustment. We also note a slight drop in accuracy when classifying urban and rural stations. However, classification performance remains high across all classifiers, with accuracy values exceeding 80 %. Additionally, we conducted tests on our machine learning models using the manually labeled stations, similar to those used for K-means evaluation. In this test, we found that the classifiers predict the label report on TOAR by data provider for 35 manually selected stations with 100 % accuracy and achieve an 87.88 % accuracy for the manual classified stations.

The implementation of the adjusted probability threshold yielded notable improvements in our classification model. While the enhancements for urban and rural station predictions were modest, the impact on suburban area classification was substantial. This is particularly significant given the inherent challenges in accurately identifying suburban zones. When compared to the unsupervised K-means clustering method, our supervised approaches demonstrated superior performance across all categories. The contrast was especially pronounced in the classification of suburban areas, where K-means exhibited a markedly low accuracy of just 15.84 %, low precision of 33.33 % and low F1-score of 21.47 %, see Table . In contrast, our supervised methods achieved significantly higher accuracy rates, higher precision, and higher F1-score underscoring their effectiveness in navigating the complexities of urban-suburban-rural distinctions.

The supervised machine learning approach demonstrates remarkable performance, achieving prediction accuracies > 84 % for urban and rural stations when applied to previously unseen test data. This can already be used to accurately predict “urban” and “rural” labels. While the model's performance in identifying suburban areas initially showed slightly lower accuracy, this challenge was effectively addressed through the adjusted probability threshold.

Despite the promising results, the classification results are far from perfect. This can be partially attributed to inherent inaccuracies within the dataset itself. To investigate this issue, we conducted a detailed review and manual inspection of the 30 randomly selected misclassifications. For these stations, which are listed in Table , we visually inspected the areas around the stations on Google Maps , using a zoom level of 11 or greater. While 6 of these cases revealed wrong classifications by our best ML model, the model's classification is actually more accurate than the label that was reported by the data providers in 16 cases. In the remaining 8 cases, neither the reported nor the ML model derived label was correct. In one of these cases, both the reported and ML based label was urban, while the station site is apparently located in a rural area. In the other case, visual inspection would place the station in the suburban class, while the reported category is urban and the ML model classifies the station as rural. However, for stations misclassified by our machine learning model, we observed ambiguous features that could misguide our model. For instance, surrounding neighborhoods of areas reported as urban but classified as rural by our model exhibited rural characteristics, such as lower population density (which is one of the important features used in the training dataset, see Fig. b) and more green space. This result, while initially counter-intuitive, can be explained by the strong robustness of tree-based ensemble methods to noisy datasets. Algorithms such as Random Forest, CatBoost, and LightGBM are specifically designed to mitigate the effects of noise and outliers through randomization and averaging techniques . However, these 30 data points are insufficient to draw definitive conclusions.

To further lend confidence to our results, we evaluated the 75-percentile statistics of the primary air pollutant concentrations NO_x and PM_2.5 from the TOAR database. We chose the 75th percentile, because urban areas typically exhibit fresh pollution with many high concentration events. This percentile captures such characteristics while being more robust than either the maximum value or a higher percentile. While data on these species is incomplete, there are sufficient measurements from several regions to yield a meaningful statistic. Figure  shows box and whisker plots of the 75-percentiles of NO_x and PM_2.5 concentrations aggregated for the year 2015 for the three classes. As expected, urban stations typically show substantially higher concentration levels compared to suburban sites, while rural stations show the lowest concentrations.