Comprehensive monitoring of NO

Air pollution is the leading environmental risk factor globally

During the last few decades, several approaches have been developed to estimate NO

Physics-based urban air quality models can generate hourly pollutant concentration estimates, overcoming the temporal limitation of microscale LUR models. Currently, these systems usually consist of the coupling between a regional chemical transport model, which accounts for the long-range transport of pollutants, and an urban-scale dispersion model. The latter can be based on semi-empirical relations, such as Gaussian dispersion models and mass exchange global parameterizations

In order to reduce model uncertainties, data fusion methods can be employed to post-process model outputs and obtain more reliable NO

By combining model and observational data, advanced data fusion methods can provide typically unbiased estimates of pollutant concentrations at the street scale. However, another piece of information that is of crucial importance is the uncertainty in the estimated concentrations, as it can help with decision-making or support the design of environmental epidemiological studies

Our study presents a data fusion methodology considering a microscale LUR model, in addition to the hourly monitoring data, to bias correct hourly NO

The paper is structured as follows: the observational data and study domain are described in Sect.

We compare two data fusion methods (UK-DM and UK-DM-LUR; illustrated in Fig.

Workflow of the two studied data fusion methodologies. Hourly data from monitoring stations are combined with hourly dispersion model results (UK-DM) and the time-invariant microscale LUR basemap (UK-DM-LUR). PBL stands for planetary boundary layer.

Barcelona (Fig.

Hourly NO

Domain of study and location of the referenced monitoring stations. The map has been generated using the ggplot2

Two different NO

Hourly high-resolution concentrations of surface NO

At the regional scale, CALIOPE-Urban relies on the regional air quality modeling system CALIOPE

In this work, CALIOPE-Urban employs a non-uniform mesh that is refined at the edge of traffic roads and coarser in low-gradient regions of NO

A nonlinear microscale LUR model based on passive dosimeter campaigns is used to produce an observation-based climatological view of the NO

The target variable of the microscale LUR model is the time-averaged concentrations of the two different NO

Sampler locations of the two different NO

The potential predictors of the microscale LUR model are shown in Table

The microscale LUR model contemplates the use of these 21 potential predictors.

A recursive feature elimination method has been applied to remove highly correlated or uninformative features. We have used the simple backward-selection algorithm implemented in the R package caret

To account for nonlinear relations among the predictors and the target variable, we used the gradient boosting machine (GBM) algorithm implemented in the R package gbm

One could think that skipping over the LUR computation by directly using all of its time-invariant information (passive dosimeter campaigns, urban geometry, traffic-related data, and annually averaged model results) as covariates in the universal kriging methodology would simplify the workflow. However, there are two main drawbacks to doing so. On the one hand, in contrast to GBM, universal kriging assumes linear relations between covariates and the observed NO

The microscale LUR model and the hourly CALIOPE-Urban outputs are combined with observational NO

As a Gaussian process, universal kriging estimates the variance of its predictions (

To normalize the distribution of the NO

Assuming a normal distribution of the error, the probability of exceedance (

Statistical performance is assessed by leave-one-out cross-validation (LOOCV), which consists of performing the data fusion by considering all of the monitoring stations, except for one that is kept to cross-validate the results. For each LOOCV, we present the coefficient of efficiency (COE), the root mean square error (RMSE), the mean bias (MB), and the correlation coefficient (

In the universal kriging context, the variogram describes the spatial autocorrelation structure of the residual random field. In our case, the limited number of monitoring stations makes it challenging to extract a meaningful spatial structure. For this reason, we estimate the residual variogram based on the dosimeter campaigns. This decision, however, entails a substantial limitation due to the assumption of a static variogram. We rely only on the IDAEA-CSIC campaign (discarding the xAire campaign for the variogram derivation) to avoid extra premises for the combination of campaigns. Additionally, we considered an isotropic variogram. All of these postulates impact the variance error estimated by universal kriging

The variogram is fitted using the Matérn model with Stein's parameterization implemented in the R package automap

The correlation coefficient (

The results are organized into two sections. First, in Sect.

The GBM-based microscale LUR model is evaluated using two nested

Scheme of the outer 10-fold CV and the inner 4-fold CV applied for the GBM training.

The results are given in Table

Statistical results of the microscale LUR model in nested CV. The 2017 annual mean concentration of NO

Table

In Fig.

Comparing these results with previous works, the resulting correlation coefficient (

We proceed to train the microscale LUR model with the residual correction using all available sampling sites. Figure

The influence of each predictor in the final microscale LUR model has been computed based on the methodology proposed by

In order to quantify the added value of including the microscale LUR basemap in the data fusion methodology, two different post-processes (see Fig.

Statistical results for each station after applying UK-DM and UK-DM-LUR to 2019 hourly data in LOOCV. In addition, we show the statistical results for the CALIOPE-Urban estimates at each station. All stations refer to the average over all stations.

Hourly statistical results for the raw CALIOPE-Urban, UK-DM, and UK-DM-LUR models are shown in Fig.

The uncertainty in the universal kriging predictions is estimated from the (multi-)linear regression and the spatial interpolation variances, as formulated in Eq. (

Percentages of observations falling in the

To better understand the behavior of uncertainty estimates, in Fig.

PDF of the hourly bias, normalized by the universal kriging standard deviation, for all monitoring stations in LOOCV during 2019. The PDFs correspond to the reference normal distribution (

In addition, in Fig.

PDFs by observed NO

We first analyze the annual mean concentration levels of NO

Statistical results using the 12 monitoring stations after applying UK-DM and UK-DM-LUR directly to the annual averages in LOOCV.

Figure

As expected, the areas surrounding the monitoring stations (presented in Fig.

Regardless of the data fusion method, the most polluted regions correspond to probabilities exceeding the annual limit above 0.7, as shown in Fig.

Figure

The present work assesses the added value of including a microscale LUR basemap into a data fusion method to obtain spatially bias-corrected urban maps of NO

The statistical performance of the microscale LUR model has been assessed using a comprehensive nested CV. As expected, the obtained microscale LUR basemap (

Adding the microscale LUR time-invariant spatial information (UK-DM-LUR) has been demonstrated to significantly improve the skills of the more straightforward data fusion UK-DM method at the hourly scale, increasing the

To check the consistency of the estimated uncertainty, we have empirically validated the universal-kriging-based uncertainties through a LOOCV. Despite the predicted variance of the universal kriging being slightly overconfident and tending to degrade for extreme concentration values, we found that it is a meaningful estimation of uncertainty. The PDFs of the error are close to the normal distribution, especially for the UK-DM-LUR approach. The spatial characterization of the uncertainty adds value to the NO

In developing our microscale LUR model, a limitation arises when using campaigns conducted between February and March. Although the annualization adjustment factor corrects the NO

Local authorities frequently conduct air quality diagnoses based solely on available monitoring stations, resulting in inaccurate assessments of the situation, since numerous local pollution hotspots remain unmonitored. We have shown that data fusion methods can provide a more comprehensive analysis by minimizing the sampling bias. For instance, in 2019, only the Gràcia and Eixample stations exceeded the annual legal NO

A strong point of the presented methodology is the characterization of the NO

An assessment of the passive dosimeters data needed for the present data fusion methods is provided here. Despite the specificities of the data, this assessment is intended to aid in the transferability to other cities. First, Sect.

We have calculated the effect of using campaigns from different years at two distinct levels, namely effects on the microscale LUR performance and effects on the overall data fusion workflow performance (UK-DM-LUR).

Applying the performance evaluation procedure described in Sect.

Statistical results of the microscale LUR model in nested CV, considering both campaigns or only one of them. The 2017 annual mean concentration of NO

The microscale LUR model, based solely on the CSIC campaign, exhibits superior performance compared to the model based on both campaigns, whereas the model based solely on the xAire campaign demonstrates the opposite trend. However, there are notable differences in the number of data points and the motivation behind each campaign. The CSIC campaign deployed fewer samplers (175), which raises concerns about possible overfitting. In this line, the COE statistic shows a significant decline (

Figure

Statistical results for each station after applying UK-DM and UK-DM-LUR to 2019 hourly data in LOOCV. For the UK-DM-LUR application, we have considered developing the microscale LUR model with only one experimental campaign (UK-DM-LUR CSIC or UK-DM-LUR xAire) or both of them (UK-DM-LUR). In addition, we show the statistical results for the CALIOPE-Urban estimates at each station. All stations refer to the average over all stations.

Regardless of the configuration, UK-DM-LUR improves the UK-DM methodology (and, therefore, CALIOPE-Urban) for the COE,

Statistical results of the 15 microscale LUR models in nested CVs. The models are built by considering both dosimeter campaigns and gradually increasing the number of samplers from 140 to 790 by uniform increments of 50 random samplers, in addition to the final model with all of the samplers (844). The statistics represent the evaluation of the microscale LUR models for the training and test (with and without the correction of the residuals) sets. The 2017 annual mean concentration of NO

In the case of using two campaigns, we have computed the microscale LUR performance by gradually increasing the number of samplers from 140 to 790 with uniform increments of 50 random samplers, which results in 14 new models. In addition, we have also added the final model with all samplers (844) to make a comparison. To ensure the robustness of the results, we repeated these computations three times, randomly varying the selected samplers. Then, from these three series, the average and the standard deviation of the statistical indicators are computed. Figure

As expected, as more samplers are considered, the standard deviation of the different metrics decreases. Also, an increasing trend in COE and

The source code and the results, including the final kriging post-processed product (predicted concentrations, uncertainties, and exceedances) are publicly available via Zenodo at

AC implemented the data fusion code and generated the figures. AC and JMA conducted the study and wrote the draft. JMA and JB processed the CALIOPE-Urban data. HP supported the validation of the microscale LUR model. All authors contributed to the analysis and objectives of the document and internally reviewed and edited the text.

The contact author has declared that none of the authors has any competing interests.

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

This article is part of the special issue “Air quality research at street level (ACP/GMD inter-journal SI)”. It is not associated with a conference.

The authors would like to thank Direcció General de Qualitat Ambiental i Canvi Climàtic – Generalitat de Catalunya, for providing observational data through the XVPCA, and IDAEA-CSIC, for providing the experimental dosimeters campaign data. The BSC researchers thankfully acknowledge the computer resources at Marenostrum and the technical support provided by Barcelona Supercomputing Center.

We have received support from the Barcelona City Council through the UncertAIR project (ID 22S09501-001; Recerca Jove i emergent 2022). This research has been supported by the Ministerio de Ciencia e Innovación through the BROWNING project (grant no. RTI2018-099894-BI00), the Agencia Estatal de Investigación as part of the VITALISE project (grant no. PID2019-108086RA-I00) and the MITIGATE project (grant no. PID2020-116324RA695 I00), the H2020 Marie Skłodowska-Curie Actions (grant no. H2020-MSCA-COFUND-2016-754433), the AXA Research Fund, and the Barcelona Supercomputing Center (grant nos. RES-AECT-2021-1-0027 and RES-AECT-2021-2-0001).

This paper was edited by Christoph Knote and reviewed by two anonymous referees.