Validation of the 2021 Emission Inventory for Cuenca, Ecuador, Through Weather and Air Quality Modeling in the Framework of WRF-Chem

Abstract

The last atmospheric emission inventory for Cuenca, a city located in the Andean region of southern Ecuador, was developed for the year 2021 (EI 2021) and encompasses both primary pollutants (NO_x, CO, VOC, SO₂, PM₁₀, and PM_2.5) and greenhouse gases (CO₂, CH₄, and N₂O). We formally assessed the quality of this emission inventory by modeling air quality levels in October 2021 using the Weather Research and Forecasting with Chemistry (WRF-Chem 3.2) model at a high spatial resolution (1 km). Although we conducted simulations with different combinations of boundary conditions for chemical species and hourly profiles to disaggregate daily on-road traffic emissions, we selected two sets of numerical experiments to report. The results indicated that most of the assessed meteorological and air quality variables were modeled acceptably, suggesting that the EI 2021 emission inventory is a reasonable estimation of real emissions. The results also indicated the current capacity of WRF-Chem to model the atmosphere in a complex Andean city using the “one atmosphere” approach, highlighting the variables with good, fair, and poor modeling performance. We propose future research directions to improve emission inventories and the performance of atmospheric modeling in the Equatorial Andean region. Finally, the results and spatial distribution of the EI 2021 were compared to the emission data from the last version of the EDGAR Emissions Dataset (which has a spatial resolution of 11.1 km), one of the most used global emission datasets. We concluded that, for the Equatorial Andean region and for modeling purposes, the EDGAR Dataset results should be reviewed, both to account for the effects of height above sea level on the magnitude of primary pollutant emissions and their spatial configuration.

1. Introduction

Atmospheric emission inventories provide relevant information on the amount or rates of air pollutant and greenhouse gas (GHG) emissions produced by different sources over a specific region in the past, present, or future [1,2]. Atmospheric emission inventories have been used both for policy and research purposes. They provide temporal behavior data that can be used to track the progress in achieving reductions in emissions, e.g., refs. [3,4]. On the other hand, emission inventories provide input for chemical transport models, which simulate the behavior of pollutants in the atmosphere for both current and hypothetical emission scenarios, as well as for atmospheric forecasting [5,6,7]. Emission data are probably the most important input for chemical transport models [8].

The building and use of accurate atmospheric emission inventories are directly related to several Sustainable Development Goals (SDGs), such as SDG 3 (Good Health and Well-Being), by identifying the main air pollution sources and guiding air quality regulations; SDG 11 (Sustainable Cities and Communities), by supporting city-level policies for cleaner transport, energy, and waste management; and SDG 13 (Climate Action), by quantifying GHGs and other pollutants involved in the energetic balance of the atmosphere, identifying their emissions sources and sinks, and tracking mitigation progress [9].

International initiatives aim to develop and make the results of emission inventories available to inform scientists and policymakers. The EDGAR Emissions Database for Global Atmospheric Research is a dataset that covers time-series emission inventories for anthropogenic sources of primary pollutants and GHGs for all countries, with a spatial resolution of up to 0.1° (approximately 11.1 km) [10]. The EDGAR Emissions Database is one of the most widely used global atmospheric emission inventories, providing, in addition to annual estimation of primary pollutants and GHGs, temporal profiles for disaggregating annual emissions into monthly and hourly data by country and emission source. The CAMS emission inventory contains gridded distributions of global anthropogenic and natural emissions of atmospheric pollutants and GHGs, with a resolution of up to 0.1° [11].

Although emission inventories have been developed for many years, they still can have high levels of uncertainty [12,13,14]. Apart from the uncertainties inherent in emission models as simplifications of real emissions, other factors, such as the limitations of statistical data and the lack of local emission factors, can also significantly contribute. Uncertainty is also influenced by the method used to downscale total emission results to a grid-cell level, which is accompanied by high-time resolution disaggregation in order to use the data in a chemical transport model [15]. An assessment of uncertainty is required to clarify the usefulness of emission inventories and identify the sources that should be targeted in the future [16].

The quality of an emission inventory has typically been defined by the quality of the statistical data, emission factors, and models used in its development. Additionally, when available, it should be compared with other emission inventories. Another approach to assessing the quality of an emission inventory is to use it as input for atmospheric modeling. Mathias et al. (2017) [8] highlighted the need to work with accurate emissions data and provided an outlook for improving the spatial and temporal distribution of emissions inventories for use in chemical transport models. Park et al. (2023) [14] assessed emission inventories for sulfur dioxide (SO₂) and nitrogen oxides (NO_x) from large point sources in South Korea, evaluating the modeling performance through a comparison of computed and recorded data for these pollutants. Malasani et al. (2024) [17] assessed five global emission inventory datasets to determine the spatial and temporal distribution of mercury over India and also evaluated the modeling performance by comparing modeled and measured concentrations of this pollutant.

Online coupled meteorology–atmospheric chemistry models, which consider the influence and feedback between atmospheric and air quality variables, have undergone significant evolution in recent years [18]. The online approach provides consistent treatment of physical and chemical processes for both numerical weather and air quality modeling, allowing for a “one atmosphere” approach to the analysis. If the meteorological component is properly modeled and the comparison of the modeled air pollution levels with the corresponding records indicates consistency, we can deduce that the emission inventory is reliable and provides useful information, though with some uncertainties.

For Cuenca, a city located in the Andean region of southern Ecuador, several emission inventories have been developed since 2007. The results of the 2014 emission inventory [19] were used to conduct several modeling experiments to assess the influence of the parameters and options coded in the Weather Research and Forecasting with Chemistry (WRF-Chem 3.2) model. These numerical experiments employed the online option and proposed a configuration of parameters for modeling both meteorological and air quality variables using the “one atmosphere” approach [20,21] in urban areas of the Equatorial Andean region.

The last emission inventory for this city was created using 2021 as the base year [22], which is hereafter referred to as the EI 2021. Although its uncertainty was assessed, the purpose of this contribution is to evaluate the quality and usefulness of the EI 2021 by incorporating its results as input to WRF-Chem for “one atmosphere” modeling of both meteorological and air quality variables. Additionally, the EI 2021 results are compared with the corresponding results from the EDGAR Emissions Dataset.

1.1. Location and the Air Quality-Monitoring Network of Cuenca

Cuenca, located in the Andean region of southern Ecuador (Figure 1), is characterized by a complex topography and diverse land-use configuration. The urban area is located at 2550 masl; however, the Andes Mountains, to the west of the city, have heights exceeding 4000 masl (Figure 1c). Its air quality-monitoring network has been operational since 2008, in accordance with national regulations. Currently, the air quality-monitoring network is operated by the EMOV EP (Empresa Pública Municipal de Movilidad, Tránsito y Transporte de Cuenca). Between 2012 and 2024, the MUN station (the city’s historic center) measured yearly mean PM_2.5 concentrations; they were found to range between 5.7 and 14.5 µg m⁻³ [23], which is higher than the current World Health Organization (WHO) recommendation (5.0 µg m⁻³) [24]. Typically, O₃ levels are highest in September and October, mainly due to the elevated solar radiation, and in some years, forest fires contribute to O₃ precursor emissions. More information about the air quality-monitoring network is included in Appendix A.1.

1.2. The Emission Inventory of the Year 2021 (EI 2021)

The last atmospheric emission inventory for Cuenca was compiled for the year 2021 by the EMOV EP [25] in accordance with the recommended practice of accounting for primary pollutants and GHGs [13]. Nitrogen oxides (NO_x), carbon monoxide (CO), volatile organic compounds (VOCs), sulfur dioxide (SO₂), and particulate matter with an aerodynamic diameter of 10 µm or less (PM₁₀) and 2.5 µm or less (PM_2.5) were included as primary pollutants. In addition, carbon dioxide (CO₂), methane (CH₄), and nitrous oxide (N₂O) were included as GHGs.

Sources included on-road traffic, vegetation, industries, use of solvents, service stations, domestic combustion of liquid petroleum gas (LPG), air traffic, landfills, artisanal production of bricks, dust resuspension, and mining activities in the territory of the Cantón Cuenca (Figure 1c).

The EI 2021 was developed to estimate real emissions as accurately as possible in the territory of Cuenca, with the main goals of providing a proper emission inventory that can serve as a reference for air pollution management and the development of a future atmospheric forecasting system. The EI 2021 includes hot, cold, and evaporative emissions from on-road traffic [26], the VOC emissions from vegetation (based on the model by Guenther et al. [27,28]), and emissions from landfills [29].

The EI 2021 was published as a technical report [25]. More information on the EI 2021 is provided in Appendix A.2.

2. Methods

2.1. Modeling Approach

Using WRF-Chem 3.2 [30], previous modeling experiments were conducted for Cuenca to analyze the influence of planetary boundary layer, land-surface, cumulus, and microphysics schemes [20,21]. Based on these studies, the recommended set of schemes and options for modeling both atmospheric and air-quality variables (Table 1) was used to validate the EI 2021. WRF-Chem is a state-of-the-art numerical 3D model used for research and forecasting purposes that enables online analysis by considering interactions between meteorological and air quality variables through the “one atmosphere” approach.

First, a master domain which covers the entire Ecuadorian continental territory (grid of 70 × 70, 27 km cells; Figure 1a) and two nested subdomains (Figure 1b) (first subdomain: 52 × 52, 9 km cells; second subdomain: 61 × 42, 3 km cells) were used to conduct the meteorological simulations from 1 to 30 of October 2021. A third and inner subdomain of 100 × 82 cells (1 km on each side) covering the territory of Cuenca (Figure 1d), which corresponds to the grid used for the development of EI 2021, was used to simulate meteorological and air quality variables. For vertical resolution, we used 35 vertical levels up to a top pressure of 50 hPa (altitude of approximately 20 km). The initial and boundary conditions were generated using the Global Operational Analysis (Final, FNL) [31]. The Carbon Bond Mechanism Z (CBMZ) [32] and the Model for MOSAIC (4 sectional aerosol bins) [33] were selected to speciate and represent the corresponding hourly emissions of gaseous pollutants and aerosols, respectively. The online option accounts for the direct effects of feedback between aerosols and meteorological variables.

To test the quality of the emission inventory and considering both the availability of records and pollution conditions, we selected October 2021 for modeling, which was the month with the highest O₃ records of that year [34] and had representative socioeconomic activities related to atmospheric emissions for most of that year. Typically, the maximum levels of O₃ occur in Cuenca during September and October, in part due to the highest surface solar radiation, which promotes VOCs emissions from both vegetation and anthropogenic sources, thereby promoting photochemical reactions. In addition, during these months, on-road traffic levels, both from gasoline and diesel vehicles, are high due to the reactivation of labor and academic activities after vacation.

On-road traffic emissions for October were estimated as a fraction of annual emissions using the ratio of October gasoline and diesel sales to total fuel sales at all service stations in Cuenca in 2021. Based on EMOV EP estimates of weekday-to-weekend traffic dynamics, we assumed 80% and 60% of weekday traffic on Saturdays and Sundays, respectively. These percentages were previously used and yielded coherent results when assessing the influence of parameters and schemes on the set of options listed in Table 1, which was used in this contribution.

Previous modeling exercises using WRF-Chem 3.2 in Ecuadorian cities indicated that the default VOC vertical profile would be higher for the Ecuadorian Andean region. Therefore, it was necessary to decrease the vertical concentrations originally coded in the model. In addition, because the magnitude of hourly emissions also affects modeling performance, we conducted various numerical experiments. From these results, we selected two sets of numerical experiments to report in this manuscript:

• Numerical Experiment 1 (NE 1): boundary conditions for chemical species for the inner domain were assumed to be 80% of the default concentrations specified in WRF-Chem V3.2. We used two basic hourly traffic profiles from the EMOV EP to characterize weekday and weekend on-road emissions.
• Numerical Experiment 2 (NE 2): boundary conditions for chemical species for the inner domain were assumed to be 70% of the default concentrations specified in WRF-Chem V3.2. For this experiment, we used the hourly traffic profile from the EMOV EP for weekdays and the hourly traffic profile from the EDGAR Dataset for Ecuador for weekends on-road emissions.

Appendix A.3 includes figures of the percentages used to estimate monthly and hourly on-road emissions for modeling.

Vegetation emissions were estimated on an hourly basis using the meteorological component of WRF-Chem for each day of October to simulate temperature and photosynthetically active radiation for each cell of the inner subdomain (1 km resolution). Emissions from combustion from large industrial facilities were assumed to be constant with a fixed hourly pollutant emission rate.

We conducted simulations for three-day periods, using six hours of spin-up time for each, and initialized the chemistry using the default WRF-ChemV3.2 idealized profile.

Table 1. Schemes and options for atmospheric modeling in the Ecuadorian Andean region [20,21].

Component	WRF Variable	Option	Model and References
Cumulus Parameterization	cu_physics	0	0 No Cumulus
Microphysics	mp_physics	3	WRF Single–moment 3–class [35]
Longwave Radiation	ra_lw_physics	1	RRTM [36]
Shortwave Radiation	ra_sw_physics	2	Goddard [37]
Surface Layer	sf_sfclay_physics	1	MM5 similarity [38]
Planetary Boundary Layer	bl_pbl_physics	1	Yonsei University [39]
Land Surface	sf_surface_physics	2	Noah [40]
Urban Surface	sf_urban_physics	0	No urban physics
Chemical Mechanism and Aerosol Modules	chem_opt	7	CBMZ and MOSAIC [32,33]

Table 1. Schemes and options for atmospheric modeling in the Ecuadorian Andean region [20,21].

Component	WRF Variable	Option	Model and References
Cumulus Parameterization	cu_physics	0	0 No Cumulus
Microphysics	mp_physics	3	WRF Single–moment 3–class [35]
Longwave Radiation	ra_lw_physics	1	RRTM [36]
Shortwave Radiation	ra_sw_physics	2	Goddard [37]
Surface Layer	sf_sfclay_physics	1	MM5 similarity [38]
Planetary Boundary Layer	bl_pbl_physics	1	Yonsei University [39]
Land Surface	sf_surface_physics	2	Noah [40]
Urban Surface	sf_urban_physics	0	No urban physics
Chemical Mechanism and Aerosol Modules	chem_opt	7	CBMZ and MOSAIC [32,33]

2.2. Metrics for Modeling Performance

We assessed the performance of the model in predicting the hourly surface temperature and hourly wind speed using the following metrics [41]:

$$\mathrm{G}\mathrm{E}={\displaystyle \frac{1}{\mathrm{N}}}{\sum }_{\mathrm{i}=1}^{\mathrm{N}}\left|\mathrm{P}\mathrm{i}-\mathrm{O}\mathrm{i}\right|$$

(1)

$$\mathrm{M}\mathrm{B}={\displaystyle \frac{1}{\mathrm{N}}}{\sum }_{\mathrm{i}=1}^{\mathrm{N}}\left(\mathrm{P}\mathrm{i}-\mathrm{O}\mathrm{i}\right)$$

(2)

$$\mathrm{I}\mathrm{O}\mathrm{A}=1-{\displaystyle \frac{{\sum }_{\mathrm{i}=1}^{\mathrm{N}}{\left(\mathrm{P}\mathrm{i}-\mathrm{O}\mathrm{i}\right)}^2}{{\sum }_{\mathrm{i}=1}^{\mathrm{N}}{\left(\left|\mathrm{P}\mathrm{i}-\mathrm{P}\mathrm{m}\right|+\left|\mathrm{O}\mathrm{i}-\mathrm{O}\mathrm{m}\right|\right)}^2}}$$

(3)

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{{\displaystyle \frac{1}{\mathrm{N}}}{\sum }_{\mathrm{i}=1}^{\mathrm{N}}{\left(\mathrm{P}\mathrm{i}-\mathrm{O}\mathrm{i}\right)}^2}$$

(4)

• GE: gross error, which measures the magnitude of the difference between modeled and observed values, regardless of whether the modeled values are higher or lower than the observed values.
  MB: mean bias, which quantifies the tendency of the model to over- or underestimate the values.
  IOA: index of agreement, which measures the degree to which the modeled values match the records, considering both the variance and magnitude of the differences.
  RMSE: root mean square error, which measures the overall difference between the model and the records, with a greater weight on significant differences.
  N: number of values.
  Pm: mean model value.
  Om: mean observed value.
  Pi: model value.
  Oi: record.

Table 2 indicates the ideal ranges and expected accuracies for these metrics, as suggested in [41], about the application of models under the European Union’s Air Quality Directive, which emphasizes the use of models combined with monitoring data for a range of applications about air quality management.

The model’s performance for daily rainfall was assessed using Equation (5).

$$\mathrm{P}\mathrm{d}\mathrm{c}\mathrm{m}={\displaystyle \frac{\mathrm{N}\mathrm{d}\mathrm{r}+\mathrm{N}\mathrm{d}\mathrm{w}\mathrm{r}}{\mathrm{T}\mathrm{m}\mathrm{d}}}·100$$

(5)

• Pdcm: percentage of days with or without rainfall from model.
  Ndr: number of days with rainfall (≥0.2 mm d⁻¹) in both records and model results.
  Ndwr: number of days without rainfall (<0.2 mm d⁻¹) in both records and model results.
  Tmd: total number of modeled days.

A rainfall intensity of 0.2 mm d⁻¹ was used to define a rainy day or a day with measurable precipitation [42]. For days without rainfall, we considered that the model captured the record if both were <0.2 mm d⁻¹. For days with rainfall, we considered the model to agree with the corresponding record if both values were ≥0.2 mm d⁻¹.

To assess the short-term air quality modeling performance, we used the hourly records for nitrogen dioxide (NO₂), PM_2.5, and O₃ from the MUN station to deduce the concentrations of these pollutants during specified periods of time (Table 2), according to national regulation and the WHO guidelines [24,43,44]. We considered the days having more than 70% of the corresponding hourly records. Based on Simon et al. (2012) about metrics used for air quality modeling [45], the last concentrations were used to compute the MB, RMSE, and the correlation coefficient (r):

$$\mathrm{r}={\displaystyle \frac{\frac{1}{\mathrm{N}}{\sum }_{\mathrm{i}=1}^{\mathrm{N}}\left(\mathrm{O}\mathrm{i}-\mathrm{O}\mathrm{m}\right)\left(\mathrm{P}\mathrm{i}-\mathrm{P}\mathrm{m}\right)}{\sqrt{\frac{1}{\mathrm{N}}{\sum }_{\mathrm{i}=1}^{\mathrm{N}}{\left(\mathrm{O}\mathrm{i}-\mathrm{O}\mathrm{m}\right)}^2}\sqrt{\frac{1}{\mathrm{N}}{\sum }_{\mathrm{i}=1}^{\mathrm{N}}{\left(\mathrm{P}\mathrm{i}-\mathrm{P}\mathrm{m}\right)}^2}}}$$

(6)

• r measures the strength of the linear relationship between the modeled and observed values.

To assess the long-term air quality model’s performance, we calculated the percentage of passive stations with a difference of less than 30% between modeled monthly concentrations and recorded data. Additionally, we calculated the percentage of records correctly predicted by the model based on the accuracy presented in Table 2.

Table 2. Metrics for modeling atmospheric variables [41,45].

Variable	Metric	Benchmark or Ideal Range	Accuracy
Hourly surface temperature	GE	<2 °C	±2 °C
	MB	(−0.5 °C, 0.5 °C)
	IOA	≥0.8
Hourly wind speed (10 mas)	RMSE	<2 m s−1	±1 m s−1
	MB	(−0.5 m s−1, 0.5 m s−1)
	IOA	≥0.6
Daily air quality: 24 h NO2 mean, 24 h PM2.5 mean, max. 8 h O3 mean	MB	0	±50%
	RMSE	0
	r	1
Monthly air quality: NO2 and O3			±30%

Table 2. Metrics for modeling atmospheric variables [41,45].

Variable	Metric	Benchmark or Ideal Range	Accuracy
Hourly surface temperature	GE	<2 °C	±2 °C
	MB	(−0.5 °C, 0.5 °C)
	IOA	≥0.8
Hourly wind speed (10 mas)	RMSE	<2 m s−1	±1 m s−1
	MB	(−0.5 m s−1, 0.5 m s−1)
	IOA	≥0.6
Daily air quality: 24 h NO2 mean, 24 h PM2.5 mean, max. 8 h O3 mean	MB	0	±50%
	RMSE	0
	r	1
Monthly air quality: NO2 and O3			±30%

2.3. Processing of the Data from the EDGAR Emissions Database

For comparison purposes, we identified and extracted information from cells within the global domain of the EDGAR Emissions Database (EDGARv8.1) [4,10,15,46] mesh covering the territory of the Cantón Cuenca (Figure 1c). We geoprocessed the EDGAR emission data for the year 2021. We totaled the emissions from these cells in the EI 2021 for the corresponding primary pollutants and GHGs.

3. Results

3.1. Meteorology

At the MUN station, although the model overestimated the surface temperature (MB = 0.61 °C for the NE 1 and 0.49 °C for the NE 2), the GE (1.46, 1.54 °C) and IOA (0.84, 0.83) were within the expected benchmark ranges for the two sets of experiments (

Table 3. Metrics for modeling meteorological variables. Bold numbers are in the benchmark.

Metric	MUN		SAY		Benchmark
Hourly surface temperature:
	NE 1	NE 2	NE 1	NE 2
GE	1.46	1.54	2.11	2.18	<2 °C
MB	0.61	0.49	0.74	0.62	(−0.5 °C, 0.5 °C)
IOA	0.84	0.83	0.86	0.85	≥0.8
Hourly wind speed:
	NE 1	NE 2	NE 1	NE 2
RMSE	1.04	1.08	NA	NA	<2 m s−1
MB	0.31	0.36	NA	NA	(−0.5 m s−1, 0.5 m s−1)
IOA	0.72	0.71	NA	NA	≥0.6

NA: Not assessed.

), indicating that the model captured the trends of the records (Figure 2a). However, at the SAY station, both the GE (2.11, 2.18 °C) and MB (0.74, 0.62 °C) were out of range. For both stations and the two sets of experiments, the model provided a close match between the records and modeled hourly temperatures (IOA ≥ 0.8).

At the MUN station, although the model showed a slight overestimation during afternoons (Figure 2b), the wind speed was acceptably modeled as all the metrics fell within the expected ranges (Table 3).

The model correctly predicted 78.4% and 60.3% of the hourly temperature records at the MUN and SAY stations, respectively, for the NE 1 (

Table 4. Percentage of records captured by modeling meteorological variables.

Parameter	MUN		SAY		SOL
	NE 1	NE 2	NE 1	NE 2	NE 1	NE 2
Hourly surface temperature	78.4	77.4	60.3	58.9	NA	NA
Hourly wind speed	75.8	76.1	NA	NA	NA	NA
Daily rainfall	46.7	55.7	80.0	73.3	73.3	66.7

NA: Not assessed.

). These percentages decreased to 77.4% and 58.9%, respectively, for the NE 2. Most of the wind speed records at the MUN station (75.8, 76.1%) were modeled adequately.

The model showed the highest accuracy for daily rainfall at the SAY station (80.0, 73.3%), followed by the SOL station (73.3, 66.7%) and the MUN station (46.7, 55.7%).

From 1 to 30 October 2021, 19 days had no rainfall or intensities lower than or equal to 0.2 mm d⁻¹ at the MUN station (Figure 3a). Eleven days showed measurements between 0.4 and 21.0 mm d⁻¹. In comparison, for the NE 1, the model identified 13 days with no rainfall or intensities lower than or equal to 0.2 mm d⁻¹ and 17 days with intensities between 0.4 and 30.5 mm d⁻¹. A better agreement was obtained at the SAY station, which had 20 days without rainfall or with rainfall intensities lower than or equal to 0.2 mm d⁻¹, similar to the days quantified by the model. At the SOL station, which is located over the Andean mountains, the model indicated that all the days (30) had intensities higher than 0.2 mm d⁻¹ while the records indicated 22 days. At the three stations, the ranges of the modeled rainfall were consistent with the ranges of the records.

The global solar radiation at the SOL (Figure 2c) and SAY was overestimated for the two numerical experiments.

3.2. Air Quality

Although the model provided a mean daily profile consistent with the records, it slightly overestimated the mean 24 h NO₂ levels (MB = 1.1 µg m⁻³) for the NE 1 but underestimated them (MB = −2.2 µg m⁻³) for the NE 2 (

Table 5. Short-term air quality metrics.

	NE 1	NE 2	Ideal Value	Number of Assessed Days (Total = 30 Days)	Percentage of Days for Assessing
24 h NO2 mean:
MB	1.1	−2.2	0	26	86.7
RMSE	10.3	9.6	0	26	86.7
r	0.1	0.2	1	26	86.7
24 h PM2.5 mean:
MB	−3.52	−3.76	0	29	96.7
RMSE	5.42	5.67	0	29	96.7
r	−0.11	−0.09	1	29	96.7
Maximum 8 h O3 mean:
MB	10.70	2.90	0	28	93.3
RMSE	17.00	16.90	0	28	93.3
r	0.60	0.1	1	28	93.3

, Figure 4a,b). However, in the two numerical experiments, the model did not show a strong linear relationship between the recorded and modeled values (r = 0.1 and 0.2). The model slightly underestimated the mean 24 h PM_2.5 levels (MB = −3.5 µg m⁻³ for NE 1, MB = −3.8 µg m⁻³ for NE 2) for the two sets, and there was a weak relationship between the observed and computed levels (r = −0.1 for both NE 1 and NE 2). Although the model exhibited a positive bias for modeling the maximum 8 h O₃ mean (MB = 10.7 µg m⁻³) for NE 1, the bias decreased (MB = 2.9 µg m⁻³) for NE 2. The two numerical experiments produced consistent mean daily O₃ profiles relative to the records, with better agreement for NE 2 (Figure 4f).

For NE 1, the model captured 84.6%, 75.9%, and 67.9% of the daily mean 24-h NO₂ levels, mean 24 h PM_2.5, and maximum 8 h O₃ levels, respectively (

Table 6. Percentage of records captured by modeling air quality variables.

	NE 1	NE 2
Short-term air quality:
24 h NO2 mean	84.6	84.6
24 h PM2.5 mean	75.9	72.4
Max. 8 h O3 mean	67.9	82.1
Average:	76.1	79.7
Long-term air quality:
NO2, monthly mean	81.3	56.3
O3, monthly mean	76.5	35.3
Average:	78.9	45.8

). For the NE 2, the model improved for maximum 8 h O₃ concentrations, capturing 82.1% of the records.

Although there were some differences, the model provided mean daily profiles for the three assessed pollutants that were consistent with the records (Figure 4b,d,f). On average, although the model fitted the hourly O₃ concentrations during the morning, it overestimated them at midday and in the afternoon (Figure 4f).

At the passive stations, the monthly mean NO₂ and O₃ concentrations were properly modeled in the NE 1 (81.3% and 76.5%, respectively; Figure 5a,b, Table 6). However, the performance decreased for the NE 2 (56.3% and 35.3%).

As an example of modeled air quality variables, Figure 6 illustrates the spatial distribution of the modeled hourly PM_2.5 levels on 17 October 2021 for the NE 1, which ranged between 3 and 45 µg m⁻³. The highest levels were computed at areas north and northwest of the urban area between 04:00 and 07:00 local time (L.T.), corresponding to the influence of the emissions from artisanal brick kilns (Figure A1). The concentrations decreased with atmospheric warming during daylight hours and increased again during the night, reaching up to 20 µg m⁻³ at 22:00 L.T. over the urban area.

For the NE 1, Figure 7 shows the map of modeled hourly O₃ levels on 17 October 2021. The highest levels were computed for the northwest and west areas outside of the urban area between 11:00 and 13:00 L.T., reaching levels of up to 130 µg m⁻³. At midday, the O₃ levels reached 100 µg m⁻³ over the urban area.

3.3. Comparison to the EDGAR Emissions Dataset

Table 7. Comparison to the results from the EDGAR Emissions Dataset. Year 2021.

	This Contribution	EDGAR Dataset	Difference
Primary pollutants	(t/y)	(t/y)	(%)
NOx	7554	6690	−11.4
CO	39,547	22,986	−41.9
VOC	13,214	4970	−62.4
SO2	1118	3851	244.5
PM10	1174	1558	32.7
PM2.5	822	1237	50.4
Greenhouse gases	(t/y)	(t/y)	(%)
CO2	1,142,016	1,004,333	−12.1
CH4	4997	10,815	116.4
N2O	100	392	291.9
Total greenhouse gases	(t CO2 equation/y)	(t CO2 equation/y)	(%)
CO2 equation	1,304,235	1,403,332	7.6
Greenhouse gases emissions per capita:	(t CO2 equation per capita/y)	(t CO2 equation per capita/y)	(%)
CO2 equation per capita	2.01	2.16	7.6

compares the EI 2021 results with the emissions for the same year from the most recent version of the EDGAR Emissions Dataset [10]. The total NO_X emissions were consistent, with a ratio of 1.13 between the EI 2021 and EDGAR Dataset values. Larger differences were observed for CO and VOC emissions, with ratios of 1.72 and 2.66, respectively. The ratios for SO₂ and PM_2.5 emissions were 0.29 and 0.66, respectively.

The CO₂ emissions were consistent, with a ratio of 1.14 between the EI 2021 value and the estimated value from the EDGAR Dataset. There were larger differences for CH₄ and N₂O, with ratios of 0.46 and 0.26, respectively. The ratio of total CO₂-equivalent emissions was 0.93. The per capita GHG emissions were consistent, with values of 2.01 t CO₂ equation per capita y⁻¹ for the EI 2021, and 2.16 t CO₂ equation per capita y⁻¹ for the EDGAR Dataset.

For the estimation of the GHG emissions in units of CO₂ equivalent, we used factors of 27 and 273 for the global warming potential of CH₄ (mostly non-fossil) and N₂O, respectively, which were taken from the sixth Intergovernmental Panel on Climate Change (IPCC) report [47].

Figure 8 depicts the spatial distribution of total CO emissions. The EI 2021 map, with a 1 km spatial resolution, provides a more accurate depiction of the emissions distribution. The map generated from the EDGAR Dataset, with a resolution of approximately 11.1 km, provides a coarser resolution and even mislocalized emissions, which was also the case for the other pollutants.

4. Discussion

We assessed the quality of the EI 2021 emission inventory for Cuenca by using it as input to the WRF-Chem model to simulate atmospheric and air quality variables at a high spatial resolution (1 km). To our knowledge, this is the first time that an atmospheric emission inventory from the Equatorial Andean region has been formally assessed by modeling meteorological (surface temperature, surface wind speed, and total daily rainfall) and air pollution variables (daily mean 24 h NO₂, daily mean 24 h PM_2.5, daily maximum 8 h O₃, and monthly mean NO₂ and O₃ levels). We used a state-of-the-art tool to model the direct effects of aerosols on meteorological variables and presented the results in two sets of numerical experiments, differing in boundary conditions for chemical species and in hourly profiles for disaggregating daily on-road traffic emissions.

The results and assessment depicted the current capacity of the WRF-Chem to model the atmosphere as a unique system in a complex Andean city [13,48], highlighting the variables with good, fair, and poor modeling performance. The NE 2 improved performance for the maximum 8 h O₃ daily means, although it decreased long-term air quality at passive stations compared to NE 1.

The comparison of the EI 2021 results with the emissions from the EDGAR Dataset indicated a good agreement for NO_x. In both cases, transportation and industry were the primary sources of NO_x.

The CO emissions from the EI 2021 were 1.7 times higher than those from the EDGAR Dataset, despite the latter including additional sectors, such as agriculture. We hypothesize that the difference can be partly explained by Cuenca’s elevation (2550 masl). Molina and Molina reported that, at 2240 masl, the Mexico City Metropolitan Area has 23% less oxygen (O₂) than at sea level [2]. Similarly, at the height of Cuenca, the lower O₂ availability affects combustion processes, leading to higher CO emissions [49,50].

The VOC emissions from the EI 2021 were 2.7 times higher than the estimate from the EDGAR Dataset. The main reason for this difference may be that the EDGAR Dataset covers human-made sources, whereas the EI 2021, in addition to anthropogenic sources, also includes VOC emissions from vegetation.

The SO₂ emissions from the EI 2021 were 0.29 times the EDGAR estimate. In both cases, industry and transportation were identified as the primary sources of SO₂. Because the EMOV EP conducted a dedicated sampling campaign at service stations to characterize the sulfur content of fuels used in Cuenca in 2021, we believe that the EI 2021 estimate is more reliable.

The PM₁₀ and PM_2.5 emissions from the EI 2021 were 0.75 and 0.66 times the estimates from the EDGAR Dataset, respectively. These ratios seem consistent, as the EI 2021 did not include emissions from agricultural activities.

The comparison of the EI 2021 results with the corresponding emissions from the EDGAR Dataset indicated a good agreement for CO₂. Important differences were identified for CH₄ and N₂O. The primary purpose of the EI 2021 was, in addition to estimating VOC emissions from vegetation, to focus on combustion and fugitive emissions from the primary human-made sources. Therefore, the EI 2021 did not include CH₄ sources from enteric fermentation or agricultural or livestock activities or N₂O sources from fertilizer use, all of which were included in the EDGAR Dataset.

Emission inventories are characterized by high levels of uncertainty [13]. In the case of the EI 2021, we identified components that deserve dedicated improvements in the future. In addition to the quality required for the activity data, defining emission factors for the Equatorial Andean region is a priority, considering the decrease in oxygen in the atmosphere with increasing height above sea level, which in turn increases the magnitude of emissions of primary pollutants such as CO, VOC, and particulate matter. There are very few domestic emission factors that have been characterized in Ecuador, and most of the factors studied are taken from the international literature, which has mainly focused on locations at sea level or at low elevations.

Although the modeled air quality concentrations were acceptable for most of the variables, the model’s performance can be improved. For the daily mean NO₂, although the model captured the corresponding records for 84.6% of days, it yielded a low correlation coefficient. Similarly, for the daily mean PM_2.5 levels, the model accurately predicted the values for 75.9% (NE 1) of days, although with a very low correlation coefficient.

As part of this research, we also used the hourly temporal profiles from the EDGAR Dataset exclusively [10] proposed for Ecuador to disaggregate the daily on-road traffic emissions. However, the modeled air quality levels notably worsened, both in magnitude and in the hourly distribution, compared to the results presented in this contribution. Therefore, a dedicated study is required to improve the estimation of hourly diesel-vehicle traffic, the most significant source of NO_x and PM_2.5 in the urban area of Cuenca, thereby improving the correlation coefficient of the corresponding modeled concentrations.

The RTV (Revisión Técnica Vehicular) is a technical control for the exhaust emissions and mechanical conditions of vehicles, which has been applied in Cuenca since 2008. According to national regulations, it is a prerequisite for the legal use of a vehicle throughout the year. Currently, the RTV controls hydrocarbon and carbon monoxide exhaust emissions from gasoline vehicles and opacity, a measure of how much light is blocked by particulate matter, from diesel vehicles.

The database generated by the RTV control is a valuable resource, which has been used to characterize the composition of the city’s vehicle park. This dataset includes records of the distance traveled per vehicle each year, which could enable a more accurate estimation of the on-road activity data.

There is a lack of available atmospheric records for the Andean Region of Ecuador [51,52]. Notably, the promotion of vertical sounding of meteorological and air quality species will enhance the characterization of parameters, such as the planetary boundary layer height and the vertical abundance of species, thereby providing valuable information for improving the boundary conditions used for modeling purposes and allowing for a more complete assessment of the modeling performance through the inclusion of vertical comparisons.

The EDGAR Dataset is one of the most widely used global emission resources. Georgiou et al. (2022) [7] reported the performance of an atmospheric forecasting system applied over the Eastern Mediterranean for both meteorological and air quality parameters based on the emission data from the EDGAR Dataset and WRF-Chem. Yarragunta et al. (2025) [53] utilized the EDGAR Dataset as input to WRF-Chem for atmospheric modeling, similarly assessing the performance of meteorological and air quality quantities. One advantage of the EDGAR Dataset is its consistent use of the same approach, similar to the IPCC method. However, this approach may overlook country-specific details that could enhance emission estimates and their spatial and temporal distribution [8]. The EDGAR emissions do not account for the effect of elevation on the decrease in atmospheric O_2, which affects the magnitude of emissions.

Rainfall is one of the most challenging components in atmospheric modeling, especially for complex regions as the Andean region. In this contribution, we assessed the performance of rainfall modeling based on daily intensities by evaluating the model’s ability to identify days with and without rain. A more comprehensive future assessment could compare computed precipitation with records by precipitation range based on hourly intensities. For this parameter, the modeling performance was poor or fair and needs to be particularly improved.

To assess the quality of the EI 2021, we used version 3.2 of the WRF-Chem model, which had been previously tested in Cuenca to define a recommended configuration of options that we used for this study. Apart from rainfall, we identified other variables requiring better modeling performance. Global solar radiation was overestimated, which partially explains the overestimation of O₃ levels in the afternoon. In the future, new versions and configurations of the WRF-Chem model can be tested to identify a set of schemes that improve the modeling of cumulus processes, thereby enhancing the representation of cloud coverage, rainfall, and surface solar radiation. Similarly, the potential benefits of indirect effects between aerosols and meteorological variables require dedicated assessments to improve atmospheric modeling in the Equatorial Andean region.

As we used the EI 2021 data within the WRF-Chem 3.2 framework, the modeling performance depends on the quality and limitations of both the EI 2021 to describe the emissions and the WRF-Chem 3.2 in modeling atmospheric variables. For air quality variables, emission data are a key input that capture the complex behavior of the atmosphere (e.g., wind speed, wind direction, planetary boundary layer height, stability dynamics).

5. Conclusions

The results indicated that most of the assessed meteorological and air quality variables were modeled acceptably across the two numerical sets (NE 1 and NE 2), suggesting that the EI 2021 emission inventory is a reasonable estimation of real emissions.

Our numerical experiments indicate that the modeling of air quality variables is sensitive to boundary conditions and to the hourly disaggregation of daily emissions. Although a complete sensitivity analysis is beyond the scope of this contribution, these two components warrant dedicated future studies.

Although October 2021 was selected to test the quality of the EI 2021, in the future, other months or even the whole annual period should be considered to account for the variability in emissions and meteorological conditions.

On the other hand, although the uncertainty of the EI 2021, it is a reasonable reference for the contributions of the most important sources of primary pollutants and GHGs in Cuenca. The EI 2021’s results can be used to explore strategies for improving air quality, such as changes in the composition of vehicle parks, the promotion of electric vehicles, improvements to fuel quality, and the use of exhaust filters in diesel vehicles. Additionally, the EI 2021 could serve as a reference for environmental impact assessments of policies, programs, and projects that affect air quality and to evaluate the performance of a preliminary atmospheric forecasting system.

To improve the atmospheric modeling, we suggest the following future research directions:

Since on-road traffic is one of the most important sources of both primary pollutants and GHGs, it is essential to maintain and update a map of daily traffic intensities on avenues, roads, and main streets. If possible, data should be collected to characterize the types and numbers of vehicles on the avenues and roads with the highest levels of traffic. For this purpose, camera-based traffic data collection should be implemented to provide additional information in order to improve the accuracy of hourly emissions estimation. Another option is to use data on the dynamic behavior of passengers’ cell phones as a proxy to infer the distribution of daily and hourly on-road traffic.

Additionally, it is necessary for each vehicle type to register the distance traveled per unit of consumed fuel. Apart from the information provided by the vehicle makers, we recommend conducting dedicated surveys of vehicle owners to obtain consistent information on the conditions in Cuenca.

Due to both short-term and long-term exceedances of the WHO guidelines [24], determining local PM_2.5 emission factors is a priority, particularly for diesel vehicles and fixed sources that use fossil fuels and biomass. The RTV should also include the control of the exhaust emissions of PM_2.5 and NO_x from diesel cars. In fact, the RTV database of exhaust emission measurements (e.g., refs. [54,55,56]) should be used to deduce local emission factors to complement the results from the in-route use of on-board measurement devices (e.g., refs. [57,58,59]). In particular, the viability of deducing PM_2.5 emission factors from opacity records deserves dedicated research, as the literature has reported conflicting correlation results (e.g., refs. [60,61]).

In addition, it is necessary to promote the definition of local emissions factors for NO_x and VOC as precursors of O₃. In particular, it is necessary to characterize local VOC emission factors from gasoline vehicles, vegetation, service stations, and solvents.

Regarding VOC emissions from vegetation, there are no known national or local emission factors. Additionally, there is limited information on the VOC emission capacity of some native species. Therefore, determining their emission factors [62,63] is a priority. Vegetation has high emission uncertainty. In the future, it will be necessary to characterize the spatial configuration of species and foliar biomass, as well as their dynamics throughout the year.

For future emission inventories, as a best practice, additional components can be included, such as quantification of black and elemental carbon [24].

The EI 2021 provides a coherent spatial distribution of emissions (1 km resolution), as indicated by the modeling results and performance. Apart from differences in emissions for some primary pollutants, the comparison of the geographical distribution with the EDGAR Dataset suggests that the latter would require revision to achieve performance comparable to EI 2021. For this purpose, the results from the EDGAR Dataset should be revisited to account for the effects of elevation on the magnitude and spatial distribution of primary pollutant emissions. In addition, because the EDGAR Database includes only anthropogenic sources, a vegetation VOC emission inventory is required as a relevant component for O₃ modeling.

The EDGAR results could be used to calculate emissions at a high spatial resolution (1 km), potentially using proxy variables, for atmospheric modeling of cities in the equatorial Andes.

As we use EI 2021 data within the WRF-Chem 3.2 framework, modeling performance depends on both the quality and limitations of EI 2021 and the WRF-Chem 3.2 model. Emissions are the most relevant input for modeling air pollutant dynamics, and the model’s ability to describe the atmosphere modulates those dynamics. Therefore, good modeling performance is not solely attributable to inventory quality, and poor modeling performance can result not only from poor inventory but also from the model’s inability to describe atmospheric behavior. This pioneering research in the equatorial Andes area aims precisely to promote modeling with the “one atmosphere” approach, with promising results that, of course, can be improved.

Supplementary material is available from [64].