Next Article in Journal
Event-Based Tracking of Spatiotemporally Contiguous PM2.5 Pollution Events in China
Previous Article in Journal
Weather-Regime-Based Heatwave Risk Typing and Urban Climate Resilience Assessment in New Delhi (1997–2016)
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Atmosphere
Volume 16
Issue 10
10.3390/atmos16101181
atmosphere-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Retrieval of Atmospheric Visibility and Its Driving Factors in Shanghai, China

by
Xiaowen Gui
1,2,
Jing Ren
1,
Guoyin Wang
1,2,
Yuying Wang
3,
Miao Zhang
1 and
Xiaoyan Wang
1,2,*
1
Department of Atmospheric and Oceanic Sciences and Institute of Atmospheric Sciences, Fudan University, Shanghai 200438, China
2
Shanghai Key Laboratory of Ocean-Land-Atmosphere Boundary Dynamics and Climate Change, Shanghai 200438, China
3
School of Atmospheric Physics, Nanjing University of Science and Technology, Nanjing 210094, China
*
Author to whom correspondence should be addressed.
Atmosphere 2025, 16(10), 1181; https://doi.org/10.3390/atmos16101181
Submission received: 11 September 2025 / Revised: 8 October 2025 / Accepted: 11 October 2025 / Published: 14 October 2025
(This article belongs to the Section Air Quality)
Browse Figures
Review Reports Versions Notes

Abstract

The combined effects of meteorological factors and aerosol chemical compositions on atmospheric visibility in Shanghai were investigated in this study based on the observed hourly dataset during 2022–2024. Correlation analysis and random forest modeling are employed to quantify the relative contributions of these factors. The results reveal significant negative correlations between visibility and both PM2.5 concentration and relative humidity, with partial correlation coefficient of −0.62 and −0.61. Nitrate, ammonium, and other aerosol components substantially modulate these relationships. The random forest model explains 83% of the variance when only meteorological variables are considered, increasing to 93% with the inclusion of aerosol chemical composition. Under 30 km high-visibility conditions, PM2.5 is the dominant predictor (39%) of atmospheric visibility variation, followed by relative humidity (35%). In contrast, during low-visibility conditions (lower than 7.5 km), relative humidity becomes the primary contributor (30%), the influence of PM2.5 weakens (18%), and aerosol chemical components account for a larger share (30%). These findings provide important insights into the mechanisms governing visibility variability under different environmental conditions.

1. Introduction

Atmospheric visibility refers to the maximum distance at which an object can be clearly identified by the human eye in a given direction, serving as an indicator of the transparency of the atmosphere. Visibility degradation is a direct manifestation of deteriorating air quality and is closely linked to public health. Prolonged exposure to low-visibility environments has been associated with increased risks of respiratory and cardiovascular diseases. In addition, poor visibility conditions frequently lead to disruptions such as flight delays, suspension of maritime operations, and highway traffic accidents. With the rapid development of the low-altitude economy, visibility has become a critical constraint on the safety and efficiency of low-level flight operations, and its variability may hinder the large-scale deployment of related industries. In recent years, visibility impairment caused by air pollution has become increasingly prominent, exerting a range of adverse effects on both society and the economy. Therefore, a comprehensive investigation into the influencing factors, underlying mechanisms, and mitigation strategies of visibility variation is of great significance for enhancing environmental governance, ensuring transportation safety, and supporting the development of emerging economic sectors.
Aerosol particles such as PM2.5 affect the transmission of solar radiation through scattering and absorption, leading to significant changes in both its intensity and propagation direction, thereby reducing atmospheric visibility [1]. High concentrations of PM2.5 directly scatter and absorb sunlight, reducing visibility [2]. Due to variations in particle size distribution, mixing state, and chemical composition, the relationship between PM2.5 concentration and visibility is inherently complex and nonlinear [3,4]. This complexity arises because larger and more light-scattering particles (e.g., sulfates, nitrates) disproportionately reduce visibility. Internally mixed particles with absorbing components (e.g., black carbon) exhibit compounding optical effects [5]. In addition, meteorological conditions influence visibility by affecting the formation of secondary aerosols, regional transport, and local accumulation of pollutants, all of which impact the physicochemical properties of local aerosols [6]. For example, the secondary formation of aerosols is highly sensitive to ambient temperature, humidity, and solar radiation. High humidity can promote new particle formation and gas-to-particle conversion [7,8]. Strong winds facilitate the horizontal transport of pollutants, whereas stagnant weather conditions and vertical temperature inversions contribute to the local accumulation of pollutants, exacerbating particulate pollution and thereby reducing visibility [9,10]. Hygroscopic aerosols such as sulfates and nitrates enhance aerosol growth by taking up water vapor, causing the particles to swell in size and form hygroscopic haze droplets, which reduces visibility even when the total PM mass remains constant [11,12].
Existing studies have shown distinct functional relationships between PM2.5 and visibility in different regions. In the Beijing-Tianjin-Hebei and Pearl River Delta regions, the relationship generally follows a power-law function [13,14], while in cities such as Xi’an and Mt. Tai, it tends to follow an exponential function [15,16]. According to Wang et al. [17], visibility decreases linearly with increasing PM2.5 concentration when concentrations are below 50 μg/m3, but it exhibits an exponential decline beyond this threshold. This quantitative relationship is further influenced by ambient humidity, as the hygroscopic growth of particles enhances light extinction under high relative humidity conditions [16,18,19]. Moreover, variations in wind speed, wind direction, and precipitation intensity may result in different functional forms of the PM2.5-visibility relationship even within the same region [20]. Meteorological variables and aerosol concentrations jointly affect atmospheric visibility, and their interactions are often nonlinear due to the inherent correlations between them. Either factor may dominate under different environmental conditions, adding substantial complexity to the mechanisms governing visibility variations. This complexity poses a significant challenge for accurate visibility forecasting in real-world atmospheric environments.
The IMPROVE (Interagency Monitoring of Protected Visual Environments) algorithm is a widely adopted method for estimating atmospheric visibility [21,22]. It calculates the atmospheric light extinction coefficient by quantifying the mass concentrations of key aerosol components—such as sulfates, nitrates, and organic carbon—and incorporating their hygroscopic growth effects, and subsequently estimates visibility using the Koschmieder equation [20,23,24]. This method is grounded in clear physical principles and has strong theoretical foundations, making it widely applied in extinction coefficient reconstruction studies in China [25,26,27,28,29]. However, the IMPROVE algorithm tends to significantly underestimate visibility at high levels and exhibits increasing estimation errors with rising particulate concentrations. Under conditions of elevated relative humidity, the discrepancy between reconstructed and observed extinction coefficients becomes more pronounced. Moreover, the method relies on detailed chemical speciation data, which are not routinely measured in conventional atmospheric monitoring networks, thereby limiting its applicability in regions with sparse or incomplete observational datasets [30].
Compared with the limitations of the IMPROVE method, machine learning techniques offer greater flexibility in integrating multi-source datasets and automatically learning the nonlinear characteristics of meteorological conditions, multi-pollutant interactions, and boundary layer dynamics. These models can adaptively adjust parameters through training, thereby improving predictive accuracy. Consequently, machine learning algorithms have been widely applied in recent years to visibility retrieval and forecasting, enabling seamless visibility prediction based on meteorological variables. Traditional machine learning models such as support vector machines (SVM), decision trees, random forests, and XGBoost, as well as deep learning approaches including long short-term memory (LSTM) networks, artificial neural networks, and convolutional neural networks, have demonstrated strong performance in visibility-related applications. Ortega et al. compared five deep learning models using conventional meteorological variables to simulate visibility over Florida and found that the performance of the LSTM model steadily improved as the training data size increased [31]. Kim et al. evaluated visibility forecasts produced by tree-based machine learning models and a numerical assimilation forecasting system, concluding that the XGBoost algorithm yielded significantly lower prediction errors than the numerical forecast products [32]. Yu et al. developed a visibility prediction model based on boosting algorithms by integrating air pollutant monitoring data, meteorological variables, and MODIS aerosol optical depth over Shanghai, showing that the model achieved notably higher accuracy under low-visibility conditions compared to the operational atmospheric environment modeling system in East China [33]. Zhang et al. proposed a real-time visibility retrieval framework using a stacked ensemble learning model, enabling continuous 24 h high-resolution monitoring of ground-level visibility across China [34].
In summary, atmospheric visibility is influenced by multiple factors, primarily including PM2.5 concentration, thermodynamic properties of the atmosphere, and the chemical composition of particulate matter. As China’s economic hub and one of its national central cities, Shanghai is also among the most populous cities globally, where changes in visibility exert significant impacts on public health, transportation, and emergency management. Although air quality in China has improved markedly since the implementation of the “Atmospheric Ten Measures” policy, low-visibility events remain frequent. The dominant drivers of visibility variation are still not fully understood, posing challenges for both visibility forecasting and long-term visibility improvement.
While most existing studies have focused on visibility prediction using machine learning models, this study aims to develop a visibility retrieval model by integrating air pollutant concentrations, routine meteorological variables, and aerosol chemical composition observations. The objective is to reveal the mechanisms underlying visibility variation driven by the synergistic effects of multiple factors and to quantify the relative contributions of each. The findings are expected to provide a scientific basis for targeted air quality management, improved forecasting accuracy, and enhanced emergency response capabilities.

2. Data and Methods

2.1. Data

The Fudan University Meteorological Observation Station, established in 2019, is a facility integrating teaching and research. Its construction complies with the regulations of the World Meteorological Organization and is equipped with standard instruments for conventional meteorological variables, atmospheric composition, and boundary layer turbulence structures. All observational data from the station are uniformly collected by the DZZ4 data acquisition system into the Ground-based Integrated Observation Business Software (Ver 3.0.2.615) of the China Meteorological Administration’s Meteorological Observation Center. This study employed hourly observational data from the observation station (as shown in Figure 1), covering the period from 1 January 2022 to 30 September 2024. The station is located at 121.51° E, 31.34° N with an elevation of 4 m. The station features a grassland surface type, with urban land cover located approximately 1 km away. The hourly meteorological data used in this study include visibility, relative humidity, and precipitation, which were obtained from instruments produced by Jiangsu Radio Scientific Institute Co., Ltd. (Nanjing, China), with the DNQ1 visibility sensor [35,36], DHC2 humidity sensor [37], and DSG1 present weather detector, respectively. The DNQ1 typically employs the forward scattering method to measure visibility, converting the extinction coefficient into the meteorological optical range, which is commonly referred to as the visibility value. The DHC2 uses a capacitive humidity sensor, in which a specially designed polymer thin-film capacitor changes its dielectric constant in response to the water vapor content in the surrounding air. The observation site is equipped with two parallel precipitation measurement systems: a tipping-bucket rain gauge and a siphon rain gauge. A daily precipitation of 1 mm is defined as effective precipitation. Precipitation is considered to have occurred if either rain gauge records effective precipitation. When precipitation is detected, if the relative humidity is below 75%, the data quality is deemed problematic, and all data from that time period are excluded. All instruments at the automatic weather station undergo regular annual calibration.
From November 2023 to May 2024, an intensive observation campaign on the physicochemical properties of atmospheric aerosols was conducted at the site. The aerosol chemical composition observation instruments were installed in a container within the observation field, co-located with the meteorological instruments in a 25 × 25 m outdoor observation area. During this period, hourly data on aerosol chemical composition were collected. The aerosol chemical speciation monitor (ACSM, Q-ACSM), equipped with a PM2.5 lens system, a capture vaporizer, and a quadrupole mass spectrometer was used to measure mass concentrations of non-refractory aerosol chemical species in PM2.5, including organics (Org), nitrate (NO3−), sulfate (SO42−), ammonium (NH4+), and chloride (Chl) [38]. In addition, black carbon (BC) mass concentrations were measured using a seven-wavelength aethalometer (AE-33, Magee Scientific), equipped with a PM2.5 cyclone inlet [39,40]. The AE-33 calculates BC mass concentration based on the attenuation of light intensity caused by the absorption of light by black carbon particles deposited on the filter as air passes through it. Before the enhanced observations, the ACSM and AE-33 had been calibrated. The chemical component data (measured at 15 min intervals) and black carbon data (measured at 5 min intervals) were averaged to hourly averages to match the temporal resolution of the meteorological observation data. To characterize ambient air pollution around the observation site, hourly PM2.5 concentration data from 1 January 2022 to 30 September 2024, were retrieved from two nearby national air quality monitoring stations located in Yangpu and Hongkou districts (as shown in Figure 1). The average PM2.5 concentration from the two stations was used to represent pollutant levels in the vicinity of the Fudan University Jiangwan Campus.

2.2. Methods

The random forest regression algorithm is an ensemble learning method that constructs multiple decision trees to perform classification or prediction [41,42]. It is capable of handling high dimensional features and is robust to multicollinearity, while also providing automatic evaluation of feature importance and exhibiting strong resistance to outliers and noise [43,44]. The primary reason for selecting this model in our study is its ability to effectively capture complex nonlinear relationships among variables without requiring predefined functional forms—a capability that has been validated through preliminary scatter plots and locally weighted regression (LOESS) fitting. To assess the influence of aerosol chemical components on atmospheric visibility, two prediction models were constructed based on the random forest algorithm, with and without the inclusion of chemical composition variables. To better capture the periodic and nonlinear temporal patterns in visibility, two key time-related features, day of year (DOY) and hour, were extracted. DOY reflects seasonal variability, while hour (ranging from 0 to 23) captures diurnal cycles. The dataset was randomly split into training and testing sets at a ratio of 8:2 using the train_test_split function, with a fixed random seed of 42 to ensure reproducibility. The model was further subjected to 5-fold cross-validation for internal validation. Additionally, to ensure the reliability of statistical inferences, an in-depth analysis of the model residuals was conducted. Specifically, the presence of heteroscedasticity was examined through the Breusch-Pagan test, and the presence of autocorrelation was verified via the Durbin-Watson test. The model was configured with 100 decision trees, and early stopping verification showed that the model’s performance stabilized beyond this number. Other key parameters were as follows: maximum tree depth was unlimited; the splitting criterion was minimization of mean squared error (MSE); and default values were used for other parameters, including min_samples_split = 2 and min_samples_leaf = 1. Model performance was evaluated using the coefficient of determination (R2) and root mean square error (RMSE).
To construct a high-quality analytical dataset, systematic preprocessing was applied to address missing and abnormal values in both the meteorological and aerosol chemical composition datasets. This included data cleaning, time series continuity enhancement, and refined feature processing strategies, whereby different treatments were applied to different types of variables. Since the visibility sensor used in this study has an upper detection limit of 30 km, whereas actual atmospheric visibility may exceed this threshold. A bilinear time series interpolation method was adopted to fill in missing values. This interpolation technique estimates intermediate visibility values by applying bidirectional linear weighting between the two nearest known time points, effectively preserving and extending the linear trend of visibility variation. In addition, a dynamic enhancement algorithm was developed to process visibility outliers. This algorithm adjusts the correction strength based on data trends: for rising trends, a larger multiplier is combined with a dynamic adjustment coefficient based on the difference between extreme and adjacent normal values; for falling or stable trends, a smaller multiplier is used in combination. Using this strategy, visibility values capped at 30 km in the raw observations were interpolated and dynamically adjusted up to 50 km to support downstream random forest modeling.
To evaluate the model’s sensitivity to input variables and verify its robustness, a comprehensive sensitivity analysis was conducted, which encompassed feature importance assessment, feature perturbation analysis, as well as evaluations of multicollinearity and outlier robustness. Specifically, the intrinsic feature importance evaluation method based on impurity reduction in Random Forest was employed to quantify the contribution of each input variable to visibility prediction, calculating the mean decrease in impurity (e.g., Gini impurity) caused by each feature across all decision trees and then normalizing the results to derive the relative importance of each feature; the sensitivity of individual features to model predictions was further validated by randomly shuffling the values of single features (Permutation Importance) and observing the corresponding changes in model performance (R2); Variance Inflation Factor (VIF) analysis was performed to detect potential multicollinearity among meteorological variables; moreover, the model’s resilience to outliers and data noise was jointly verified through model retraining after introducing controlled random noise to a defined proportion of the training set and preprocessing extreme values using the Median Absolute Deviation method during the data preparation stage.

3. Results

3.1. Temporal Variations in Visibility and Related Factors

Observations from 2022 to 2024 indicate that atmospheric visibility in Shanghai exhibits clear seasonal variation, with higher values in summer and lower values in winter. The average visibility during summer reached 27.79 km, compared to 21.74 km in winter. As shown in Table 1, the seasonal distribution of valid visibility samples reveals that 80% of summer observations reached or exceeded the instrument’s upper detection limit of 30 km, with an annual average exceedance rate of 62%. The visibility data distribution exhibits significant left-skewness (negative skew) and leptokurtosis (peakedness), with a skewness value below zero and kurtosis above zero. The skewness reaches its maximum in summer (−0.34), likely due to frequent low-visibility events caused by higher relative humidity during this season. Given the overall high visibility levels in Shanghai and the substantial proportion of values reaching the instrument’s ceiling, this subset of data was analyzed separately in subsequent sections. PM2.5 concentrations were lower during summer and autumn and increased significantly in winter and spring, with seasonal means of 20.41 μg/m3 and 39.93 μg/m3 for summer and winter, respectively. Relative humidity showed the opposite trend, being higher in summer and autumn and lower in winter and spring.
Figure 2 illustrates the diurnal variation patterns of visibility and its two primary influencing factors, i.e., PM2.5 concentration and relative humidity. Visibility exhibits a unimodal diurnal cycle, with the highest values typically occurring between 10:00 and 15:00, reaching up to 24.26 km, and the lowest values appearing between 04:00 and 06:00, dropping to 13.08 km. Overall, visibility is higher during the daytime and lower at night. Relative humidity follows a bimodal pattern, reaching its minimum in the early afternoon (14:00–15:00) and increasing significantly during nighttime hours. PM2.5 concentration displays a unimodal distribution, with the lowest concentrations observed in the early morning (04:00–06:00) and a peak around 09:00–10:00, followed by a gradual decline throughout the afternoon. In general, the highest visibility coincides with the lowest PM2.5 concentrations in the afternoon, indicating a negative diurnal correlation. Visibility and relative humidity exhibit opposite peak-trough timing, also suggesting a strong negative correlation. Figure 3 presents the Pearson correlation coefficients and partial correlation between visibility and PM2.5 concentration, as well as visibility and relative humidity, after excluding data capped at the instrument’s maximum detection threshold. Both variables show negative correlations with visibility, with stronger correlations observed during nighttime hours. The correlation coefficients between visibility and PM2.5, as well as relative humidity, showed significant improvement after controlling for the other variable, from −0.41 to −0.62 and −0.37 to −0.61, indicating a strong mutual influence between relative humidity and PM2.5. The partial correlation coefficient of visibility-PM2.5 and visibility-RH exhibited consistent diurnal variation patterns, reaching their maximum absolute values during 09:00–12:00 (i.e., −0.70 and −0.65) and minimum values during 17:00–19:00 (i.e., −0.53 and −0.53).
As shown in Figure 3, the correlation between visibility and both PM2.5 concentration and relative humidity exhibits a pronounced diurnal pattern. Due to the hygroscopic growth of particulate matter, the relationship between PM2.5 and relative humidity further increases the complexity of visibility variation. Figure 4 illustrates the modulation effect of relative humidity on the visibility–PM2.5 relationship. A general decreasing trend in visibility is observed with increasing PM2.5 concentration, consistent with the negative correlation discussed earlier. At a fixed PM2.5 concentration, visibility declines progressively as relative humidity increases. For example, during winter, when PM2.5 concentration is approximately 50 μg/m3, visibility can exceed 25 km under low-humidity conditions (RH < 50%), but drops to below 5 km when relative humidity exceeds 90%.
The hygroscopic growth properties of particulate matter are significantly influenced by its chemical composition. Figure 5 further analyzes the impact of PM2.5 chemical component variation on the visibility–PM2.5 and visibility–relative humidity relationships. Since the absolute concentrations of individual chemical species generally increase linearly with total PM2.5 concentration, the analysis of their influence on the visibility–PM2.5 relationship was conducted using the relative proportion of each component within PM2.5. In contrast, the visibility–humidity relationship was examined using the absolute concentrations of the chemical species. As shown in Figure 5, although organic aerosols (Org) account for the largest fraction of PM2.5, they exert minimal influence on the visibility–PM2.5 relationship. In comparison, the proportions of nitrate, sulfate, and black carbon have a pronounced effect. Under low PM2.5 concentration conditions (below 50 μg/m3), a decrease in nitrate fraction or an increase in black carbon fraction significantly reduces visibility, while a higher sulfate fraction is associated with improved visibility. Regarding the visibility–relative humidity relationship, increases in the concentrations of ammonium, nitrate, and black carbon lead to substantial visibility degradation at the same humidity levels.

3.2. Distributions of Relative Humidity and PM2.5 Concentration Under Extreme Visibility Conditions

Given that over 60% of the visibility observations in Shanghai reach the instrument’s upper threshold of 30 km, this section further investigates the distribution characteristics of relative humidity and PM2.5 concentration under conditions of maximum visibility (30 km) and low visibility (defined as <7.5 km based on haze criteria), in order to explore differences in the dominant factors influencing extreme visibility levels. Figure 6 presents the probability density distributions of relative humidity and PM2.5 concentration under both high- and low-visibility conditions. As shown, PM2.5 concentration exhibits a left-skewed distribution under high-visibility conditions, with all values below 75 μg/m3 and a sharp unimodal peak centered around 10–15 μg/m3. Under low-visibility conditions, the PM2.5 distribution also appears left-skewed but spans a broader range with significantly higher concentrations, generally between 20 and 80 μg/m3, and maxima exceeding 175 μg/m3. In contrast, the probability distribution of relative humidity demonstrates the opposite pattern. During high-visibility episodes, humidity exhibits a broad peak centered at moderately high levels, whereas under low-visibility conditions, the distribution is strongly right-skewed, with values concentrated in the high-humidity range and consistently exceeding 50%. These results suggest that high humidity is a key driver of visibility degradation in Shanghai, while exceptionally high visibility is primarily associated with low PM2.5 concentrations.
To address the potential multicollinearity between relative humidity and PM2.5 in their influence on visibility, Table 2 reports the ridge regression coefficients derived from models incorporating both variables. When all samples are considered, the coefficient for PM2.5 exceeds that of relative humidity, indicating that particulate matter concentration is the primary driver of visibility variation. However, stratified analysis reveals contrasting patterns: under low-visibility conditions, the coefficient for relative humidity substantially surpasses that of PM2.5, highlighting the dominant role of atmospheric moisture in governing visibility degradation. Conversely, under high-visibility conditions, the coefficient associated with PM2.5 is greater, underscoring the leading contribution of particulate matter to visibility variability in relatively clean environments. Figure 7 shows the scatterplot of PM2.5 versus relative humidity under high-visibility (30 km) conditions, along with the corresponding mass fractions of major aerosol chemical components. Under low-PM2.5 and low-humidity conditions, ammonium mass fractions are relatively low; under low-PM2.5 but high-humidity conditions, the nitrate fraction decreases while the sulfate fraction increases. Organic carbon, black carbon, and chloride exhibit minimal influence on high-visibility conditions.

3.3. Visibility Retrieval Based on Random Forest Regression

As demonstrated in the previous analyses, atmospheric visibility is closely related to PM2.5 concentration, relative humidity, and the chemical composition of particulate matter, with these relationships exhibiting pronounced diurnal and seasonal variations. Based on these findings, a visibility prediction model was developed using PM2.5, relative humidity, precipitation, aerosol chemical components, date, and hour as input features, with visibility as the target variable. The model was evaluated for predictive performance and subjected to feature importance analysis. Given that chemical composition data are not part of routine atmospheric monitoring and are relatively difficult to obtain, yet have been shown to significantly influence visibility, two separate visibility retrieval models were constructed: one including aerosol chemical components and one excluding them. This approach allows for a quantitative assessment of the relative contribution of chemical composition to visibility variation.
Due to differences in the observational periods of chemical composition data and conventional meteorological data such as visibility, two random forest models were constructed using datasets from distinct time spans. Model 1 was developed using hourly data from 1 January 2022 to 30 September 2024, including PM2.5 concentration, relative humidity, precipitation, date, and time information. Model 2 was constructed using hourly data from November 2023 to May 2024, which included PM2.5 chemical composition and mass concentration, along with relative humidity and other meteorological parameters. As shown in Figure 8a, Model 1 achieved an R2 of 0.83 and a residual error of 2369 m, indicating that a model incorporating only particle mass concentration, precipitation, relative humidity, and temporal features (date and hour) could explain 83% of the variance in visibility. Figure 9a presents the global feature importance derived from Model 1. PM2.5 was the most influential factor, contributing 40% of total importance and exhibiting a significant negative correlation with visibility. Relative humidity accounted for 35% of the importance, ranking second. High humidity likely reduces visibility indirectly by promoting aerosol hygroscopic growth or fog formation. Figure 9b further distinguishes feature importance under high- and low-visibility conditions in Model 1. Under high-visibility conditions, PM2.5 emerged as the primary influencing factor (38%), followed by relative humidity (31%). In contrast, during low-visibility conditions, relative humidity became the dominant factor (32%), while the importance of PM2.5 declined to 31%, and the importance of temporal features increased. These results are consistent with earlier qualitative analyses, reinforcing that the dominant visibility drivers shift with environmental conditions.
Building on the previously developed random forest visibility retrieval framework, as shown in Figure 8b, Model 2 was constructed by incorporating aerosol chemical components (Org, NO3−, SO42−, NH4+, Chl, and BC) as additional predictors. Compared with Model 1, the inclusion of chemical composition increased the model’s explanatory power from 83% to 93%, indicating that these added features provided more comprehensive coverage of the key factors influencing visibility. In particular, the addition of light-absorbing species (e.g., black carbon) and hygroscopic components (e.g., sulfate and ammonium) significantly enhanced the model’s predictive accuracy, better aligning with the underlying atmospheric physical processes. As shown in Figure 9c, relative humidity emerged as the most important predictor in Model 2, contributing 41% of the total importance and surpassing PM2.5 (39%). This shift reflects the dual role of relative humidity: directly affecting visibility by modulating PM2.5 concentration and indirectly influencing visibility through the hygroscopic behavior of aerosol chemical components, particularly those capable of water uptake. Among the chemical species, ammonium exerted the greatest influence on visibility variation. Figure 9d further presents feature importance under different visibility regimes. Compared with high-visibility conditions, the importance of PM2.5 sharply decreased under low-visibility conditions (from 39% to 18%), suggesting that the extinction effect of particles is more prominent in clean air. Consistent with Model 1, relative humidity remained the dominant factor in low-visibility cases, while PM2.5 became secondary. However, the overall contribution of chemical components increased significantly under low-visibility conditions. Black carbon, sulfate, and nitrate aerosols had a pronounced impact on visibility reduction, while the relative importance of total PM2.5 mass concentration was diminished.
To verify heteroscedasticity and autocorrelation in the two models, Table 3 presents their Durbin-Watson test statistics: Model 1 and Model 2 yielded values of 2.3 and 2.1, respectively, both falling between 2 and 2.5, indicating no significant autocorrelation in either model. Meanwhile, the Durbin-Watson statistics for heteroscedasticity evaluation were 0.16 for Model 1 and 0.13 for Model 2—both exceeding 0.05, which suggests the presence of heteroscedasticity in both models. However, the Random Forest model is inherently less sensitive to heteroscedasticity. The presence of multicollinearity among selected meteorological variables was assessed through Variance Inflation Factor (VIF) calculations. The results in Table 4 indicate that some meteorological features in Model 2 (i.e., NO3 and NH4) exhibit VIF values exceeding 10, suggesting moderate multicollinearity. However, the Bootstrap sampling mechanism inherent in the Random Forest model can effectively mitigate the impact of such multicollinearity on model stability.

4. Conclusions and Discussion

This study focused on Shanghai to investigate the relationships between atmospheric visibility and multiple influencing factors, including PM2.5 concentration, relative humidity, and aerosol chemical composition. A high-resolution, hourly visibility retrieval model was developed using the random forest algorithm, aiming to reveal the underlying mechanisms of visibility variation driven by the synergistic effects of meteorological conditions and aerosol composition and to quantify the relative contributions of different influencing factors.
The results show that atmospheric visibility exhibits a significant negative correlation with both PM2.5 concentration and relative humidity, with stronger negative correlations observed during nighttime hours. This diurnal asymmetry is largely attributed to the enhanced hygroscopic growth of aerosols during the day, which increases the complexity and nonlinearity of the relationship between visibility and either PM2.5 concentration or relative humidity alone. The proportions and absolute concentrations of sulfate, nitrate, ammonium, and black carbon aerosols were found to play a key role in modulating the visibility–PM2.5 and visibility–humidity relationships.
Under high-visibility conditions, PM2.5 concentrations exhibited a strongly left-skewed distribution, while relative humidity followed a broad-peak pattern. In contrast, during low-visibility conditions, PM2.5 concentrations were broadly distributed, and relative humidity showed a right-skewed distribution, indicating that visibility under clean-air conditions is primarily controlled by aerosol concentration, while visibility degradation under polluted or humid conditions is mainly driven by elevated relative humidity. The random forest model showed an R2 of 0.83 when only PM2.5, relative humidity, time, and date were included, which improved to 0.93 upon incorporating aerosol chemical composition. Feature importance analysis further revealed that PM2.5 was the dominant factor under high-visibility conditions (39%), followed by relative humidity (35%). However, under low-visibility conditions, relative humidity became the primary factor (30%), while the importance of PM2.5 decreased to 18%. The roles of black carbon, sulfate, nitrate, and ammonium became more prominent, with a combined contribution of approximately 30%.
The results in this study are consistent with previous studies, in which visibility degradation was found to be jointly controlled by PM2.5 concentration and relative humidity [45]. Stronger negative correlations during nighttime were attributed to enhanced aerosol hygroscopic growth, as similarly reported in Beijing and Korea [17,46,47]. The greater influence of sulfate, nitrate, ammonium, and black carbon under polluted or humid conditions was also observed, in line with findings that secondary inorganic aerosols and absorbing species contribute to nonlinear visibility loss [48,49]. Importantly, the shift in dominance from PM2.5 under clean conditions to humidity and aerosol composition under haze was quantified, and a predictive framework was provided through the random forest model.
Uncertainty remains in both data and methodology. Although the high-resolution visibility retrievals and random forest model achieved high R2 values, prediction errors persisted under extreme polluted or clean-air conditions, suggesting limited robustness in boundary cases. The reliance on ground-based optical and meteorological measurements also introduces potential biases due to calibration accuracy and instrument stability, which may affect the reliability of the outcomes. While the results confirm the significant roles of PM2.5 and relative humidity in visibility degradation, the numerical findings should be interpreted with caution. Additional independent validation and more comprehensive observations are required to strengthen confidence in the conclusions.
The strong influence of sulfate, nitrate, ammonium, and black carbon on visibility has been widely reported, yet the relative contributions of PM2.5 and humidity are likely shaped by Shanghai’s specific emission mix, meteorological conditions, and geographic setting. This limits the generalizability of the conclusions to other regions. Future research should incorporate multi-city comparisons and multi-platform observations, including satellite and network data, to assess broader applicability. Moreover, the current study provides only a partial treatment of aerosol–radiation interactions and hygroscopic processes. Although this study is based solely on observational data from a single case in Shanghai, the diversified comprehensive observational data confirm that even without detailed chemical composition information, conventional meteorological observations (such as relative humidity and PM2.5 concentration) can effectively capture the main variability characteristics of visibility, providing important reference for subsequent studies. This study highlights the dominant role of hygroscopic growth and secondary aerosol components, though it lacks a mechanistic investigation into the coupled feedbacks between aerosol radiative forcing (e.g., black carbon heating altering boundary layer stability) and humidity dynamics. This simplification restricts a deeper physical understanding of the observed diurnal asymmetries and nonlinear visibility responses, highlighting the need for integrated observational-modeling approaches to unravel these complex processes. More detailed mechanistic studies combining observational and modeling approaches are necessary to improve physical understanding and enhance predictive capacity across diverse environments.

Author Contributions

X.G., G.W. and X.W. designed research. J.R., X.G., M.Z. and Y.W. performed the analyses and wrote this paper. All authors have read and agreed to the published version of the manuscript.

Funding

This research has been funded by the National Natural Science Foundation of China (grant nos. 42375183, 42288101, 42522504).

Data Availability Statement

Hourly PM2.5 concentration observations at Yangpu and Hongkou districts of Shanghai were obtained from the website of China General Environmental Monitoring Station of China (www.cnemc.cn). Further inquiries can be directed to the corresponding author.

Acknowledgments

We thank the support of MAP-AQ Asian Office, and FDU-IRDR-ICoE-RIG-WECEIPHE.

Conflicts of Interest

The authors declare that they have no conflict of interest.

References

  1. Xin, J.; Gong, C.; Liu, Z.; Cong, Z.; Gao, W.; Song, T.; Pan, Y.; Sun, Y.; Ji, D.; Wang, L. The observation-based relationships between PM2.5 and AOD over China. J. Geophys. Res. Atmos. 2016, 121, 10, 701–710, 716. [Google Scholar] [CrossRef]
  2. Hu, S.; Zhao, G.; Tan, T.; Li, C.; Zong, T.; Xu, N.; Zhu, W.; Hu, M. Current challenges of improving visibility due to increasing nitrate fraction in PM2.5 during the haze days in Beijing, China. Environ. Pollut. 2021, 290, 118032. [Google Scholar] [CrossRef]
  3. Covert, D.S.; Charlson, R.; Ahlquist, N. A study of the relationship of chemical composition and humidity to light scattering by aerosols. J. Appl. Meteorol. Climatol. 1972, 11, 968–976. [Google Scholar] [CrossRef]
  4. Lin, J.; van Donkelaar, A.; Xin, J.; Che, H.; Wang, Y. Clear-sky aerosol optical depth over East China estimated from visibility measurements and chemical transport modeling. Atmos. Environ. 2014, 95, 258–267. [Google Scholar] [CrossRef]
  5. Won, W.-S.; Oh, R.; Lee, W.; Ku, S.; Su, P.-C.; Yoon, Y.-J. Hygroscopic properties of particulate matter and effects of their interactions with weather on visibility. Sci. Rep. 2021, 11, 16401. [Google Scholar] [CrossRef] [PubMed]
  6. Kim, B.-Y.; Cha, J.W.; Chang, K.-H.; Lee, C. Estimation of the Visibility in Seoul, South Korea, Based on Particulate Matter and Weather Data, Using Machine-learning Algorithm. Aerosol Air Qual. Res. 2022, 22, 220125. [Google Scholar] [CrossRef]
  7. Wang, X.; Zhang, R. How did air pollution change during the COVID-19 outbreak in China? Bull. Am. Meteorol. Soc. 2020, 101, E1645–E1652. [Google Scholar] [CrossRef]
  8. Wang, X.; Dickinson, R.E.; Su, L.; Zhou, C.; Wang, K. PM2.5 pollution in China and how it has been exacerbated by terrain and meteorological conditions. Bull. Am. Meteorol. Soc. 2018, 99, 105–119. [Google Scholar] [CrossRef]
  9. Sun, Y.; Wang, X. Meteorological factor contributions to the seesaw concentration pattern between PM2.5 and O3 in Shanghai. Front. Environ. Sci. 2022, 10, 1015723. [Google Scholar] [CrossRef]
  10. Wang, X.; Zhang, R.; Tan, Y.; Yu, W. Dominant synoptic patterns associated with the decay process of PM2.5 pollution episodes around Beijing. Atmos. Chem. Phys. Discuss. 2020, 2020, 1–35. [Google Scholar]
  11. Zhang, X.; Ding, X.; Talifu, D.; Wang, X.; Abulizi, A.; Maihemuti, M.; Rekefu, S. Humidity and PM2.5 composition determine atmospheric light extinction in the arid region of northwest China. J. Environ. Sci. 2021, 100, 279–286. [Google Scholar] [CrossRef]
  12. Wang, X.; Wang, K.; Su, L. Contribution of atmospheric diffusion conditions to the recent improvement in air quality in China. Sci. Rep. 2016, 6, 36404. [Google Scholar] [CrossRef]
  13. Fu, X.; Wang, X.; Hu, Q.; Li, G.; Ding, X.; Zhang, Y.; He, Q.; Liu, T.; Zhang, Z.; Yu, Q. Changes in visibility with PM2.5 composition and relative humidity at a background site in the Pearl River Delta region. J. Environ. Sci. 2016, 40, 10–19. [Google Scholar] [CrossRef] [PubMed]
  14. Zhao, P.; Zhang, X.; Xu, X.; Zhao, X. Long-term visibility trends and characteristics in the region of Beijing, Tianjin, and Hebei, China. Atmos. Res. 2011, 101, 711–718. [Google Scholar] [CrossRef]
  15. Cao, J.-J.; Wang, Q.-Y.; Chow, J.C.; Watson, J.G.; Tie, X.-X.; Shen, Z.-X.; Wang, P.; An, Z.-S. Impacts of aerosol compositions on visibility impairment in Xi’an, China. Atmos. Environ. 2012, 59, 559–566. [Google Scholar] [CrossRef]
  16. Zhao, T.; Yang, L.; Yan, W.; Zhang, J.; Lu, W.; Yang, Y.; Chen, J.; Wang, W. Chemical characteristics of PM1/PM2.5 and influence on visual range at the summit of Mount Tai, North China. Sci. Total Environ. 2017, 575, 458–466. [Google Scholar] [CrossRef]
  17. Wang, X.; Zhang, R.; Yu, W. The effects of PM2.5 concentrations and relative humidity on atmospheric visibility in Beijing. J. Geophys. Res. Atmos. 2019, 124, 2235–2259. [Google Scholar] [CrossRef]
  18. Deng, X.; Tie, X.; Wu, D.; Zhou, X.; Bi, X.; Tan, H.; Li, F.; Jiang, C. Long-term trend of visibility and its characterizations in the Pearl River Delta (PRD) region, China. Atmos. Environ. 2008, 42, 1424–1435. [Google Scholar] [CrossRef]
  19. Yang, Y.; Liu, X.; Qu, Y.; An, J.; Jiang, R.; Zhang, Y.; Sun, Y.; Wu, Z.; Zhang, F.; Xu, W. Characteristics and formation mechanism of continuous hazes in China: A case study during the autumn of 2014 in the North China Plain. Atmos. Chem. Phys. 2015, 15, 8165–8178. [Google Scholar] [CrossRef]
  20. Wang, J.; Zhang, Y.-F.; Feng, Y.-C.; Zheng, X.-J.; Jiao, L.; Hong, S.-M.; Shen, J.-D.; Zhu, T.; Ding, J.; Zhang, Q. Characterization and source apportionment of aerosol light extinction with a coupled model of CMB-IMPROVE in Hangzhou, Yangtze River Delta of China. Atmos. Res. 2016, 178, 570–579. [Google Scholar] [CrossRef]
  21. Chen, D.; Zhao, Y.; Zhang, J.; Yu, H.; Yu, X. Characterization and source apportionment of aerosol light scattering in a typical polluted city in the Yangtze River Delta, China. Atmos. Chem. Phys. 2020, 20, 10193–10210. [Google Scholar] [CrossRef]
  22. Deng, J.; Zhang, Y.; Hong, Y.; Xu, L.; Chen, Y.; Du, W.; Chen, J. Optical properties of PM2.5 and the impacts of chemical compositions in the coastal city Xiamen in China. Sci. Total Environ. 2016, 557, 665–675. [Google Scholar] [CrossRef] [PubMed]
  23. Ma, Q.; Wu, Y.; Fu, S.; Zhang, D.; Han, Z.; Zhang, R. Pollution severity-dependent aerosol light scattering enhanced by inorganic species formation in Beijing haze. Sci. Total Environ. 2020, 719, 137545. [Google Scholar] [CrossRef]
  24. Wang, H.; Shi, G.; Tian, M.; Zhang, L.; Chen, Y.; Yang, F.; Cao, X. Aerosol optical properties and chemical composition apportionment in Sichuan Basin, China. Sci. Total Environ. 2017, 577, 245–257. [Google Scholar] [CrossRef] [PubMed]
  25. Guo, L.; Cui, Y.; He, Q.; Gao, W.; Pei, K.; Zhu, L.; Li, H.; Wang, X. Contributions of aerosol chemical composition and sources to light extinction during haze and non-haze days in Taiyuan, China. Atmos. Pollut. Res. 2021, 12, 101140. [Google Scholar] [CrossRef]
  26. Lan, Z.; Zhang, B.; Huang, X.; Zhu, Q.; Yuan, J.; Zeng, L.; Hu, M.; He, L. Source apportionment of PM2.5 light extinction in an urban atmosphere in China. J. Environ. Sci. 2018, 63, 277–284. [Google Scholar] [CrossRef]
  27. Liao, W.; Zhou, J.; Zhu, S.; Xiao, A.; Li, K.; Schauer, J.J. Characterization of aerosol chemical composition and the reconstruction of light extinction coefficients during winter in Wuhan, China. Chemosphere 2020, 241, 125033. [Google Scholar] [CrossRef] [PubMed]
  28. Xia, Y.; Tao, J.; Zhang, L.; Zhang, R.; Li, S.; Wu, Y.; Cao, J.; Wang, X.; Ma, Q.; Xiong, Z. Impact of size distributions of major chemical components in fine particles on light extinction in urban Guangzhou. Sci. Total Environ. 2017, 587, 240–247. [Google Scholar] [CrossRef]
  29. Yang, L.; Zhou, X.; Wang, Z.; Zhou, Y.; Cheng, S.; Xu, P.; Gao, X.; Nie, W.; Wang, X.; Wang, W. Airborne fine particulate pollution in Jinan, China: Concentrations, chemical compositions and influence on visibility impairment. Atmos. Environ. 2012, 55, 506–514. [Google Scholar] [CrossRef]
  30. Xin, J.; Wu, X.; Zhang, W.; Kong, L.; Ma, Y.; Ma, Y. A review on the extinction contribution of aerosol components in the United States and China. Chin. J. Atmos. Sci. 2024, 48, 273–286. [Google Scholar]
  31. Ortega, L.C.; Otero, L.D.; Solomon, M.; Otero, C.E.; Fabregas, A. Deep learning models for visibility forecasting using climatological data. Int. J. Forecast. 2023, 39, 992–1004. [Google Scholar] [CrossRef]
  32. Kim, B.-Y.; Belorid, M.; Cha, J.W. Short-term visibility prediction using tree-based machine learning algorithms and numerical weather prediction data. Weather Forecast. 2022, 37, 2263–2274. [Google Scholar] [CrossRef]
  33. Yu, Z.; Qu, Y.; Wang, Y.; Ma, J.; Cao, Y. Application of machine-learning-based fusion model in visibility forecast: A case study of Shanghai, China. Remote Sens. 2021, 13, 2096. [Google Scholar] [CrossRef]
  34. Zhang, X.; Gui, K.; Zeng, Z.; Fei, Y.; Li, L.; Zheng, Y.; Peng, Y.; Liu, Y.; Shang, N.; Zhao, H. Mapping the seamless hourly surface visibility in China: A real-time retrieval framework using a machine-learning-based stacked ensemble model. NPJ Clim. Atmos. Sci. 2024, 7, 68. [Google Scholar] [CrossRef]
  35. Yu, Y.; Ren, Z.; Meng, X. Reconstruction of daily haze data across China between 1961 and 2020. Int. J. Climatol. 2022, 42, 5629–5643. [Google Scholar] [CrossRef]
  36. Li, X.; Tang, G.; Li, L.; Quan, W.; Wang, Y.; Zhao, Z.; Liu, N.; Hong, Y.; Ma, Y. Multilevel air quality evolution in Shenyang: Impact of elevated point emission reduction. J. Environ. Sci. 2022, 113, 300–310. [Google Scholar] [CrossRef] [PubMed]
  37. Weiwei, L.; Xiaohua, L.; Zuoyang, T.; Xiaohua, X.; Yong, X. Fault Analysis and Maintenance of DZZ Series of Automatic Weather Stations. Meteorol. Environ. Res. 2018, 9, 35–37. [Google Scholar]
  38. Wang, Y.; Li, Z.; Wang, Q.; Jin, X.; Yan, P.; Cribb, M.; Li, Y.; Yuan, C.; Wu, H.; Wu, T. Enhancement of secondary aerosol formation by reduced anthropogenic emissions during Spring Festival 2019 and enlightenment for regional PM2.5 control in Beijing. Atmos. Chem. Phys. Discuss. 2020, 2020, 1–28. [Google Scholar]
  39. Dong, H.; Zou, Y.; Yan, M.; Sun, H.; Chen, J.; Yan, Y.; Zhu, C.; Hao, C.; Chen, Z. Epidemiological characteristics of RSV in pediatric inpatients with lower respiratory tract infections in Suzhou and their correlation with meteorology and atmospheric pollutants. BMC Infect. Dis. 2025, 25, 662. [Google Scholar] [CrossRef]
  40. Chen, W.; Cao, X.; Ran, H.; Chen, T.; Yang, B.; Zheng, X. Concentration and source allocation of black carbon by AE-33 model in urban area of Shenzhen, southern China. J. Environ. Health Sci. Eng. 2022, 20, 469–483. [Google Scholar] [CrossRef]
  41. Gong, H.; Wang, G.; Wang, X.; Kuang, Z.; Cheng, T. Trajectories of Terrestrial Vegetation Productivity and Its Driving Factors in China’s Drylands. Geophys. Res. Lett. 2024, 51, 10. [Google Scholar] [CrossRef]
  42. Ding, W.; Qie, X. Prediction of air pollutant concentrations via RANDOM forest regressor coupled with uncertainty analysis—A case study in Ningxia. Atmosphere 2022, 13, 960. [Google Scholar] [CrossRef]
  43. Gocheva-Ilieva, S.G.; Voynikova, D.S.; Stoimenova, M.P.; Ivanov, A.V.; Iliev, I.P. Regression trees modeling of time series for air pollution analysis and forecasting. Neural Comput. Appl. 2019, 31, 9023–9039. [Google Scholar] [CrossRef]
  44. Sun, H.; Gui, D.; Yan, B.; Liu, Y.; Liao, W.; Zhu, Y.; Lu, C.; Zhao, N. Assessing the potential of random forest method for estimating solar radiation using air pollution index. Energy Convers. Manag. 2016, 119, 121–129. [Google Scholar] [CrossRef]
  45. Cheng, Y.; Zheng, G.; Wei, C.; Mu, Q.; Zheng, B.; Wang, Z.; Gao, M.; Zhang, Q.; He, K.; Carmichael, G. Reactive nitrogen chemistry in aerosol water as a source of sulfate during haze events in China. Sci. Adv. 2016, 2, e1601530. [Google Scholar] [CrossRef] [PubMed]
  46. Joo, S.; Shin, J.; Tesche, M.; Dehkhoda, N.; Kim, T.; Noh, Y. Increased number concentrations of small particles explain perceived stagnation in air quality over Korea. Atmos. Chem. Phys. 2025, 25, 1023–1036. [Google Scholar] [CrossRef]
  47. Wen, W.; Tharaka, W.A.N.D.; Liu, L.; Ma, X.; Shen, L.; Deng, Z. Comparative analysis of the impact of PM2.5 components on visibility, extinction and oxidative potential in beijing, China. Air Qual. Atmos. Health 2025, 18, 2343–2356. [Google Scholar] [CrossRef]
  48. Shang, X.; Zhang, K.; Meng, F.; Wang, S.; Lee, M.; Suh, I.; Kim, D.; Jeon, K.; Park, H.; Wang, X. Characteristics and source apportionment of fine haze aerosol in Beijing during the winter of 2013. Atmos. Chem. Phys. 2018, 18, 2573–2584. [Google Scholar]
  49. Wu, C.; Wang, G.; Wang, J.; Li, J.; Ren, Y.; Zhang, L.; Cao, C.; Li, J.; Ge, S.; Xie, Y. Chemical characteristics of haze particles in Xi’an during Chinese Spring Festival: Impact of fireworks burning. J. Environ. Sci. 2018, 71, 179–187. [Google Scholar] [CrossRef]
Atmosphere 16 01181 g001
Figure 1. Spatial distribution of observational data sites. Blue markers represent the two PM2.5 monitoring sites, while the red five-pointed star indicates the location of Fudan University, which provided visibility, meteorological, and chemical composition observation data for this study.
Figure 1. Spatial distribution of observational data sites. Blue markers represent the two PM2.5 monitoring sites, while the red five-pointed star indicates the location of Fudan University, which provided visibility, meteorological, and chemical composition observation data for this study.
Atmosphere 16 01181 g001
Atmosphere 16 01181 g002
Figure 2. Diurnal Variations in Visibility, PM2.5, and Relative Humidity in Shanghai. Dashed lines represent data including visibility up to 30 km, solid lines represent data excluding maximum visibility. Hourly observed dataset from 1 January 2022 to 30 September 2024 were used here.
Figure 2. Diurnal Variations in Visibility, PM2.5, and Relative Humidity in Shanghai. Dashed lines represent data including visibility up to 30 km, solid lines represent data excluding maximum visibility. Hourly observed dataset from 1 January 2022 to 30 September 2024 were used here.
Atmosphere 16 01181 g002
Atmosphere 16 01181 g003
Figure 3. Diurnal Variation in the Correlation Coefficients Between Visibility and PM2.5 and Between Visibility and Relative Humidity in Shanghai, Excluding Data with Visibility of 30 km. Dashed lines represent the Pearson correlation coefficients between the two variables, while solid lines indicate the partial correlation coefficients controlling for the third variable.
Figure 3. Diurnal Variation in the Correlation Coefficients Between Visibility and PM2.5 and Between Visibility and Relative Humidity in Shanghai, Excluding Data with Visibility of 30 km. Dashed lines represent the Pearson correlation coefficients between the two variables, while solid lines indicate the partial correlation coefficients controlling for the third variable.
Atmosphere 16 01181 g003
Atmosphere 16 01181 g004
Figure 4. Scatter Plot of Visibility Versus PM2.5 Concentration, Colored by Relative Humidity (Unit: %).
Figure 4. Scatter Plot of Visibility Versus PM2.5 Concentration, Colored by Relative Humidity (Unit: %).
Atmosphere 16 01181 g004
Atmosphere 16 01181 g005
Figure 5. Modulation of the Relationships Between Visibility and PM2.5 (Top Two Rows) and Between Visibility and Relative Humidity (Bottom Two Rows) by PM2.5 Chemical Components. In the Visibility-PM2.5 Scatter Plots, color Represents the Fraction of Each Chemical Component (Unit: %); in the Visibility-Relative Humidity Scatter Plots, color Represents the Concentration of Each Chemical Component (Unit: μg/m3).
Figure 5. Modulation of the Relationships Between Visibility and PM2.5 (Top Two Rows) and Between Visibility and Relative Humidity (Bottom Two Rows) by PM2.5 Chemical Components. In the Visibility-PM2.5 Scatter Plots, color Represents the Fraction of Each Chemical Component (Unit: %); in the Visibility-Relative Humidity Scatter Plots, color Represents the Concentration of Each Chemical Component (Unit: μg/m3).
Atmosphere 16 01181 g005
Atmosphere 16 01181 g006
Figure 6. Probability Distributions of Relative Humidity and PM2.5 Concentration During Maximum Visibility (30 km) and Low Visibility (<7.5 km).
Figure 6. Probability Distributions of Relative Humidity and PM2.5 Concentration During Maximum Visibility (30 km) and Low Visibility (<7.5 km).
Atmosphere 16 01181 g006
Atmosphere 16 01181 g007
Figure 7. Scatter Plot of PM2.5 and Relative Humidity at Maximum Visibility, with Color Representing the Fraction of Each Chemical Component in Particulate Matter (Unit: %).
Figure 7. Scatter Plot of PM2.5 and Relative Humidity at Maximum Visibility, with Color Representing the Fraction of Each Chemical Component in Particulate Matter (Unit: %).
Atmosphere 16 01181 g007
Atmosphere 16 01181 g008
Figure 8. Performance Evaluation of the Random Forest Model: (a) Scatter Plots of Predicted and Observed Visibility by Model 1 and Model 2 (Unit: m); (b) Residual Distributions of Model 1 and Model 2 (Unit: m).
Figure 8. Performance Evaluation of the Random Forest Model: (a) Scatter Plots of Predicted and Observed Visibility by Model 1 and Model 2 (Unit: m); (b) Residual Distributions of Model 1 and Model 2 (Unit: m).
Atmosphere 16 01181 g008
Atmosphere 16 01181 g009
Figure 9. Feature Importance of the Random Forest Model: (a,c) show the global feature importance for Model 1 and Model 2, respectively; (b,d) show the grouped feature importance for high and low visibility conditions for Model 1 and Model 2, respectively.
Figure 9. Feature Importance of the Random Forest Model: (a,c) show the global feature importance for Model 1 and Model 2, respectively; (b,d) show the grouped feature importance for high and low visibility conditions for Model 1 and Model 2, respectively.
Atmosphere 16 01181 g009
Table 1. Sample Size Statistics of Hourly Scale Effective Visibility Data (2022–2024) and Statistical Characteristics of Hourly Visibility Distribution.
Table 1. Sample Size Statistics of Hourly Scale Effective Visibility Data (2022–2024) and Statistical Characteristics of Hourly Visibility Distribution.
SeasonTotal Samples (h)30 km Visibility (h)Proportion of 30 km Samples (%)Skewness *Kurtosis *
Spring4979210158.4%−0.25 (−1.37)1.96 (3.56)
Summer5270307780.5%−0.34 (−2.71)2.09 (9.53)
Autumn4861391269.6%−0.03 (−1.70)1.86 (4.52)
Winter4624321742.2%−0.05 (−0.79)1.88 (2.20)
Total19,73412,30762.4%−0.09 (−1.41)1.89 (3.61)
* Skewness (>0: right-skewed; <0: left-skewed; =0: symmetric). Kurtosis (>0: more peaked; <0: flatter; =0: normal). Main values exclude visibility = 30 km; parenthetical values include all visibility samples.
Table 2. Regression coefficients of the visibility ridge regression model based on relative humidity and PM (ridge parameter = 1).
Table 2. Regression coefficients of the visibility ridge regression model based on relative humidity and PM (ridge parameter = 1).
λ = 1 All CaseVis < 7500Vis > 7500
Intercept46,10216,28443,405
PM2.5−232.12−13.55−223.65
RH−202.76−114.51−160.45
Table 3. The Statistical Characteristics of the Two Random Forest Models.
Table 3. The Statistical Characteristics of the Two Random Forest Models.
R2RMSE (m)MAE (m)MAPEDurbin-Watson *Breusch-Pagan *
Model 10.832369.31471.3112.3%2.20.16
Model 20.931652.91020.99.8%2.10.13
* If the Durbin-Watson statistic is close to 2, it indicates no significant autocorrelation; if it significantly deviates from 2, it suggests the presence of autocorrelation. For the Breusch-Pagan test, a p-value < 0.05 indicates significant heteroscedasticity, whereas a p-value ≥ 0.05 suggests no significant heteroscedasticity.
Table 4. The Variance Inflation Factors (VIFs) of the Variables in the Two Models.
Table 4. The Variance Inflation Factors (VIFs) of the Variables in the Two Models.
VariablesPM2.5RHDOYhourprecOrgNO3SO4NH4ChlBC
Model 14.66.23.52.84.0
Model 23.85.31.11.02.83.816.36.212.63.33.7
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Gui, X.; Ren, J.; Wang, G.; Wang, Y.; Zhang, M.; Wang, X. Retrieval of Atmospheric Visibility and Its Driving Factors in Shanghai, China. Atmosphere 2025, 16, 1181. https://doi.org/10.3390/atmos16101181

AMA Style

Gui X, Ren J, Wang G, Wang Y, Zhang M, Wang X. Retrieval of Atmospheric Visibility and Its Driving Factors in Shanghai, China. Atmosphere. 2025; 16(10):1181. https://doi.org/10.3390/atmos16101181

Chicago/Turabian Style

Gui, Xiaowen, Jing Ren, Guoyin Wang, Yuying Wang, Miao Zhang, and Xiaoyan Wang. 2025. "Retrieval of Atmospheric Visibility and Its Driving Factors in Shanghai, China" Atmosphere 16, no. 10: 1181. https://doi.org/10.3390/atmos16101181

APA Style

Gui, X., Ren, J., Wang, G., Wang, Y., Zhang, M., & Wang, X. (2025). Retrieval of Atmospheric Visibility and Its Driving Factors in Shanghai, China. Atmosphere, 16(10), 1181. https://doi.org/10.3390/atmos16101181

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop