The nine sampling sites selected for this study encompassed typical river sections from upstream to downstream in the Zhengzhou segment of the Yellow River Basin (Fig. 1). Specifically, the sampling sites comprised natural ecological areas (TH, SD, HT), transportation hubs (HW), tourist hotspots (FS, JY, MD), and village neighborhoods (SL, HY). The selection of sampling sites took into account diverse hydrological characteristics and coastal geomorphology, reflecting the impacts of both human activities and natural ecosystems on bioaerosols. The descriptions of abbreviations used in this article are provided in Table S1. In this study, bioaerosols were collected on March 14 to 15, 2023, and March 9 to 10, 2024. This period coincided with spring, when intensive human activities enhanced bioaerosol production and the likelihood of human exposure. The meteorological condition parameters on the sampling site days were shown in Table S2.

The Andersen eight-stage microbial air sampler (PSW-8) was employed for the collection of bioaerosols. The size range of particles collected at each stage is detailed in Table S3. The sampler was decontaminated using alcohol-soaked cotton before sampling. Throughout the sampling process, the sampler’s flow rate was maintained at 28.3 L/min, with three consecutive 3-minute collections performed at each sampling site., The collected microorganisms were subsequently stored under low-temperature and light-protected conditions. Environmental parameters, including temperature, relative humidity, wind speed, and light intensity, were recorded concurrently using a dew point thermohygrometer (WD-35612, OAKTON, USA), a portable anemometer (HD2303, Delta OHM, Italy), and an irradiance meter (HD2302.0, DeltaOHM). Air samples were collected, sealed, and immediately transported to the laboratory for analysis. The air-exposed petri dishes were incubated at 37°C for 48 hours to assess bacterial growth. The average number of bacteria collected at each sampling point was determined and used in Eq. (1) to calculate the bioaerosol concentration.

where C_i is the airborne bacterial concentrations in CFU/m³, N_i is the effective number of bacteria on Petri dishes in CFU, t is the sampling time in minutes, and Q is the gas flow rate during sampling in L/min.

Water samples were collected utilizing a plexiglass water collector (CSQ, Henan, China). Under sterile conditions, the samples were serially diluted, plated on appropriate agar media, and incubated for bacterial growth. The number of bacterial colonies formed was then used to calculate the concentration of culturable bacteria in the water.

After incubation at 37 °C for 48 hours, colonies were collected in PBS buffer for sequencing. DNA was extracted and the V4–V5 regions of 16S rDNA were amplified by PCR. Products were purified and checked by agarose gel electrophoresis. Qualified products were sequenced using Illumina MiSeq and processed with Trimmomatic, Cutadapt, and FLASH for quality control. Operational Taxonomic Units (OTUs) were clustered at 97% sequence similarity using UPARSE, and taxonomic classification was performed using the RDP Classifier against the SILVA database with a confidence threshold of 70% [17]. Feature sequences were annotated using the SILVA database with a Bayesian classifier to assign taxonomic information. Community compositions were analyzed at multiple taxonomic levels (phylum to species) using QIIME2 (version 2020.6) and the results were visualized using R.

Random forest analysis (RFA), renowned for its accuracy and stability in assessing group differences, was used to evaluate species contributions to sample classification. Plots were generated using R (version 3.1.1) with the Random Forest package (version 4.6–10). The ecological functions of microorganisms were predicted using FAPROTAX, while their potential pathogenicity was assessed using BugBase.

In this study, exposure and risk were assessed using a model developed by the United States Environmental Protection Agency (USEPA) and parameters from the Exposure Factors Handbook (USEPA, 2011). The model assumes inhalation and dermal contact as the primary exposure routes, with constant exposure frequency and duration. To adapt the model to the Yellow River context, parameters were adjusted based on Chinese population data and local surveys. Exposure levels were calculated as the average daily dose (ADD). The respiratory inhalation dose and skin absorption dose of bioaerosols were calculated using Eq. (2), (3), respectively.

For specific symbol meanings and parameter details, refer to Table S4, S5. For non-carcinogenic risk assessment, the ADD was compared to the chronic reference dose (RfD) to evaluate potential health risks. The risk of a non-carcinogenic pollutant was expressed as a hazard quotient (HQ). An HQ value exceeding 1 indicates a potentially significant risk [18]. The assessment of HQ was calculated using Eq. (4).

In this study, we employed the Hybrid Single-Particle Lagrangian Integrated Trajectory (HYSPLIT) model, developed by the National Oceanic and Atmospheric Administration (NOAA), to calculate and analyze the dispersion trajectories of the Yellow River bioaerosols. The HYSPLIT model is a comprehensive system for simulating transport, dispersion, and deposition processes, covering a wide range of physical processes and emission sources. The model tracks the directional movement of airborne particles or gases, offering critical insights into the atmospheric transport and dispersion of pollutants [19].

Radial Basis Function Neural Networks (RBFNN), a type of feedforward neural network, are widely used for function approximation, pattern recognition, and classification. Specifically, the nonlinear relationship between bacterial biological characteristics and risk indicators can be effectively modeled by RBFNN using radial basis functions in its hidden layer [20]. Moreover, the local receptive field characteristic of RBFNN incorporates the concept of distance in decision-making, which facilitates the consideration of similarities or differences between samples during the recognition process. This feature enhances the network’s ability to identify characteristic bacteria with higher interpretability of HQ. In this study, RBFNN is employed to identify risk marker bacterial genera in the Yellow River bioaerosols.

Statistical analyses, including normality and chi-square tests, were performed using SPSS version 27.0 (SPSS, Inc., Chicago, IL, USA). The significance level was set at α = 0.05. Principal component analysis (PCA) was performed in R (version 3.1.1) with the ggplot2 package. Canonical Correlation Analysis (CCA) was carried out in R using the vegan package (version 2.3.0).

Fig. 2a illustrates the culturable bacterial concentrations in bioaerosol samples collected from multiple sites along the Yellow River. In 2023, the concentrations ranged from 96 to 519 CFU/m³, with a mean of 226 ± 138 CFU/m³. Concentrations of culturable bacteria in bioaerosols reached a maximum at the TH (519 CFU/m³) in 2023. In 2024, the range was broader, from 153 to 908 CFU/m³, and the mean concentration increased to 456 ± 218 CFU/m³. The highest concentration was reached at HY (908 CFU/m³) in 2024. A paired-sample t-test indicated a statistically significant difference in bacterial concentrations of bioaerosols between 2023 and 2024 with a strong effect size (p < 0.05, Cohen’s d = 1.037). Notably, the mean concentration in 2024 was approximately double the value recorded in 2023.

The culturable bacterial concentrations (Fig. 2b) in water samples varied across sampling points. In 2023, the concentrations ranged from 1.15×10² to 3.25×10⁴ CFU/ml, with an average concentration of 5.89×10³ ± 1.08×10⁴ CFU/ml. In 2024, the range was 3.5×10² to 1.62×10⁴ CFU/ml, and the average concentration was 5.37×10³ ± 5.95×10³ CFU/ml. Paired-sample t-tests revealed no statistically significant difference in bacterial concentrations of water samples between 2023 and 2024 (p > 0.05, Cohen’s d = 0.041). The presence of enteric bacteria was identified in water samples collected in both years (Fig. 2c).

The particle size distribution of bioaerosols exhibits variations across different sampling times and locations (Fig. 2d, 2e). In this study, a cutoff of 3.3 μm was used to categorize bioaerosol particles with diameters less than 3.3 μm as fine particle bioaerosols (FPB) [21]. In 2023, the percentage of FPB at each sampling point ranged from 16% to 89%, with an average of 47% (Fig. 2f). In 2024, the range was from 7% to 75%, with the average percentage increasing to 50% (Fig. 2g). In 2023, the highest FPB percentages were observed at HW (89%), SD (69%), and TH (52%). For 2024, the highest percentages were found at HT and HW (both 75%), followed by HY (73%).

Canonical Correspondence Analysis (CCA) revealed that wind speed did not significantly influence the distribution of bacterial communities (p > 0.05) (Fig. 4). Humidity and temperature were identified as the primary factors shaping the microbial community structure. However, their interaction exhibited a negative effect. In contrast, humidity demonstrated a synergistic relationship with microbial concentrations in water. The CCA biplot (Fig. 4) further indicated that meteorological conditions likely play a dominant role in regulating bioaerosol community dynamics, while microbial inputs from the water column had a secondary influence. Specifically, Bacillus and Enterobacter showed positive correlations with humidity and water microbial concentrations but were negatively associated with temperature. Additionally, Lysinibacillus was positively correlated with temperature, whereas Pseudomonas exhibited a negative correlation with light intensity.

FAPROTAX analysis (Fig. 5a, 5b) indicated that the dominant functional genes in bioaerosols during 2024 were primarily associated with chemoheterotrophy, aerobic chemoheterotrophy, and fermentation. Additionally, bioaerosol samples collected from FS, TH, and JY in 2023 were found to harbor a significant number of human pathogens. Phenotypic predictions using BugBase (Fig. 5c, 5d) revealed that bioaerosol samples collected in 2024 exhibited a higher overall potential for generalized pathogenicity compared to 2023. In 2023, bioaerosol samples from HW, TH, and HT exhibited higher potential pathogenicity than those from other sites. In 2024, samples from FS, SD, SL, and MD showed higher potential pathogenicity than those from other sites.

Based on the average daily dose results (Fig. 5e, 5f), the inhalation hazard quotient (HQ_inhalation) in 2024 (HQ_male: 4.21 × 10⁻²; HQ_female: 3.36 × 10⁻²) was higher than in 2023 (HQ_male: 2.09 × 10⁻²; HQ_female: 1.67 × 10⁻²). Similarly, the dermal contact hazard quotient (HQ_skin) in 2024 (HQ_male: 5.10 × 10⁻⁷; HQ_female: 4.76 × 10⁻⁷) was also greater than that in 2023 (HQ_male: 2.53 × 10⁻⁷; HQ_female: 2.36 × 10⁻⁷). Overall, the risk assessment results indicated a trend of increased respiratory and dermal exposure risks in 2024 compared to 2023. Additionally, the risk of bioaerosol exposure was consistently higher in males than in females. Although the hazard quotient at each sampling site remained below 1, prolonged exposure may pose health risks to humans.

In this study, a RBFNN model was constructed with bacterial genus concentrations as covariates and the HQ_inhalation as the dependent variable. The model was utilized to identify high-risk marker bacteria based on the importance of independent variables (Fig. 5g). The results revealed a hidden layer consisting of six nodes, designated as H(1) through H(6). Notably, the weight associated with H(6) was 2.655, the highest among all weights from the hidden layer to the output layer. This indicates that variations in the activation of H(6) are likely to have a substantial impact on the predicted value of HQ_inhalation. The importance of the independent variables was evaluated, with Enterobacteriaceae identified as the most significant contributor (0.415), followed by Bacillus (0.175) and Micrococcaceae (0.166). These findings imply that Enterobacteriaceae may serve as a high-risk biomarker within the bioaerosols of the Yellow River.

The HYSPLIT analysis showed that bioaerosol samples from nine sampling sites in the Zhengzhou section of the Yellow River were substantially influenced by riverine air mass transport (Fig. 6). After one hour of simulated dispersion, bioaerosols at all sampling sites exhibited an elliptical distribution extending towards the downwind direction. The microbial dispersal area was the largest at HY (417.86 km²) and the smallest at TH (39.02 km²). The microbial concentration at the dispersion source of each sampling site exceeded a threshold of 1.0 × 10⁻⁶ mg/m³, with the maximum concentration observed at TH (3.0 × 10⁻⁶ mg/m³). The microbial concentration was lower in the area farther away from the source point of dispersal.

This study revealed a mean culturable bacterial concentration of 318 ± 208 CFU/m³ in bioaerosols over the Yellow River. This concentration range aligns with levels reported in marine bioaerosols from the Qingdao coastal area (25 to 561 CFU/m³) but is markedly lower than the bacterial concentrations observed in urban bioaerosols from Lanzhou (2 × 10⁵ to 1.1 × 10⁶ CFU/m³) [6, 22]. These disparities likely reflect the distinct environmental conditions and pollution levels between aquatic and urban ecosystems. For instance, the elevated concentrations in Lanzhou may be attributed to frequent dust events and stagnant meteorological conditions, whereas the lower levels in Qingdao could stem from dilution effects driven by marine climates [23, 24]. Notably, comparative analysis of the two-year dataset demonstrated an upward trend in bacterial concentrations in both air and water samples from the Yellow River in 2024, potentially linked to increased tourism activity and agricultural operations following the relaxation of pandemic-related restrictions [25].

Studies have demonstrated that bioaerosols are predominantly concentrated in the coarse mode (> 3.3 μm) at a large scale. For example, in the coastal region of Qingdao, the bioaerosol concentration peaks in the size fraction > 7.0 μm [6]. In contrast, the particle size distribution observed in this study reveals that fine particles constitute nearly 50% of the total bioaerosols. High humidity and sea salt particles in coastal regions are more conducive to the formation and suspension of coarse particles, whereas the dry climate and frequent sand activities in the Yellow River region may result in a higher proportion of fine particles [26]. Previous research has indicated that bioaerosol particles, being small in volume and light in mass, can enter biological bodies through the skin, mucous membranes, and respiratory tract [27]. Size and shape differences lead to distinct penetration and deposition mechanisms of bioaerosols in different parts of the human respiratory system. Particles smaller than 2.5 μm could more easily pass through the human pharynx to the trachea and alveoli and other tissues, causing serious harm to the human body [28]. Additionally, small-sized bioaerosol particles can travel long distances in the wind, posing a risk to large surrounding populations [15]. Therefore, analyzing the small-sized particulate matter of bioaerosols was of great practical significance. The percentage of fine particle bioaerosols exceeded 70% at both the 2023 HW (88.79%) and 2024 HT (75.00%). Severe soil erosion and overcrowding were the main causes of this phenomenon [29]. This suggested that the surrounding population may be exposed to significant bioaerosol inhalation risk.

As an important component of atmospheric ecosystems, changes in the species composition and functional genes of bioaerosols significantly influence the cycling of matter and energy flow in ecosystems [30]. Functional prediction analysis revealed that chemotaxis-associated heterotrophic functional genes dominated the bioaerosol communities in the Yellow River, a pattern potentially linked to the high salinity and severe pollution pressures characteristic of this environment. These genetic traits confer heightened adaptability and metabolic versatility to the microbial communities [31]. High-throughput sequencing results showed that Bacillus was detected at all sampling sites, and its spore-forming mechanism may help it survive and dominate the extreme environment in the Yellow River region [32]. In addition, differences in dominant genera were observed between sampling sites, potentially reflecting the differentiation of ecological functions under distinct environmental conditions. For instance, Enterobacter harbors diverse functional genes enabling organic pollutant degradation [33], while Stenotrophomonas exhibits plant growth-promoting traits that enhance host tolerance to environmental adversities [34].

Phenotypic prediction using BugBase indicated that bioaerosol samples collected in 2024 exhibited a broader pathogenic potential compared to those from 2023. Elevated bioaerosol concentrations may increase the exposure dose of pathogenic bacteria, thereby heightening the risk of infection [35]. Furthermore, the presence of toxin genes in Bacillus further amplifies this pathogenic risk [36]. In this study, Pseudomonas, Enterobacter, Stenotrophomonas, and Lysinibacillus were identified as potential pathogens. Pseudomonas, an opportunistic pathogen, can invade multiple organ systems, causing respiratory, gastrointestinal, urinary tract, and skin infections [37]. Multidrug-resistant Pseudomonas aeruginosa has become a major public health concern worldwide due to its escalating prevalence [38]. Enterobacter can directly invade the lower respiratory tract through inhalation, leading to community-acquired and hospital-acquired pneumonia [39]. Recent studies have also highlighted its role in urinary tract infections, skin and soft tissue infections, and intra-abdominal infections [40]. Stenotrophomonas, a well-documented pathogen, is associated with various infections, including respiratory, urinary tract, and bloodstream infections [41]. In immunocompromised populations, Lysinibacillus may colonize the respiratory tract and trigger inflammatory responses, resulting in symptoms such as coughing, wheezing, and dyspnea [42]. Given these risks, regular monitoring of bioaerosols in the Yellow River Basin, coupled with enhanced personal protection and public health education, is essential to improve regional ecological and public health conditions.

The weather conditions exerted an influence on the production, release, dispersal, and deposition of microorganisms, as well as on the diversity and abundance of airborne biological particles [43, 44]. The negative interactive effect between temperature and humidity in our results may indicate a characteristic climatic contradiction during spring in the Yellow River Basin. Rising temperatures intensify soil moisture evaporation, thereby causing a decline in air humidity [45]. This phenomenon would likely inhibit the aerosolization process of water bodies, while facilitating the dust-mediated dispersal of xerotolerant spore-forming microorganisms [32].

This study found that Stenotrophomonas was directly proportional to temperature, a relationship that has been reported in numerous studies [46]. In contrast, Enterobacter exhibited an inverse relationship with temperature. Some studies have shown [47] that cold air may promote the release and transportation of bacteria. Therefore, the effect of temperature on different microorganisms was not consistent [48]. The combined effect of humidity and temperature is likely to produce a more pronounced impact on these microbial characteristics [49]. Most Gram-negative bacteria in bioaerosols have a longer survival time under lower relative humidity conditions. In contrast, Gram-positive bacteria were more likely to survive at higher relative humidity levels [50]. Different microorganisms responded differently to humidity, and even some did not respond [51]. Although this study showed that the correlation between wind speed and the survival of dominant strains in bioaerosols was not significant, it was likely due to the small differences in wind speeds at the sampling sites and the insufficient sample size. Numerous studies have demonstrated that wind is a driving force for atmospheric microbial production and dispersal [52, 53]. In the present study, it was found that Bacillus had an inverse relationship with light exposure. Studies [54] have shown that microbial inactivation is directly proportional to the intensity of ultraviolet (UV) radiation.

Our analysis indicated that both bioaerosol and riverine bacterial concentrations were elevated at TH. It is hypothesized that riverine bacteria, acting as a source of bioaerosols during river flow, are dispersed into the atmosphere [55]. Simultaneously, atmospheric bioaerosols may influence bacterial concentrations in the water through the alteration of the water cycle [56, 57]. This reciprocal relationship underscores the complex interactions between the atmospheric and aquatic ecosystems, highlighting the potential for bioaerosols to serve as vectors for microbial exchange between these environments.

In this study, it was found that men had a higher respiratory exposure risk and dermal contact risk to bioaerosols compared to women, which was similar to previous research findings [58, 59]. This was primarily attributed to men’s generally larger lung capacity and breathing rate, as well as their preference for outdoor activities [60]. Although this study indicated that the non-carcinogenic exposure risk of bioaerosols surrounding the Yellow River was within an acceptable range, the intake of pathogenic bacteria may still lead to the onset and exacerbation of diseases, given the potential for prolonged exposure among local residents [61].

Previous studies have demonstrated the feasibility and potential of RBFNN in identifying biomarkers and assessing health risks [62]. For instance, Xie et al. reported that RBFNN achieved a significantly higher accuracy (97.62%) in detecting biomarkers for food adulteration compared to Partial Least Squares Discriminant Analysis and Linear Discriminant Analysis [63]. Similarly, R S Aswal developed a predictive model to assess the hazards of potentially toxic elements in groundwater, revealing that RBFNN outperformed Multilayer Perceptron models, with coefficient of determination (R²) values ranging from 0.912 to 0.976 [64]. These findings collectively highlight the superior capability of RBFNN in complex pattern recognition and high-precision prediction tasks. In this study, the established model successfully identified Enterobacteriaceae as a high-risk indicator (weight = 0.415). Unlike deep learning models such as Convolutional Neural Networks, the hidden layer weights of RBFNN can be directly linked to the ecological and health risk contributions of bacterial genera, providing a transparent basis for developing targeted intervention strategies [65].

Most bioaerosols exist in the atmosphere in the form of attachment to particulate matter and are transported among different regions through the diffusion and transport of air masses [2]. Cáliz employed HYSPLIT to track air mass trajectories, thereby identifying potential source regions for microbial communities transported to receptor sites [66]. Based on wind direction data from each sampling site and the population distribution along the riverbank, it was observed that there were residential villages downwind of SL. Moreover, the downwind areas of HY and SD were tourist attractions frequented by many visitors. Consequently, at several sampling sites, the potential risks of bioaerosols to human health must be approached with caution.