Spatiotemporal distribution trends and driving factors of avian influenza in poultry in China

Abstract

Background

Avian influenza viruses can cross species barriers to mammals, spreading among wildlife, domestic animals and humans. Poultry, as a primary host, plays a crucial role in the transmission of avian influenza.

Methods

This study analyzed the temporal patterns and hotspot regions of avian influenza in poultry, examined the distribution trends of major subtypes, identified driving factors on the avian influenza transmission, and assessed their contributions and interactive effect using the XGBoost model.

Results

Poultry avian influenza in China mainly occurred from November to March, with hotspot areas concentrated in southeastern China. The H5 subtype persisted throughout the study period, and its epidemic center gradually shifted eastward. The H7 and H9 subtypes showed different temporal trends in circulation, while H9N2 accounted for the highest number of reported cases. Human activity-related variables, including road network density, population density, GDP and poultry trade points, contributed the most to poultry avian influenza outbreaks. Climate factors ranked the second, with winter precipitation contributing the most. Livestock and poultry farming density ranked the third and waterfowl species density ranked the last. The increase of human activities, livestock and farming densities contributed to a sharp increase in avian influenza risks when the levels of these factors were at low-medium level, while the contribution became stable when the human activities exceeded to certain threshold, and became negative when the livestock and farming densities were extremely high. Further, poultry avian influenza risk was amplified under the combined effects of livestock and poultry farming systems, high waterfowl species density, and strong trade connectivity, suggesting a coupled wild bird-poultry-trade network transmission pattern.

Conclusion

These findings highlight that socioeconomic activities are the major contributors for poultry avian influenza, while climate and waterfowl species density acted as important modifying factors for avian influenza, in particular with livestock and poultry farming density. The study provides a useful framework for identifying potential transmission interfaces and supports more targeted surveillance and prevention strategies under the One Health perspective.

1. Introduction

Avian influenza (AI) poses a major threat to poultry production and public health worldwide, particularly in China, where dense poultry farming, extensive live poultry trade, and frequent interactions between domestic poultry and wild waterfowl create favorable conditions for viral transmission [1], [2], [3], [4]. The avian influenza virus is commonly transmitted among birds and poultry and is categorized into low-pathogenic avian influenza (LPAI) and highly pathogenic avian influenza (HPAI) based on its virulence in avian species. Certain avian influenza virus strains are capable of crossing species barriers, leading to occasional spillover infections in mammals such as dairy cattle, minks, red foxes and even humans [5], [6], [7].

Avian influenza virus subtypes varies with time and space due to internal factors, mainly mutation and genetic reassortment, and external selective pressure [8], [9], [10]. In 2009, clade 2.3.2 of the H5N1 virus emerged among wild birds at Qinghai Lake and subsequently circulated in both wild birds and poultry. Between 2010 and 2019, eleven H5 avian influenza outbreaks were reported in wild birds in China, with 27 viral isolates identified as belonging to clade 2.3.2.1 and clade 2.3.4.4 [11]. Notably, clade 2.3.4.4 exhibits strong reassortment capacity, with H5N6 and H5N8 as its major subtypes. Since 2014, gs/Gd-lineage clade 2.3.4.4 HPAI H5 viruses have spread from East Asia to North America, West Asia, and Europe, causing multiple outbreaks in wild birds and poultry worldwide [12]. After 2014, H5N6 gradually replaced H5N1 as the dominant subtype infecting poultry in southern China [13], and severe human cases caused by these novel H5N6 viruses have also been reported [14]. In China, H7N9 viruses initially emerged mainly as LPAI. However, H7 subtype viruses are widely present in wild waterfowl and can spill over to poultry and mammals, posing serious threats to poultry industry and public health [15], [16]. Studies have shown that LPAI H7N9 viruses in China have evolved into two distinct lineages—the Yangtze Delta and Pearl Delta—based on their HA and NA genes [17]. LPAI H7N9 viruses circulating in poultry can mutate during transmission and transform into HPAI H7N9 [18], [19], [20], [21]. Between 2013 and 2017, five epidemic waves of H7N9 occurred in China, during which HPAI H7N9 first emerged in poultry in the fifth wave and subsequently transmitted to humans [22], [23]. Since the 1990s, the H9 subtype of LPAI has also become widespread in poultry populations across several countries. H9 subtype AIVs have been reported in poultry and wild birds across multiple regions, including North America, Europe, Asia, Africa, and the Pacific, with H9N2 as the predominant subtype and h9.4.1 and h9.4.2 as the major lineages. In China, the dominant lineage is h9.4.2, which has further evolved into two major sublineages, h9.4.2.5 and h9.4.2.6, with h9.4.2.5 currently being the most prevalent [24]. Studies have shown that some H9 lineages are more widely distributed in domestic poultry than in wild waterfowl. Analysis of H9 viruses isolated in China during 2008–2011 revealed that over 90% of viruses from the h9.4.1, h9.4.2, and h9.3.3.1 lineages were derived from domestic terrestrial birds, suggesting that these lineages may have gradually adapted to terrestrial poultry under farming conditions [25].

The factors influencing the occurrence and spread of avian influenza are highly complex. Wild waterfowl are the natural hosts of avian influenza viruses, and the activities of wild birds are considered a key factor in the long-distance transmission of the virus [26], [27]. Large-scale, intercontinental spread of avian influenza viruses aligns closely with migratory bird routes [12], [28], [29]. These viruses typically reside latently in wild waterfowl and spread to other continents through seasonal migrations [30], [31]. Additionally, the spatial connectivity between natural bird habitats and human-reared facilities, as well as the food resources provided by farms, often attract wild birds to forage, drink, and roost within livestock facilities [32]. Birds carrying avian influenza viruses can transmit the virus to livestock and poultry through direct contact or indirect pathways, such as fecal-oral transmission. Consequently, live poultry farming, poultry meat, and livestock trade also serve as significant pathways for carrying and spreading avian influenza viruses [33], [34]. Poultry farms and live poultry markets are areas with high prevalence of avian influenza, and may serve as key hotspots for the persistence and spread of the virus [35], [36]. In particular, live poultry markets in East and Southeast Asia provide conditions conducive to the spread and cross-mutation of various avian influenza viruses [37]. Previous survey in ten live poultry markets in southern China has detected twelve types of avian influenza viruses, and these strains exhibit rapid genetic variation and show significant divergence from the strains identified in the previous year. Avian influenza viruses have frequently been detected in pig populations, with subtypes such as H5N1 and H9N2 showing increasing mammalian adaptation potential [38], [39]. Because pigs are susceptible to both avian and mammalian influenza viruses, exposure to avian influenza viruses in pigs may further facilitate viral reassortment and cross-species transmission. In addition, recent surveillance studies have detected highly pathogenic H5 viruses in wild mammals in Europe, further highlighting the growing importance of mammalian hosts in avian influenza ecology [40], [41].

The natural environment influences the occurrence of avian influenza in two ways. It directly affects the activity and transmission of pathogens, and indirectly affects avian influenza by shaping the behavior patterns of hosts [42]. Most studies suggest that avian influenza exhibits seasonality, with cold winter temperatures prompting wild waterfowl to congregate at habitats for extended periods, resulting in continuous virus transmission among birds and even outbreaks among humans. Additionally, due to food shortages and higher energy demands under extreme cold conditions, potential hosts are more vulnerable to avian influenza virus infection [43]. Research has confirmed that low temperatures and low humidity are more conducive to influenza virus survival [44], including H7N7 [45], H5N1 [46], [47], [48], H9N2 [49], [50]. However, studies in East Asia have revealed that low temperatures coupled with high humidity may facilitate the transmission of H7N9 [51], [52]. Furthermore, precipitation has a positive correlation with avian influenza incidence. This is partly because moist surfaces and rainwater-contaminated poultry drinking water after precipitation create conditions favorable for virus spread. In addition, the breeding and population size of wild waterfowl residing around wetlands are also affected by precipitation levels [53], [54]. In the Murray–Darling Basin of Australia, heavy rainfall events have been identified as key environmental drivers of avian influenza outbreaks in poultry, primarily involving the endemic H7 and H5 subtypes. Increased rainfall promotes wild waterfowl breeding, leading to a surge in juvenile birds, which subsequently congregate near poultry farms during dry periods. This facilitates viral spillover, with outbreak risk peaking approximately two years after major rainfall events [55]. Wind may accelerate the migration of airborne influenza viruses, facilitating their spread [56].

In summary, avian influenza subtypes change with time and space, and the combination of environment, socioeconomic and human activities together influencing the outbreak, in particular for poultry avian influenza. Although the previous studies have provided valuable insights on the pathogenesis of avian influenza and its meteorological driving features through mainly local scale case studies, the interactive roles of climate variability, wild waterfowl distribution, and anthropogenic factors in shaping the spatial-temporal dynamics of avian influenza in poultry across China remain poorly understood. In this study, we firstly displayed the spatiotemporal epidemic patterns of H5, H7, and H9 subtypes in poultry from 2010 to 2025 across China, and then conducted a multi-factor analysis on the epidemic of these subtypes at the national scale. By integrating multiple environmental and socio-economic drivers across the entire country, this study captures the spatiotemporal feature of the poultry avian influenza dynamics, and enables a more comprehensive understanding of epidemic patterns and their underlying determinants. The results of this study could provide scientific evidence for effective prevention and control strategies on poultry avian influenza.

2. Data and methods

2.1. Data selection

Based on the exploration of the epidemiological patterns of avian influenza in poultry and the existing literature, this study considered winter as a sensitive period for the outbreak of avian influenza when constructing climate environmental indicators. Also, waterfowl density is included considering that wild-domestic interaction is the major route for avian influenza spillover from wild to poultry. Additionally, the study took into account the socioeconomic and human activities that potentially driving the transmission in the poultry system, including various aspects of socioeconomic and human activities, such as transportation, livestock farming, and poultry trade. Ultimately, 8 natural environmental factors and 7 socioeconomic factors were selected (Table 1).

Table 1.

Indicators selected for the study.

Name	In-detail explanation, unit	Time period	Spatial resolution	Reference
Temperature	Annual number of days with temperature > 30 °C, days	Training and testing: 2010–2017 External validation: 2021	5 km	[23], [26], [28], [32]
Winter precipitation	Total precipitation from November to March, 10 mm	Training and testing: 2010–2017 External validation: 2021	5 km	[[51], [57]]
Atmospheric Pressure	Annual average atmospheric pressure, hpa	Training and testing: 2010–2017 External validation: 2021	5 km	[43]
Wind speed rate	Annual number of days with wind speed rate > 4 m/s, days	Training and testing: 2010–2017 External validation: 2021	5 km	[[58], [59]]
Elevation	Digital elevation model, meters	Training and testing: 2010–2017 External validation: 2021	1 km	[[60], [61]]
Vegetation	Normalized Difference Vegetation Index	Training and testing: 2010–2017 External validation: 2021	1 km	[61]
Wetland	Binary presence of wetlands, 1 for presence and 0 for absence	Training and testing: 2021 External validation: 2021	10 m	[[62], [63]]
Waterfowl	Waterfowl species density, species/km2	Training and testing: 2022 External validation: 2022	5 km	[[2], [16], [64]]
Economy	Gross domestic product, million yuan	Training and testing: 2010–2017 External validation: 2021	1 km	[65]
Road network	Road network density, km/km2	Training and testing: 2010, 2014, 2017 External validation: 2021	5 km	[[66], [67], [68]]
Poultry trading points	Number of poultry trading points, count	Training and testing: 2023 External validation: 2023	5 km	[34], [36], [37]
Population density	Annual population density, person/km2	Training and testing: 2010–2017 External validation: 2021	1 km	[[69], [70]]
Chicken farming	Annual chicken farming density, animal/km2	Training and testing: 2010, 2015 External validation: 2020	10 km	[[71], [72]]
Duck farming	Annual duck farming density, animal/km2	Training and testing: 2010, 2015 External validation: 2020	10 km/5 km	[[34], [36], [66]]
Pig farming	Annual pig farming density, animal/km2	Training and testing: 2010, 2015 External validation: 2020	10 km	[39], [40], [41]

2.2. Date sources and processing

2.2.1. Avian infection data

Based on the epidemiological characteristics of avian influenza and the risk factors associated with poultry avian influenza outbreaks, a spatiotemporal database was established for poultry avian influenza cases and environmental factors in China from 2010 to 2025. The avian influenza case data were obtained from the Food and Agriculture Organization (FAO), compiling reported poultry avian influenza cases across China during 2010–2025. These data included cases of the H9N2, H7N9, H5N6, H5N1, H5N2, H5N8, H5N3 and H7N2 subtypes.

Due to the mandatory vaccination for H7 since the end of 2017, the avian influenza cases declined and only 217 avian influenza cases were reported from 2018 to 2025. To avoid the bias caused data imbalance, improve the stability of the validation and reduce the influence of interannual fluctuations, all outbreak records from 2018 to 2025 were combined into a single external validation dataset. Environmental and socioeconomic variables from 2021, the year with the highest number of reported cases during this period, were used as representative background conditions for external validation. The external validation dataset was mainly used to evaluate the temporal generalization and transferability of the model.

2.2.2. Migratory water birds species density [73]

Water-birds are the main host for avian influenza virus, and migratory water-birds in particular are responsible for the long distance transmission of the virus. Data on the distribution of key bird species (Bird Species Distribution Maps of the World) were sourced from the BirdLife International platform (http://datazone.birdlife.org/species/requestdis), with metadata provided by BirdLife International and the Handbook of the Birds of the World (2022). This study focused on eight orders of aves that contain waterbird species potentially linked to the transmission of avian influenza, there are, the orders of Gaviiformes [74], Podicipediformes [75], Procellariiformes (also known as Tubinares) [76], Pelecaniformes [[77], [78]], Anseriformes [[75], [79], [80]], Ciconiiformes [81], Gruiformes [82], and Charadriiformes (which includes gulls and terns) [[80], [83]], were selected [[16], [84]]. Species that have clear records in the present or recent past are selected, that is, in the “Presence” attribute of the dataset, “Extant” category was selected. In addition, this study specifically focused on bird species with pronounced seasonal migratory behaviors under the attribute of “Season”. Three categories under this attribute were selected: “Breeding Season”, “Nonbreeding Season” and “Passage”. After the screening above, the layers of the eight orders was put together as the migratory water birds species density, representing the diversity of the susceptible birds.

2.2.3. Climate data

China's surface meteorological data from 2010 to 2021 were sourced from the National Climatic Data Center (ftp://ftp.ncdc.noaa.gov/pub/data/noaa/isd-lite/), including variables such as temperature (10 °C), precipitation (10 mm), atmospheric pressure (hPa), and wind speed rate (10 m/s). These data were processed and rasterized to create climate environment indices based on meteorological station data across China. Climate indices were developed based on meteorological station data across China and were interpolated into 5-km raster layers.

2.2.4. Road and economic density data

Road network data were obtained from the Open Street Map (OSM) platform, covering road network information for China in 2010, 2014, 2017, and 2021. Firstly, a 5-km grid network is created, and the area of each individual grid cell, as well as the total length of roads within each grid cell, is calculated separately. The road network density is then ascertained by determining the ratio of the road length to the grid cell area within each grid.

Population density data were downloaded from the WorldPop platform (https://hub.worldpop.org/), with a resolution of 1 km, covering the years 2010–2021. GDP data were derived from “Forecasting China's GDP at the pixel level using nighttime lights time series and population images” and China's spatial GDP distribution kilometer-grid dataset, both with a spatial resolution of 1 km. All datasets were resampled to a spatial resolution of 5 km.

2.2.5. Poultry trade data

Poultry trade data for China was collected and organized through a keyword-based searching strategy using web-scraping techniques on the Gaode (Amap) platform, which provides real-time Points of Interest (POI) information relevant to live poultry trade across China. Firstly, designate “live poultry”, “live chickens”, “broiler chickens”, “meat ducks”, “pigeons”, and “quails” as the primary set of keywords. Subsequently, combine this primary set with “farming”, “slaughter”, “market”, and “base” to form a secondary set of keywords. Utilize these secondary keywords to extract relevant data on live poultry trade across the nation, and subsequently eliminate any duplicate entries. Given the substantial volume of data acquired through web scraping, additional process of data cleansing was conducted to refine the dataset. Data containing keywords but irrelevant to the study's scope was removed to ensure the dataset's accuracy. After filtering out noise, a total of 15,482 Points of Interest (POI) highly relevant to the research objective were retained. Then, a 5-km grid network was created to calculate the number of trade locations within each grid cell, and a 5-km raster layer was generated.

2.2.6. Poultry and livestock data

Poultry and livestock data were sourced from the FAO's Gridded Livestock of the World datasets (GLW3 and GLW4; https://data.apps.fao.org/catalog/dataset). These datasets provided the chicken and pig farming densities for 2010, 2015, and 2020, as well as the duck farming densities for 2010 and 2015, at a spatial resolution of approximately 10 km. For 2020, the duck farming density data were sourced from the Annual Global Gridded Livestock Mapping dataset with a spatial resolution of 5 km [85]. All datasets were resampled to a spatial resolution of 5 km.

2.2.7. Topographic and land cover data

The digital elevation model (DEM) data were obtained from the Shuttle Radar Topography Mission (SRTM) and are based on the WGS84 ellipsoid projection. The wetland classification data were sourced from the Black Soil and Wetland SubCenter, National Earth System Science Data Center, National Science & Technology Infrastructure of China (http://northeast.geodata.cn). This dataset has a spatial resolution of 10 m and was further resampled to 5 km.

Normalized Difference Vegetation Index for 2010–2017 and 2021 were obtained from the Resource and Environment Science Data Center of the Chinese Academy of Sciences (http://www.resdc.cn/DOI), with an original spatial resolution of 1 km, and were further resampled to 5 km.

To align data scales across different sources, a log(x + 1) transformation was used to normalize variables with highly skewed distributions. These variables included population density, GDP, road density, and poultry and livestock density. This transformation reduced the impact of extreme high values on model training. Additionally, all spatial variables were unified to a 5 km spatial resolution to ensure consistent model inputs.

2.3. Methods

2.3.1. Spatial analysis methods

• 1)Getis-Ord Gi*

The Getis-Ord Gi* is a local spatial autocorrelation method used to measure spatial association by calculating the Gi statistic for each feature. It assesses whether the value of a given feature is significantly higher or lower compared to its neighboring features. A high z-score and a low corresponding p-value indicate a significant cluster of high values, forming a hotspot. Conversely, a low z-score indicates a cold spot, representing a significant cluster of low values [86]. In this study, the Getis-Ord Gi* statistic was applied to identify statistically significant spatial clusters of avian influenza in poultry across the study area, including hotspots and coldspots, in order to reveal underlying spatial patterns and potential risk areas.

$$G_i^{\ast }=\frac{\sum _{j=1}^n{w}_{\mathit{ij}}{x}_j-\overline{X}\sum _{j=1}^n{w}_{\mathit{ij}}}{S\sqrt{\frac{\left[n\sum _{j=1}^n{w}_{\mathit{ij}}^2-{\left(\sum _{j=1}^n{w}_{\mathit{ij}}\right)}^2\right]}{n-1}}}$$ (1)

$$\begin{array}{c}\overline{X}=\frac{\sum _{j=1}^n{x}_j}{n}\end{array}$$ (2)

$$\begin{array}{c}\mathrm{S}=\sqrt{\frac{\sum _{j=1}^n{x}_j^2}{\mathrm{n}}-{\left(\overline{X}\right)}^2}\end{array}$$ (3)

where $x_j$is the attribute value of feature $\mathrm{j}$, $w_{\mathit{ij}}$ represents the spatial weight between feature $\mathrm{i}$ and feature$\mathrm{j}$, and $\mathrm{n}$ denotes the total number of features.

• 2)Centroid-Standard Deviational Ellipse

The study further assessed the spatial distribution characteristics of major subtypes of avian influenza in poultry using the Centroid-Standard Deviational Ellipse model. Centroid analysis provides an explanation for the spatial variation patterns of unevenly distributed attributes [87]. The model can analyze the mean center location of disease cases and assess the direction of epidemic spread. The standard deviational ellipse is employed to illustrate the distribution range and direction around the mean center of multivariate spatial data, helping to identify and interpret clustering tendencies and distribution patterns [88]. By applying the Centroid-SDE model, the spatial variation of major avian influenza subtypes in China can be examined through the movement trajectory of the centroid, revealing the spatiotemporal directional trends in the distribution of different avian influenza virus subtypes.

2.3.2. Driving mechanism analysis

• 1)Spatial sampling

To address the class imbalance caused by a high proportion of zero values in the avian influenza data, and to reduce spatial confounding near case locations, a spatially constrained sampling method was used to construct the dataset. First, grid cells with a dependent variable greater than 0 were defined as case samples, while those equal to 0 were defined as non-case samples. Next, a 10 km buffer was created around each case sample, and any non-case samples within these buffers were excluded. This step reduced spatial overlap and environmental similarity between cases and nearby background samples. Consequently, it minimized the impact of spatial proximity on model training while retaining a sufficient number of samples.

• 2)XGBoost model

In this study, the Extreme Gradient Boosting (XGBoost) algorithm was used to construct the avian influenza risk model. XGBoost is an ensemble learning method based on Gradient Boosting Decision Trees (GBDT). It performs well in classification and regression problems and is widely used in ecological risk and disease prediction studies [[89], [90]]. To avoid temporal leakage, the dataset was split based on independent time periods. Data from 2010 to 2015 were used for model training, and data from 2016 to 2017 were used for independent testing. Additionally, an independent dataset from 2018 to 2025 was used for external validation to evaluate the long-term spatiotemporal generalization capability of the model. Furthermore, due to the significant class imbalance in the avian influenza data, the scale_pos_weight parameter was adjusted to balance the positive and negative samples, thereby enhancing the model's ability to recognize the minority class. During training, an early stopping strategy was applied to prevent overfitting, and the optimal number of iterations was determined based on the performance on the test set. The objective function of the XGBoost model is as follows:

$$\begin{array}{c}\text{Objective}=\sum _{i=1}^{n}l\left(y_i\right),\left(\widehat{y_i}\right)+\sum _{k=1}^k\Omega \left(f_k\right)\end{array}$$ (4)

where $\mathrm{l}\left({\mathrm{y}}_{\mathrm{i}}\right),\left(\widehat{{\mathrm{y}}_{\mathrm{i}}}\right)$represents the loss function that measures the difference between the true value${\mathrm{y}}_{\mathrm{i}}$and the predicted value $\widehat{{\mathrm{y}}_{\mathrm{i}}}$, and $\mathrm{\Omega }\left({\mathrm{f}}_{\mathrm{k}}\right)$ is the regularization term that penalizes the complexity of the model to prevent overfitting. $\mathrm{K}$is the number of decision trees, and ${\mathrm{f}}_{\mathrm{k}}$ denotes the individual tree structure.

• 3)SHapley Additive exPlanations

Ensemble algorithms in machine learning models increase model complexity but often reduce interpretability. To address this, the study introduces the Shapley Additive Explanations (SHAP) method to interpret the results of the XGBoost model. SHAP explains black-box models based on Shapley values, calculating the contribution or importance of each feature to the prediction outcome, thereby helping to determine how feature variables influence the dependent variable [91]. To further identify risk thresholds, Generalized Additive Models (GAMs) were used to smoothly fit the SHAP dependence plots. The potential risk turning thresholds were then identified based on the intersection points where the SHAP values crossed zero.

$$\begin{array}{c}\mathrm{g}\left(x\right)={\Phi }_0+\sum _{i=1}^M{\Phi }_i{x}_i\end{array}$$ (5)

where $\mathrm{g}\left(x\right)$represents the explanation model, ${\Phi }_0$ is the model constant, and ${\Phi }_i$is the Shapley value of the i-th feature. When ${\Phi }_i>0$, the feature is considered to contribute positively to the prediction value. $x_i$ is a binary variable indicating whether the i-th feature is included in the model, where $x_i=1$ means the feature is selected, and $x_i=0$ means it is not included.

• 4)Model evaluation and validation

Model performance was comprehensively evaluated using multiple metrics, including ROC-AUC, PR-AUC, and specificity. Due to the potential class imbalance in the avian influenza data, PR-AUC was introduced as a supplementary metric because it is more sensitive to imbalanced datasets [92]. Meanwhile, specificity was utilized to control the false positive rate, ensuring high accuracy when the model excludes low-risk areas.

Next, to test the temporal generalization ability of the model, independent dataset from future years were used as a validation set. This step verified the robustness of the model during temporal extrapolation and assessed its ability to stably identify high-risk areas across different periods. The stability and generalization performance of the model across time scales were comprehensively evaluated by comparing its performance on the training set, the test set, and the external validation set. Considering that spatial epidemiological data often exhibit spatial clustering, Moran's I was used to analyze the spatial autocorrelation of the model residuals. If the residuals show a random distribution with no significant spatial autocorrelation, it indicates that the model has adequately explained the primary spatial structures in the distribution of avian influenza. This further enhances the reliability and scientific validity of the prediction results.

3. Results

3.1. Spatiotemporal distribution trends

From 2010 to 2025, there were a total of 2429 reported cases of avian influenza in poultry in China, with overall prevalence being unstable. Significant surges in poultry infection cases occurred in 2011, 2013, and 2014. The highest number of cases was reported in 2011, with a total of 599 cases, more than half of which occurred in March. During the 2010–2025 period, the predominant subtypes of avian influenza cases were H5, H7, and H9, with H5 subtype accounting for 509 cases, H7 subtype for 785 cases, and H9 subtype for 1135 cases. Among these, the H9N2 subtype had the highest number of cases, followed by H7N9 and H5N6 subtypes, which reported 784 and 362 cases, respectively. Additionally, cases of the H5N1 subtype were reported during most years of the study period. As shown in Fig. 1, the occurrence of avian influenza cases exhibited clear seasonal fluctuations, with substantial differences in case numbers among years. From 2010 to 2025, cases predominantly occurred between November and March, accounting for 65% of the total cases, with winter being the peak season. Notably, March had the highest total number of cases, comprising 24% of all reported cases. From an interannual perspective, case numbers remained relatively high during 2010–2017. After this period, the overall number of cases showed a declining trend, although local fluctuations were still observed in some years. Since 2018, the total number of cases has remained at a relatively low level, but seasonal peaks have persisted. Overall, the epidemic peak periods were generally consistent across years.

Fig. 1.

Temporal change of avian influenza cases in poultry in China (2010–2025). Poultry cases refer to domestic birds infected with avian influenza viruses. The reported hosts were mainly ducks and chickens, with a few cases detected in geese, pigeons, and quails. The identified viral subtypes included H9N2, H7N9, H5N1, H5N6, H5N2, H5N8, H5N3, and H7N2.

The visualization results of poultry case data indicate that there are more cases in eastern China (Fig. 2a). The southeastern region, particularly in provinces such as Jiangsu, Anhui, Zhejiang, Jiangxi, Hubei, and Guangdong, is a hotspot for outbreaks, with H5, H7, and H9 subtypes all present (Fig. 2b).

Fig. 2.

Distribution (2a) and hotspot (2b) of avian influenza cases and in China. (2b): Z-scores indicate significance at 99% (|Z| ≥ 2.58), 95% (|Z| ≥ 1.96), and 90% (|Z| ≥ 1.65); non-significant cells are shown as zero. A significantly positive Z-score indicates a high-value cluster at that location and its neighborhood, which is identified as a hotspot.

Among these, the H5 subtype exhibits a broader distribution range and a stronger tendency for prevalence, predominantly occurring in the western and southern regions of China, while cases in the northeastern areas are sporadic. In contrast, the H9 subtype is rarely reported in the western regions. Compared to H5 and H9, the H7 subtype cases are more concentrated in southeastern China, particularly in Hubei, Jiangxi, and Fujian provinces (Fig. 3a, 3b).

Fig. 3.

The spatial and temporal shifts in the distribution of the major avian influenza subtypes in China. (3a): Standard deviation ellipse illustrating the spatial distribution of subtypes. (3b): Proportional distribution of subtypes across regions. (3c): Trajectory of centroid shifts in subtype distribution. (3d): Trajectory of centroid shifts in the distribution of H5 subtypes. (3e): Trajectory of centroid shifts in the distribution of H7 subtypes. (3f): Trajectory of centroid shifts in the distribution of H9 subtypes.

Fig. 3c shows the migration trajectories of the spatial mean centers of different poultry avian influenza subtypes (H5, H7, and H9) in China during 2010–2025. Overall, the mean centers of all subtypes were mainly distributed in central and eastern China, but their migration directions and spatial ranges differed considerably. During 2010–2025, H5 subtype cases were reported almost every year. The epidemic mean center generally showed a diffusion trend from southwestern China toward central and eastern China (Fig. 3d). In 2011, the mean center was located near the eastern edge of the Qinghai-Tibet Plateau. It then gradually shifted toward central China and became mainly concentrated in south-central China during 2018–2025. Starting in 2013, H7 subtype cases began to emerge and continued until 2019. The mean centers were mainly distributed in eastern and northern China, showing an overall northward migration trend from southern China (Fig. 3e). From 2013 to 2017, the H7 mean center gradually moved toward the middle and lower reaches of the Yangtze River and the North China Plain, reaching its northernmost position around 2019.

In contrast, the H9 subtype showed a relatively smaller migration range. Its mean centers were mainly concentrated in southeastern coastal and southern China, with relatively stable spatial changes overall (Fig. 3f). During 2010–2014, the H9 mean center was mainly located along the eastern coastal region of China. After that, poultry H9 cases temporarily disappeared. Since 2018, sporadic H9 cases reappeared in southern China, and continued until 2024. In 2025, the number of cases increased slightly.

3.2. Driving factors analysis

3.2.1. Importance and nonlinear effects of driving factors

SHAP analysis revealed substantial differences in the contributions of different driving factors to avian influenza risk (Fig. 4). Overall, human activity-related variables showed the highest contribution to model predictions, accounting for 69.1% of the total importance. Among them, population density, road network density, GDP, and number of poultry trading points contributed 24.9%, 22.7%, 17.4%, and 4.1%, respectively, indicating that human activity intensity and transportation-trade networks play important roles in shaping the spatial risk of poultry avian influenza. Climate variables contributed approximately 14.4% overall, among which total precipitation from November to March was the most important climatic factor, contributing 9.1%. Livestock and poultry farming-related variables contributed about 10.8%, with pig farming density and chicken farming density showing relatively high importance. Landscape ecological variables accounted for 7.2% of the total contribution, and DEM and waterfowl species density showed noticeable effects on risk distribution (Fig. 4a).

Fig. 4.

(4a): Bar plot of mean absolute SHAP values, showing the average contribution of each factor to the model output, which indicates overall variable importance. (4b): Bee swarm plot of SHAP values, where each point represents one sample: the horizontal position indicates the SHAP value (positive values increase the prediction, negative values decrease it), and the color shows the original feature value. This plot reveals both the magnitude and direction of each factor's effect on the prediction. (4c): Partial dependence plots (PDPs) for the top factors, illustrating the marginal effect of each factor on the predicted response while holding other factors constant. The red dashed lines represent the transition thresholds where SHAP values crossed zero.

The SHAP beeswarm plots and dependence plots further revealed nonlinear relationships between the variables and avian influenza outbreaks (Fig. 4b, c). Overall, high values of human activity-related variables, including population density, road network density, GDP, and number of poultry trading points, were generally associated with higher positive SHAP values. The associations of these factors with poultry avian influenza exhibits a similar trend of rapid growth followed by stabilization, indicating that at early stage, enhanced transportation accessibility, population aggregation, and intensive economic activity increase the risk of poultry avian influenza significantly, while once exceeding a certain threshold, the impact of these factors on avian influenza risk plateaus at a constant level.

In contrast, climatic and ecological variables showed more complex nonlinear response patterns. For total precipitation from November to March, SHAP values gradually shifted from negative to positive when precipitation exceeded 1263 mm and the influence reached the maximum at 2743 mm, suggesting that humidity in general facilitates the formation of avian influenza risk. The number of days with temperature > 30 °C showed a clear bimodal relationship with avian influenza risk. When the number of high-temperature days was between 5 and 18 days, the explanatory power on avian influenza increase significantly. Another surge was between 25 and 39 days, and then followed by a gradual decline under extremely hot conditions. This pattern suggests that moderate high-temperature conditions may favor avian influenza risk, whereas prolonged extreme heat may suppress virus transmission. Similarly, increasing number of days with wind speed rate > 4 m/s was generally associated with increasing SHAP values.

Livestock farming-related variables showed similar pro-inverted U-shaped patterns. When the livestock farming density was low, increasing density significantly promoted risk elevation, and the influence reached to a peak and declines. When the livestock farming density reached to a threshold, 20,270 for pig, 254,177 for chicken and 108,573 for duck, higher densities were associated with a reduced risk of avian influenza.

For landscape ecological variables, increasing waterfowl species density was generally associated with increased SHAP values when the density reached 40 species, and the high-value regions mainly distributed within the positive SHAP range and reaching the maximum contribution at approximately 73 species. In contrast, increasing DEM was generally associated with decreasing SHAP values. When elevation was below 376 m, it is still positively contributing to the avian influenza. While after this threshold, DEM serves as a limiting factor for avian influenza transmission. Although wetland showed relatively low overall importance, regions with wetlands were still more likely to correspond to positive SHAP values, indicating a potential association between wetland environments and the spatial distribution of avian influenza.

3.2.2. Interactive effects among driving factors

To further explore potential nonlinear interaction effects between the host (waterbirds, livestock/poultry farming and trade) and other factors, two-dimensional SHAP interaction analyses were conducted. The results further demonstrated that avian influenza outbreaks were not driven by a single environmental factor, but were jointly influenced by farming activities, wild hosts, climatic conditions, and socioeconomic activities (Fig. 5).

Fig. 5.

Multivariable partial correlation dependency plot. The red dashed line indicates the transition boundary where the SHAP value equals zero.

First, significant interaction effects were observed between poultry farming activities and wild waterbirds. A clear nonlinear relationship was found between chicken farming density and waterfowl species density (Fig. 5a). Under high chicken farming density conditions, SHAP values increased substantially as waterfowl species density increased from low to medium-high levels, indicating that chicken farming regions with greater wild waterbird presence tended to exhibit higher avian influenza risk. In contrast, in areas with relatively low chicken farming density, the overall risk contribution remained limited even when waterfowl species density was high. A similar pattern was also observed between duck farming density and waterfowl species density (Fig. 5b). However, the highest-risk regions were mainly concentrated under the combination of moderate duck farming density and relatively high waterfowl species density, whereas the contribution gradually declined in extremely high-density duck farming areas.

Within livestock farming systems, clear interaction enhancement effects were also detected among different animal groups. When pig farming density was lower than approximately 400 heads/km², the overall SHAP contribution remained relatively low. However, when pig farming density further increased and spatially overlapped with regions of high chicken farming density, SHAP values increased rapidly and formed continuous high-risk areas (Fig. 5c). This pattern suggests that regions characterized by the coexistence of multiple high-density livestock and poultry farming systems may provide more complex ecological conditions for avian influenza transmission.

In addition, climatic conditions also showed clear moderating effects on avian influenza risk in poultry farming systems. A nonlinear interaction was observed between chicken farming density and the number of days with temperature > 30 °C (Fig. 5d). When chicken farming density exceeded approximately 20,000 animal/km² and the number of high-temperature days remained within a moderate range of 15–25 days, SHAP values reached the highest level. However, as the number of high-temperature days further increased, the risk contribution gradually declined.

Among the human activity-related factors, poultry trade networks showed strong risk amplification effects. Regions characterized by intensive chicken farming together with more trading points showed higher SHAP values, suggesting that trade activities may further enhance virus transmission risk within intensive farming systems (Fig. 5e). In addition, number of poultry trading points and waterfowl species density also showed an interactive amplification pattern (Fig. 5f). Although regions simultaneously characterized by high waterfowl species density and high trade connectivity were relatively limited, these areas still had high SHAP values.

The interaction between chicken farming density and GDP further indicated that socioeconomic conditions strongly moderated avian influenza risk (Fig. 5g). Under conditions of high chicken farming density, low-GDP regions corresponded to higher SHAP values, whereas the risk contribution gradually decreased with increasing GDP levels. This result implied that the risk effect associated with poultry farming is not constant, but jointly influenced by regional economic development and farming management conditions.

3.2.3. Model validation

The XGBoost model showed good predictive performance across the training, testing, and external validation datasets. In the training dataset, the ROC-AUC and PR-AUC values reached 0.97 and 0.95, respectively, indicating that the model effectively captured the complex nonlinear relationships between avian influenza occurrence and environmental variables. In the testing dataset, the ROC-AUC and PR-AUC values were 0.90 and 0.82, respectively, while the external validation dataset showed ROC-AUC and PR-AUC values of 0.90 and 0.83 (Fig. 6). The Specificity values for the training, testing, and external validation datasets were 0.97, 0.90, and 0.89, respectively, indicating relatively high classification accuracy. Overall, the model demonstrated good temporal generalization and long-term transferability, and was able to consistently capture the relationships between avian influenza occurrence and major risk factors across different periods.

Fig. 6.

ROC and precision-recall curves of the XGBoost model across training, testing, and external validation datasets.

To further examine whether residual spatial autocorrelation remained after model fitting, Moran's I analysis was conducted on model residuals. The results showed that Moran's I was close to the theoretical expectation under spatial randomness (Moran's I = −0.0024, p = 0.50), indicating that residuals were randomly distributed in space and no significant spatial autocorrelation remained. This suggests that the model adequately captured the major spatial structure underlying the distribution of avian influenza risk.

4. Discussion

This study systematically analyzed the temporal patterns and hotspot regions of avian influenza cases among poultry in China from 2010 to 2025. The research specifically focused on the spatiotemporal distribution trends of three major subtypes, H5, H7, and H9, and delved into the driving factors associated with the spread of avian influenza.

Our findings indicate that avian influenza cases predominantly occur in the winter months from November to March [62], with the southeastern region of China identified as a hotspot. From 2010 to 2025, the H9N2 subtype accounted for the largest number of cases, and its epidemic mean center remained mainly concentrated in southern and southeastern coastal China. The H5 subtype persisted in the southern areas, which is largely consistent with previous research findings [13]. We observed that, in addition to the persistent circulation of the H5 subtype, H9 subtype temporarily declined around 2014, coinciding with the emergence and increased circulation of the H7 subtype. After 2018, H9 subtype cases reappeared, whereas H7 subtype cases gradually declined and eventually disappeared. Similarly, Bi et al. reported that after 2016, H9N2 became the predominant avian influenza virus subtype circulating in poultry in China, while the prevalence of H7N9 and H5N6 gradually decreased [93]. The near disappearance of the H7 subtype after 2017 was likely closely associated with the nationwide implementation of the H7 avian influenza vaccination strategy in poultry systems. Unlike many countries in Europe and North America that primarily control highly pathogenic avian influenza through culling infected and suspected poultry populations, China has adopted a large-scale mandatory vaccination strategy for HPAI prevention and control, for H5 subtype since 2005 and for H7 subtype since 2017 [94]. Previous studies also demonstrated that the introduction of the H5/H7 bivalent vaccine substantially reduced H7N9 circulation in poultry and nearly eliminated human H7N9 infections in China [95]. Moreover, recently updated trivalent H5/H7 vaccines have shown effective antigenic matching and complete protective efficacy against emerging 2.3.4.4b H5N1 strains in chickens [94]. However, despite the effectiveness of vaccination, H5 subtype viruses continue to evolve rapidly under complex ecological and immunological pressures [[96], [97]]. Continuous genetic evolution and antigenic drift of H5 viruses may reduce vaccine matching over time, highlighting the necessity for regular vaccine updates and long-term surveillance to prevent potential re-emergence and spread of novel variants [[98], [99], [100]].

The spatial risk of avian influenza was jointly influenced by climatic conditions, livestock and poultry farming, wild hosts, and trade networks. Human activity-related variables, including population density, road network density, and GDP, had the highest contributions in spatial risks of poultry avian influenza. Climate factors ranked the second, with precipitation contributing the most. Livestock and poultry farming density ranked the third, and landscape ecological factors ranked the forth. One thing needed to refer was, although waterfowl species density, chicken and duck farming density and number of days with temperature > 30 °C showed relatively lower individual contributions, they exhibited significant moderating effects in the interaction analysis.

Population density showed the highest contribution, indicating that densely populated areas may be associated with stronger poultry consumption, higher market demand, and more intensive live poultry circulation activities, thereby increasing opportunities for virus transmission [66]. Road network density was the second influential predictors. Transportation networks can represent the intensity of live poultry transportation, human mobility, and market connectivity [[67], [68]]. Previous studies have shown that poultry transportation networks are one of the major pathways for the interregional spread of avian influenza, and highly connected transportation systems may increase the likelihood of virus transmission between regions [101]. Although poultry trading point contributing less, it acted as a key interactive variable, the interaction between poultry farming density and waterbird density was associated with higher disease risk. This mainly because poultry trade acts as a bridge for avian influenza transmission between migratory birds and domestic breeding populations [102]. All of the four human activity factors showed similar influencing curves on avian influenza risks, sharp increase when the human activity intensity was low and then remained stable. Regions with lower levels of economic activities were generally more dependent on traditional free-range or small- to medium-scale farming systems, and often show relatively weaker biosecurity facilities, surveillance systems, and farming standardization, and thus provide more favorable conditions for sustained virus circulation and spread [[71], [72], [103]].

The livestock and poultry farming density mainly positively influence the avian influenza risk. In particular, low to medium size of farming density were related to sharply increase in avian influenza risk, and the influence declined but still positive when the density reach to a threshold. While when the magnitude of the farming size reach to a certain scale, the association became negative, this could related to the high-level biosecurity and environmental management, and high herd immunity and vaccine coverage in large-scale breeding area [104].

For climatic and ecological variables, precipitation and wind speed in general could increase the risk of avian influenza. Because most of the avian cases were in south China and in autumn and winter when the migratory wild birds inhabit in the south. For one thing, the moist surfaces and rainwater-contaminated poultry drinking water after precipitation could create favorable conditions for virus spread among the poultry system and from wild birds to poultry [51], [52]. For another thing, winter rainfall boosts migratory bird survival by providing abundant food resources in south China [105]. Wind is also positively contributing the avian influenza risk through the enhanced migration of airborne influenza viruses [56].

When the waterfowl species density was low, it acted as a limiting factor but no obvious influence trend was observed. When the species density reached to 40, a clear increasing trend can be found, and the influence became positive afterwards. This was credible because the most of the waterbirds are susceptible to avian influenza viruses [[10], [106]], when the diversity increased, the susceptible host increased in general. This conclusion contrary to the dilution effect has been confirmed in other bird related disease, such as West Nile virus, Leucocytozoon and also avian influenza [[107], [108]].

The clear interaction enhancement effect between chicken/duck farming density and waterfowl species density indicated that, under conditions of high chicken farming density, the risk contribution increased substantially with increasing waterfowl species density. This pattern suggests that virus introduction risk is more likely to be amplified when wild waterbirds spatially overlap with intensive poultry farming activities. This result is generally consistent with the concept of the wild bird-poultry transmission interface [[109], [110]], and further indicates that wetland edges, lakeside regions, and areas with concentrated waterfowl farming may serve as important spatial nodes for cross-host virus transmission [[111], [112]]. In addition, high-risk regions were mainly concentrated under the combination of moderate duck farming density and relatively high waterfowl species density, whereas the contribution did not further increase in extremely high-density duck farming areas. This pattern indicated that, compared with highly intensive farming systems, medium-scale or semi-open waterfowl farming areas are more likely to contact natural wetland environments, thereby increasing opportunities for virus exchange between wild birds and domestic ducks [[63], [103], [113]].

The coexistence of multiple livestock and poultry farming systems may further enhance transmission risk. The results showed that SHAP values increased rapidly and formed continuous high-risk regions when both pig farming density and chicken farming density reached high levels simultaneously. Although the current study cannot directly demonstrate the involvement of pigs in virus reassortment or cross-host transmission, from a spatial epidemiology perspective, regions characterized by the coexistence of multiple high-density livestock and poultry farming systems may reflect more intensive agricultural production activities, more complex host structures, and more frequent trade activities, thereby creating environmental conditions that are more favorable for virus circulation and transmission [[39], [114], [115], [116]].

This study was the first to conduct a detailed analysis of the changes in the major subtypes of avian influenza in poultry in China, reveal the spatiotemporal distribution patterns of these subtypes and explore the interactive influencing features combining human activity, climatic factor, livestock/poultry and landscape ecological factors. This study has certain limitations. The interaction results in this study reflect nonlinear associations learned by the model rather than strict causal relationships. In addition, the virus infectivity is different among genome types and subtypes, and due to the data limitation, we could not include these in the current study. Moreover, data bias and the heterogeneity caused by the local monitoring capacity and the reform of the surveillance system could influence the accuracy of the model, and we did not take this into consideration.

5. Conclusion

From 2010 to 2025, poultry avian influenza cases in China mainly occurred during the winter season from November to March. The epidemic mean centers of different subtypes were generally concentrated in central and eastern China and showed clear patterns of temporal migration and changes in subtype distribution over time. Among them, the H5 subtype gradually shifted toward central China, while the H7 subtype showed a northward migration trend. In contrast, the H9 subtype remained mainly concentrated in southern and southeastern coastal China, and the H9N2 subtype accounted for the largest number of cases. The driving factor analysis indicated that the spatial risk of avian influenza was jointly shaped by human activities, poultry farming systems, climatic conditions, and wild host distributions. Variables related to human activities showed the strongest contributions, highlighting the important roles of transportation connectivity and socioeconomic activities in virus transmission. Avian influenza risk was further amplified under the combined effects of livestock and poultry farming systems, high waterfowl species density, and strong trade connectivity, suggesting that the wild bird-poultry-trade network interface may represent a key pathway for avian influenza virus transmission in poultry. Overall, the outbreak and spread of avian influenza represent a complex socio-ecological process shaped by the coupling of natural ecological conditions and human activity networks.

CRediT authorship contribution statement

Feifei Li: Writing – original draft, Methodology, Conceptualization. Linsheng Yang: Writing – review & editing, Conceptualization. Hairong Li: Writing – review & editing, Conceptualization. Li Wang: Writing – original draft, Methodology, Conceptualization. Lijuan Gu: Data curation. Svetlana Malkhazova: Data curation.

Declaration of competing interest

The authors declare the following financial interests/personal relationships which may be considered as potential competing interests:

Data availability

Data will be made available on request.