Vector-borne diseases (VBDs), a category of zoonosis, are transmitted to humans and animals (primarily ruminants) through vectors including mosquitoes, ticks, and fleas. Recently, VBDs have emerged as a significant threat to human health in temperate areas (1). In Eastern, Western and Southern Europe, including Italy, the West Nile virus (WNV) is the most widespread mosquito-borne zoonosis (2). In its transmission cycle, birds act as primary and amplifying hosts, with mosquitoes transmitting the virus to other birds; humans and other mammals are considered dead-end hosts, as they generally do not contribute to virus transmission, often remaining asymptomatic (3).

In Europe, the Culex pipiens [from now on Cx. pipiens, (4)] is recognized as the main WNV vector (5, 6). The presence of Cx. pipiens in the world’s temperate climatic regions and its ability to transmit zoonotic pathogens, other than WNV, such as Rift Valley fever (7), Usutu and Japanese encephalitis (8), made Cx. pipiens one of the most important mosquito species regarding public health. Culex pipiens is a species able to feed on birds and mammals, including humans (9, 10), and it breeds in a wide variety of environments, either in rural and urban cycles, tolerating also human-altered ones (11).

This ability to adapt to a wide range of environments rely on its characteristics to tolerate a number of conditions, including the adaptability to many hosts for feeding, the flexibility to laying eggs in presence of water ponds or humid soil, and the environmental tolerance for resting and breeding (11). This complexity contributes to the species characteristics in terms of trophic behavior and vectorial capacities. A fine understanding of the Cx. pipiens’ habitat suitability that facilitates survival, reproduction and dispersal becomes of paramount importance for determining the risk of local establishment, persistence and spread, developing efficient (i.e., species-, place- and time specific) vector monitoring.

In Italy, Cx. pipiens is the species most frequently found to carry WN virus (5, 12). Across the country, Cx. pipiens population is characterized by seasonal dynamics across the year, with its maximum population abundance during summer (13, 14). The intensity and inter-annual variations of vector populations have been frequently associated to temperature, rainfall, humidity, vegetation, i.e., to climatic-environmental drivers. Temperature has been most frequently reported as the most influential variable affecting mosquito population dynamics (15–17). Temperature drives the vector competence, by accelerating the virus replication within the insects, prolonging their breeding season (18, 19), increasing mosquito abundance (20) and their infection rate (21). Rainfall has been found to play a significant role in many studies (19, 22), although its impact remains a topic still controversial in literature (23). On one hand, rainfall creates pools of water, which serve as suitable breeding sites for mosquitoes, thereby increasing species abundance. On the other hand, excessive rainfall can damage larval habitats flushing aquatic environments. Additionally, air humidity in preceding months is another factor associated with the abundance of mosquitoes (13). The length of daylight was another abiotic factor associated with Cx. pipiens population growth in Italy (22, 23). Vegetation and vegetation indices are other parameters correlated with the vectors’ behavior and their biological cycle (23, 24), although not always resulted relevant (19): the presence and density of green biomass provide sugar feeding supplies for adult mosquitoes, potential resting and protection from climatic conditions. A combination of these (and other) drivers have been used to classify Italian territory into different ecoregions (25).

In species distribution modeling studies aiming to associate spatial environmental characteristics with the presence/abundance of vectors, the Random Forest machine learning method is probably the most commonly used approach (26–28) among others, alongside with MaxEnt (29, 30). For an in-depth description of Machine Learning (ML) based species distribution modeling, a comprehensive review is available in Zhang and Li (31). In another research, three supervised learning models, k-nearest neighbor (kNN), artificial neural network (ANN), and support vector machine (SVM) were used to predict mosquito abundance based on socioeconomic and landscape patterns (32). Other authors used ML methods to predict West Nile virus outbreaks or WNV infection rates in Culex mosquitoes with eco-climatic drivers (33, 34).

The most common statistical methodologies focused on epidemiological modeling of main VBDs in Europe rely on pre-calculated indices and factors, such as NDVI (Normalized Difference Vegetation Index), presence of standing water indicated by NDWI (Normalized Difference Water Index) or soil moisture levels (Moisture Index) (24, 35). Those indices are well from derived from Earth Observation (EO) data: EO data has played a crucial role in the study of VBDs, particularly in the realm of epidemiological modeling and in understanding the environmental factors influencing disease transmission dynamics. EO data, collected by sensors onboard satellites, provides data about Earth’s surface across a number of wavelengths. This data provides measurements of surface temperature, chlorophyll presence, water presence, soil characteristics, land cover, among others relevant features of Earth surface supporting mosquito populations life cycle, hence possibility of VBDs transmission (36). The frequency, consistency and regularity of data acquisition generates continuous datasets able to depict environmental features crucial for understanding mosquito habitats, either as larval sites, breeding or resting places. By analyzing satellite data alongside epidemiological information on diseases, environmental risk factors associated with VBDs can be identified, assisting in pinpointing areas with high vector concentrations or conducive environmental conditions for disease transmission (24). In turn, these results enhance understanding of disease transmission mechanisms and interactions between the environment and human health, thereby facilitating the development of more effective prevention and control strategies.

To date, new large EO datasets are available at unprecedented spatial, spectral and temporal resolutions, as those produced by the European Copernicus program. The satellite data of the Sentinel-2 mission, in particular, are dedicated to land and vegetation monitoring, and through the Multi-Spectral Instrument (MSI) carried onboard, they return an optical multispectral “photograph” of the territory they fly over every 5 days. The Sentinel-2 constellation comprises Sentinel-2A (in orbit since June 23, 2015) and Sentinel-2B (in orbit since March 7, 2017) satellites, orbiting the Earth simultaneously on the same sun-synchronous orbit, offset by 180 degrees. The MSI sensors acquire the light reflected or emitted from the planet’s surface in 13 spectral bands and at 10, 20- or 60- meters spatial resolutions (37). After 6 years of regular complete acquisitions, these datasets offer new opportunities to understand the landscape in which host and vector proliferate and interact. The micro-scale level at which it is now possible to study the presence of mosquitoes is relevant for both larval and adult stadium, considering the relatively short flight distance (usually a few hundred metres) during its adult lifespan (38).

This substantial volume of data calls for additional methods of analysis and deep learning, exploiting patterns and dependencies in the provided raw data to extract information, without relying on specific a priori hypothesis (32). Deep neural network (DNN) architecture is composed of neurons, synapses, weights, biases, and functions, coarsely mimicking the functioning of the human brain: DNN uses multiple layers (intermediate understanding) to progressively succeed in its task, i.e., extract higher-level features from the raw input (39). Convolutional neural networks (CNNs) are a type of deep neural network designed primarily for pattern recognition in visual data. They are widely used in computer vision applications, such as image recognition and object detection (40): in image processing, lower layers may identify edges, while higher layers may identify more complex and abstract features (e.g., digits or letters or faces). The hidden layer learns features from the input images (by means of a set of parameters), and it is subject to a function (non-linear activation) which reduces computational complexity while retaining multi-scale information. The sequence of these layers and functions allows the modeling to learn increasingly complex and abstract features at different scales of the original image (39). During the DNN training phase, the parameters (weights, biases) are iteratively adjusted so that the output layers of the DNN best approximate the ground-truth target (41).

The application of deep learning to high spatial resolution data to predict the potential distribution of mosquito species (42) or to predict outbreaks of VBDs (43), is in its nascent stages (44). Few studies deal with animal, disease and plant distribution modeling through DNN (45–47).

In this paper, we present the combination of deep learning methods and high spatial resolution (20 m) remotely sensed imagery with naive spectral bands applied to the distribution of presence/absence Cx. pipiens species occurrences in central Italy (whose territory covers a wide range of environmental conditions, i.e., wide domain of eco-climatic values). Aims of this paper were:

To provide pictures (maps) of the spatio-temporal distribution of Cx. pipiens abundance in central Italy (Abruzzo and Molise regions), as derived from a 4 years in-field sampling.

This study demonstrates successful synergy among entomological field activities, new high-resolution satellite imagery (Sentinel-2) and advanced algorithms (AI Deep Learning) in predicting the presence/absence of Cx. pipiens in central Italy. The algorithms were trained using a dataset that combines 4 years of mosquito field collection data (2019–2022) with environmental information from multispectral bands at 20 m spatial resolution in the surroundings (4480 × 4480m) of each trap location.

Over 4 years of entomological activity, in Abruzzo and Molise regions, Cx. pipiens was found in every province of the study area, with different environmental conditions across its territory (25). This result is not surprising as Cx. pipiens is able to adapt to a wide variety of habitats and this species has already been reported as one of (if not the) most abundant mosquito species in Italy (5). Consistently with other studies (65), the fluctuations in the abundance of Cx. pipiens populations among different sites, are likely associated with the diversity of habitats and climatic/environmental conditions. Other factors, such as the presence/absence of suitable hosts (i.e., birds, horses, humans), not included in this study, may also impact mosquito abundance. Between 2019 and 2022, the yearly abundance of Cx. pipiens at each site has generally remained stable, showing a higher abundance along the coastal area (Figure 5). Fluctuations in mosquito abundance across different years could be attributed to variation in climatic factors (i.e., temperatures and rainfall). Additionally, small changes in the configuration of mosquito breeding habitats, induced by various atmospheric events, i.e., a shift in a riverbed course due to heavy precipitations, may also contribute to these fluctuations. The seasonal pattern of mosquito abundance revealed an increase during the spring months, reaching its peak in late summer, and subsequently declining in the autumn months. This population trend aligns with the typical seasonal pattern observed in the Mediterranean Basin, influenced by temperatures and photoperiod, which concurrently impact adult mosquito activity (14, 21, 27, 66).

Even though Cx. pipiens is widely distributed at our latitudes, its life cycle and the population seasonality are primarily influenced by local environmental conditions. The presence of local water pools is essential for the egg stage, and nearby suitable landscape features play a crucial role in supporting the adult stage, facilitating breeding, resting and survival (58). To fully capture these environmental prerequisites, our study employed two levels of detail in mapping areas suitable for Cx. pipiens: firstly, a fine-scale analysis of environmental conditions, as it accurately differentiates and map the mosquito-required environments for each life stage (67). Secondly, a broader-scale analysis of landscape variability near the sites was investigated, as it provides insights into the patchy territory favouring the mosquito proliferation (58, 68, 69). This study utilized a spatial resolution of 20 metres as reference, aiming to achieve a detailed representation of the landscape surrounding the entomological sites. Pixels of 20 × 20 meters were used to finely discretize the area around the sites, allowing for accurate identification of key landscape features such as presence of water, build-up areas, vegetation, grassland, etc., essential for sustaining the vector’s life stages. The satellite images subject to deep analyses were composed of 224 × 224 of those pixels, encompassing an area of 4480 × 4480 meters around each trap site, including its biodiversity composition and configuration. Each pixel in the satellite images represented the spectral reflectance for a specific date and wavelength acquisition.

The deep convolutional neural network models were provided with information on radiance wavelengths measured from satellites, without explicit land use or land cover classification, nor were pre-calculated indices, like NDVI, employed. The rationale behind this approach was to allow the algorithms to autonomously discern any relevant correlations, thereby surpassing any potential constraints of our existing knowledge. The Sentinel-2 spectral bands provide a comprehensive “picture” of the environment around the site, spanning a broader optical range than our traditional knowledge, from visible (B02—blue, B03—green, B04 red) to near infrared (B05, B06, B07, B8A for vegetation red-edge, WVP-B09 for Water Vapor), and short-wave infrared (B11, B12 snow-ice-cloud discrimination) wavelength of the electromagnetic spectrum. The multitude of these bands represents a more comprehensive set of information than the bands traditionally considered in such studies, as NDVI, Moisture Index, or Water Index, which rely on B04 and B08, B8A and B11, B03 and B08, respectively. In addition, convolutional neural networks, effectively extract meaningful features from images using a sequence of convolutional kernels and hierarchical layers (39): this approach enables the learning of local features (a group of pixels depicting a characteristic in the image), thus capturing important information in part of the image; CNNs also capture global patterns across the image (39, 40). The combination of local and global feature representation is crucial for capturing fine-scale details and broader-scale spatial conditions, making it a key component in successfully classifying mosquito habitat.

Three scenarios were tested in this study: the baseline model which considers environmental conditions at the time of mosquito collection; the multitemporal model, which considers conditions history up to 2 months prior to collection; and the MAGAT model which considers the relationship with nearby geographical sites and temperatures. The multitemporal model demonstrated the best performance, highlighting the importance of considering the temporal evolution of environmental conditions. The Sentinel-2 bands inherently contain some information about temperature variation: temperatures were included in the MAGAT model to identify similar sites, but the multitemporal model still achieved better performance. This suggests that the information contained in the multitemporal sequence of Sentinel-2 bands encompassed also the variations typically associated with temperature. The high frequency of Sentinel-2 overpasses at our latitudes is beneficial, providing updated images of the territory every 5 days. This temporal resolution is sufficient to detect and monitor the evolution of the vegetation growth, the phenological stages of crops in cultivated fields, and the impact of rainfall on vegetation and soils. These patterns are crucial for characterising the environment most suitable to Cx. pipiens presence and for monitoring population trends across seasons, assisting in optimal trap placement. In addition, 5 days interval align well with the temporal resolution of field activities. A finer resolution (less than 5 days) might offer more information, but this would approach the level of predicting daily mosquito population dynamics, which is beyond our study’s scope. The frequency of Sentinel-2 helps mitigate issues with cloudiness, a major drawback of optical measurements that rely on sunlight to collect information about the Earth’s surface. Frequent acquisitions increase the probability to have clear, cloud-free images.

The models were also tested against various time lags, meaning we considered eco-climatic conditions occurring a number of days before the collection. This “empty time” is valuable when the model is used as a base for operational predictive tools in support of surveillance activities. In both the baseline and MAGAT models, higher performances were reached by shorter temporal lags, indicating that environmental conditions closer to the collection time were the best predictors of mosquito presence/absence. On the other hand, the multitemporal model showed that the environmental conditions related to mosquito presence/absence were those occurring during the 2 months before the collection date, going back to 4 months (lag of 50 days): when a wider period of conditions were taken into account, the performance improved as the model “learned” the pattern of environmental conditions favouring the presence/absence of mosquito.

The multitemporal model utilized a sequence of 10 Sentinel-2 acquisitions: the long-term pattern identified with these 10 images provided the model with rich spectral information able to identify environmental changes that influence the presence/absence of Cx. pipiens. A similar result was obtained by Vincenzi et al. (42) for another mosquito species (Culicoides imicola) in Italy. Besides environmental data from multispectral images, the MAGAT model, which employs a graph architecture, considered also temperature and closeness to other sites to predict mosquito presence. However, its performances were closer to baseline metrics. The multitemporal model demonstrated high performances, capable of identifying the seasonal pattern of Cx. pipiens population: this result suggests that the added value of multitemporal analysis is greater than that of incorporating graphs and other predictors.

The use of Sentinel-2 data proposed in this study offers also the advantage that Copernicus represents a recently launched constellation of satellites, with missions planned to extend over the next several years. The Copernicus Sentinel-2 data, which comprise a substantial volume of information, holds significant potential for exploitation through deep learning techniques. On the other hand, the deep learning modeling adopted allows to replicate the models in other and wider geographical areas (with the same range of environmental conditions), for other vectors (relying on Sentinel-2 information), and for specific time-steps (at the beginning of the season) only, incorporating them into an operational tool.

Still, some drawbacks have to be highlighted: deep learning models require a huge amount of data for training and testing, which may not always be feasible when ground-truth data are of entomological nature, requiring in-field efforts for collection. However, this issue was considered in this study through the exploitation of pre-training strategies. Another critical issue is the requirement for high powerful supercomputers, along with the management of Sentinel-2 data, which demands advanced software and hardware resources (download, heavy files, etc.) and processing time. Additionally, the biological interpretation of the model presents another challenge: while the models can deeply understand the problem though neurons and layers and accurately predict classifications, the underlying biological relationships are not explicitly revealed. The interpretation of feature layers is not straightforward, so the deep learning methodologies could result as “black-box” outputs. This hampers the reliance on this kind of approaches from a biological point of view (70). Addressing the challenges of interpretability remains an active area of research, with advancements being crucial for the broader adoption of deep learning models in real-world applications. Overcoming these challenges required a multidisciplinary team effort in our research, bringing together different skills so to address issues like collecting ground truth data, processing biological samples, handling large amounts of data, and performing complex computations.

The findings of our study offer concrete support for the conceptualization and deployment of locally tailored entomological interventions, providing useful information for optimizing surveillance activities in the following seasons, enabling precise timing and location for trapping activities. This could be useful to support prevention and reduction of diseases transmitted by Cx. pipiens. The positive outcomes of our research pave the way for future actions, which include refining models with additional years of field data, integrating various surveillance data to improve generalization, extending the infrastructure to other diseases within the veterinary domain, investigating high-level biological features, and translating findings into an operational predictive tool for surveillance activities.

The need to optimise human and economic resources in conducting surveillance activities for vector-borne diseases necessitates efforts to identify areas at risk both spatially and temporally. The widespread availability of Earth Observation data related to vectors and diseases, coupled with the growing ability to apply sophisticated analysis and algorithms, enables the development of models and tools that facilitate and optimise the identification of such areas. This approach is based on the integration of diverse disciplines, expertise and data, leading to the achievement of the expected results. Specifically, in this research, the use of data from the Copernicus program at 20 m spatial resolution, along with deep learning models, have made it possible to solve a classification task in which entomological surveillance collections are predicted as positive or negative for Cx. pipiens, the main vector in Italy for West Nile virus, and the prediction is made with 15 days in advance. The model’s performances were satisfactory (F1 score was higher than 75.8% for any temporal lag in the baseline model; F1 score reached 81.4% in the multitemporal model and 80.9% in the MAGAT); results derived from this study will advance our ability to identify suitable times and areas for Cx. pipiens presence and high-risk exposure to VBDs within Italian landscapes.

The methodology employed here can be expanded to the national territory and to other vectors, supporting the Ministry of Health in developing strategies for the surveillance and control of the vectors and the diseases they transmit.

The data analyzed in this study is subject to the following licenses/restrictions: entomological data in Italy are not publicly available due to legal and security concerns. Requests to access these datasets should be directed to m.goffredo@izs.it.

The manuscript presents research on animals that do not require ethical approval for their study.

CI: Visualization, Software, Methodology, Data curation, Conceptualization, Writing – review & editing, Writing – original draft. LB: Visualization, Software, Methodology, Formal analysis, Writing – review & editing, Writing – original draft. MdeA: Resources, Data curation, Writing – review & editing, Writing – original draft. ST: Resources, Data curation, Visualization, Software, Formal analysis, Writing – review & editing, Writing – original draft. AdiL: Resources, Software, Formal analysis, Data curation, Writing – review & editing, Writing – original draft. Sd’A: Resources, Data curation, Writing – review & editing, Writing – original draft. AP: Software, Investigation, Formal analysis, Methodology, Writing – review & editing. AB: Visualization, Writing – review & editing. DC: Software, Data curation, Writing – review & editing. MG: Validation, Resources, Funding acquisition, Conceptualization, Writing – review & editing. SC: Methodology, Conceptualization, Writing – review & editing. AC: Validation, Methodology, Funding acquisition, Conceptualization, Writing – review & editing, Writing – original draft.