Abstract
The West Nile Virus (WNV) infection represents one of the most common mosquito-borne viral zoonosis. Its circulation is highly influenced by climatic and environmental variables influencing vector proliferation and virus replication (Paz & Semenza, 2013). Several statistical models have been developed to predict WNV circulation. Beyond this, the recent massive availability of Earth Observation data, coupled with the continuous advances in the field of Artificial Intelligence, offers new valuable opportunities in modelling animal diseases or vector population dynamics (Porrello et al., 2019).
In the work presented herein, we seek to forecast WNV circulation by supplying Deep Neural Networks (DNN) with high-resolution multi-band Copernicus Sentinel-2 and Landsat-8 imagery. While aggregated indices (e.g. NDVI) have been extensively shown to hold environmental and climatic features, here we let the DNN learn directly from raw spectral bands.
While previous DNN approaches analyze each geographical site independently (Porrello et al., 2019), we propose a spatial-aware approach that includes the characteristics of close sites. Specifically, we build upon Graph Neural Networks to aggregate features from neighbouring locations, while considering their mutual environmental relations. In particular, for each site, we measure their difference in Land Surface Temperature and Surface Soil Moisture - acquired from Copernicus Sentinel-3 and Sentinel-1 respectively - as well as their geographical (Haversine) distance. Moreover, we take into account the temporal factors that characterize the spread of the virus by injecting time-related information directly into the model.
To prove the merits of our proposal, we design an experimental setting that combines satellite imagery with “on-the-ground observations”, i.e. the veterinary cases detected in Italy in the years 2017-2019 and reported in the National Disease Notification System of the Italian Ministry of Health, as described in Candeloro et al., 2020.
By means of both qualitative and quantitative results, we show that our proposed Multi-Adjacency Graph Attention Network (MAGAT) consistently leads to higher performance than classical Machine Learning and Deep Learning baselines, especially when paired with an appropriate pre-training stage.
References
Paz, S., & Semenza, J. C. (2013). Environmental Drivers of West Nile Fever Epidemiology in Europe and Western Asia—A Review. International Journal of Environmental Research and Public Health, 10(8), 3543–3562. https://doi.org/10.3390/ijerph10083543
Porrello, A., Vincenzi, S., Buzzega, P., Calderara, S., Conte, A., Ippoliti, C., Candeloro, L., Di Lorenzo, A., & Capobianco Dondona, A. (2019). Spotting Insects from Satellites: Modeling the Presence of Culicoides Imicola Through Deep CNNs. 2019 15th International Conference on Signal-Image Technology Internet-Based Systems (SITIS), 159–166. https://doi.org/10.1109/SITIS.2019.00036
Candeloro, L., Ippoliti, C., Iapaolo, F., Monaco, F., Morelli, D., Cuccu, R., Fronte, P., Calderara, S., Vincenzi, S., Porrello, A., D’Alterio, N., Calistri, P., & Conte, A. (2020). Predicting WNV Circulation in Italy Using Earth Observation Data and Extreme Gradient Boosting Model. Remote Sensing, 12(18), Art. 18. https://doi.org/10.3390/rs12183064