Added-value of mosquito vector breeding sites from street view images in the risk mapping of dengue incidence in Thailand

Author summary The primary dengue mosquito vectors breed in containers with sufficient water and nutrition. Outdoor containers can be detected from geotagged images using state-of-the-art deep learning methods. Eight breeding site container types in Google street view images are detected using convolutional neural networks. We investigate the added value of including container information from geotagged images in predicting dengue risk. To explain the target variable dengue incidence, weather variables are added to complement the container variable predictors. Linear mixed-effects models are built to account for the effects of spatial and seasonal variation in weather and container variables on the dengue incidence. Evaluation is carried out over three provinces in Thailand: Bangkok, Nakhon Si Thammarat, and Krabi in comparison with classic linear models as well as the mixed effect models without container information. The proposed model with the container information outperforms both baseline models in all three provinces. We further perform sensitivity analysis to investigate the sensitivity of dengue incidence to the changes in the number of containers as well as the improvement in the model performance. This is the first work on dengue risk prediction models using container density information from geotagged images analysis. Dengue is an emerging vector-borne viral disease across the world. The primary dengue mosquito vectors breed in containers with sufficient water and nutrition. Outdoor containers can be detected from geotagged images using state-of-the-art deep learning methods. In this study, we utilize such container information from street view images in developing a risk mapping model and determine the added value of including container information in predicting dengue risk. We developed seasonal-spatial models in which the target variable dengue incidence was explained using weather and container variable predictors. Linear mixed models with fixed and random effects are employed in our models to account for different characteristics of containers and weather variables. Using data from three provinces of Thailand between 2015 and 2018, the models are developed at the sub-district level resolution to facilitate the development of effective targeted intervention strategies. The performance of the models is evaluated with two baseline models: a classic linear model and a linear mixed model without container information. The performance evaluated with the correlation coefficients, R-squared, and AIC shows the proposed model with the container information outperforms both baseline models in all three provinces. Through sensitivity analysis, we investigate the containers that have a high impact on dengue risk. Our findings indicate that outdoor containers identified from street view images can be a useful data source in building effective dengue risk models and that the resulting models have potential in helping to target container elimination interventions.