A computer program capable of identifying, from aerial images, water reservoirs on roofs or slabs and outdoor swimming pools has been developed by Brazilian researchers with the help of artificial intelligence. The proposal is to use this type of image as an indicator of areas particularly vulnerable to mosquito infestation. Aedes aegyptiwhich transmits diseases such as dengue fever, zika and chikungunya. Moreover, the strategy appears as a potential alternative for a dynamic socio-economic mapping of cities — a gain for various public policies.
The research, supported by Fapesp, was carried out by professionals from USP (University of São Paulo), UFMG (Federal University of Minas Gerais) and Sucen (Superintendence for the Control of Endemic Diseases) from the Department of the health of the State of São Paulo. The results were published in the journal OLP A.
“What we did at that first moment was create a model based on aerial imagery and computing to detect water reservoirs and pools, and use them as a socio-economic indicator,” explains the engineer Francisco Chiaravalloti Neto, professor in the Department of Epidemiology, Faculty of Public Health, USP.
In the published article, he and his colleagues point out that previous surveys have already shown that poor areas of municipalities are often more prone to dengue fever. That is, the use of a relatively dynamic socio-economic status update model – especially compared to the census, which is carried out every ten years and subject to delays – would help to prioritize prevention efforts against dengue, zika and chikungunya.
“This is one of the first stages of a larger project”, emphasizes Chiaravalloti Neto. Among other things, the group aims to integrate other elements to be detected in the images and to quantify the real infestation rates of Aedes aegypti in a given region to refine and validate the model. “We hope to create a flowchart that can be applied in different cities to find risk areas without the need for home visits, a practice that wastes a lot of time and public money,” explains Chiaravalloti Neto.
In a previous study, the group had already used artificial intelligence to identify water reservoirs and swimming pools in Belo Horizonte (MG). The researchers began by presenting these satellite images of the mining town to a computer algorithm and indicated which ones had these facilities. Through a process of deep learning (or deep learning), the program began to identify patterns in the images that indicated the presence of a pool or reservoir of water. Over time, the system was able to differentiate these structures in the photos on its own.
“It’s really a machine learning process, a subfield of artificial intelligence,” says Jefersson Alex dos Santos, a professor in UFMG’s Department of Computer Science and founder of the Recognition and Patterns Laboratory. Earth observation.
For the present research, the professionals have delimited four regions of Campinas characterized by different socio-economic contexts, according to the IBGE (Brazilian Institute of Geography and Statistics). A drone with a high-resolution camera flew over these areas and took a series of photos. Thus, a database was created for water tanks and another for swimming pools.
The next step was to perform the learning technique transfer. “We trained this model in Belo Horizonte and applied it to Campinas,” explains Santos. With the images obtained in the city of São Paulo, the models became more reliable for the region, reaching an accuracy of 90.23% for the detection of swimming pools and 87.53% for the detection of exposed water tanks.
With the algorithm properly trained, the researchers used other images to calculate the concentration of exposed reservoirs and pools of water in these four previously selected regions of Campinas. By crossing this information with the IBGE data, it was noticed that the socio-economic indices were lower in areas with a greater concentration of water reservoirs and higher where there were more swimming pools.
Since less structured regions are more prone to infestation by Aedes aegypti, this model would already help to fight against the diseases spread by it. “Although this is not yet the definitive methodology, we can already think of a practical and relatively simple use of large-scale software development, with the aim of mapping the neighborhoods most at risk of epidemics of dengue,” Santos points out.
Chiaravalloti Neto points out that the models created could be useful beyond the control of dengue, zika and chikungunya: “The update of the socio-economic indices of different parts of Brazil occurs every ten years, with the census. With this technique, we would be able to renew this data more quickly, which can be used to fight against different diseases and problems.”
According to him, future work could find other markers from aerial images and, thus, refine these algorithms to ensure even greater reliability.
Drone or Satellite?
Although the aerial photos of Campinas were obtained with a drone, it is expected that in the future the strategy tested in this research will only use satellite images. “We used a drone because it was a study, but scanning with this equipment is expensive”, analyzes Chiaravalloti Neto.
“They also have less autonomy. To carry out a large-scale project, which includes large cities, we will need satellite images”, adds Santos. In the Belo Horizonte study, satellite images were used successfully – they need high resolution for the computer to identify patterns. According to Santos, access to this type of image is fortunately expanding.
Although this type of methodology seems expensive, it generates potential savings by eliminating the need for face-to-face visits to map dengue-prone areas house-by-house. Instead, health workers would take advantage of information obtained remotely — and processed with artificial intelligence — to get to priority locations with more confidence.
The current model is able to detect water tanks, but not if they are properly sealed. The same goes for swimming pools: he identifies them, but does not know if they are well maintained or closed. “This methodology could be refined to differentiate well-maintained structures from those that would actually serve as breeding grounds for mosquitoes,” says Chiaravalloti Neto. Blaming these patterns and other structures related to greater mosquito infestation would make the algorithm even more reliable in defining public health measures.
Currently, researchers are setting up a series of traps to Aedes aegypti in about 200 blocks of Campinas and assessing in detail the conditions of the properties and the presence of different mosquito breeding sites. Socio-economic indicators will also be examined. The next step will be to assess the aerial images of these regions with the same logic used in the aforementioned research to classify the degree of risk of the presence of the Aedes aegypti and diseases transmitted by it.
“By observing these blocks, we intend to build a dengue control prioritization model for the whole city and, later, for the rest of Brazil,” concludes Chiaravalloti Neto.
In addition to funding from Fapesp, the researchers had access to resources from the Serrapilheira Institute, the CNPq (National Council for Scientific and Technological Development), the USP Research Dean’s Office and the Support Foundation for research of Minas Gerais (Fapemig). Sucen also provided structural support.
The authors involved are: Higor Souza Cunha, Brenda Santana Sclauser, Pedro Fonseca Wildemberg, Eduardo Augusto Militão Fernandes, Jefersson Alex dos Santos, Mariana de Oliveira Lage, Camila Lorenz, Gerson Laurindo Barbosa, José Alberto Quintanilha and Francisco Chiaravalloti-Neto.