Dengue incidence rate estimation using aerial and street-level urban imagery with deep learning models

• Autor(a) principal: Andersson, Virgínia Ortiz
• Data de Publicação: 2019
• Tipo de documento: Tese
• Idioma: por
• Título da fonte: Repositório Institucional da UFPel - Guaiaca
• Texto Completo: http://guaiaca.ufpel.edu.br/handle/prefix/6406
• topic: CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO; Computação; Dengue fever estimation; Visual computational sociology; Street-level images; Aerial images; Deep convolutional neural networks; Machine learning; Computer vision; Estimativa de dengue; Sociologia computacional visual; Imagens no nível da rua; Imagens aéreas; Redes neurais convolucionais; Aprendizado de máquina; Visão computacional

Motivated by sociological theories that present the physical appearance of the urban environment as an influential factor in the behavior of inhabitants, the new Visual Computational Sociology research area has investigated computer vision models to infer latent variables such as demographic, socioeconomic, cultural, and health indicators from aerial and street-level urban imagery. Just like crime events can be inferred from the appearance of the urban environment, occurrences of diseases, such as dengue fever, can be explained from visual data as well. This work proposes the use of aerial and street-level images to estimate dengue fever incidence rates, in an automated way, to increase the estimation effectiveness of dengue and its variants in urban regions. Specifically, it was proposed using computer vision techniques capable of extracting attributes from urban images automatically and neural network models for multiple regression to estimate latent variables of dengue incidence using urban environment visual attributes as predictors. For this, experiments were carried out using street-level and aerial images, together with historical dengue fever data obtained from the Brazilian capitals Rio de Janeiro (RJ), São Paulo (SP), and Salvador (BA). Results showed evidence that: (i) street-level image features can be used for estimating dengue incidence rates, although models using aerial image features present better results; (ii) the combination of aerial and street-level features contribute to better results in estimating dengue incidence rates; (iii) models generalize poorly to other cities, slightly improving the results when using transfer-learning techniques and multiple cities in training and (iv) Deep Convolutional Neural Networks (Deep Convnets) are suitable for use in the proposed model, since its features presented better results compared to designed descriptor techniques. At last, it is expected that the proposed models will contribute to an improvement in the state of the art of dengue estimation models, and the obtained results contribute to public health policies in urban centers, through better results or in optimizing their accomplishment.