Modeling Urban Areas Epidemiological Risk Exposure Using Multispectral Spaceborne Data

In recent decades, the world has been fast urbanizing. More than half of the world’s human population now live in urban areas. Such high density of urban population is resulting in air and water pollution, land degradation, and infectious diseases spread risks prominence. However, the increasing quality (in terms of finer spatial and temporal resolution)and quantity of Earth Observation (EO) satellite data provide new perspectives for analysing these phenomena. Within the specific domain of epidemiological risks dynamics in urban areas which is the focus of this work, the use of multispectral optical EO sensor data has created new opportunities. These data through their visible, near, mid, far and thermal infrared bands provide planetary-­‐scale access to environmental variables such as temperature, humidity, and vegetation types, location and conditions. Since these environmental variables affect the development of vectors causing infectious disease (e.g., mosquitoes), there is the possibility to use EO data to estimate them, and obtain disease risk models. The Ae. aegypti mosquito species transmits Zika, Dengue, and Chikungunya, diseases widespread in more than 100 world countries, and is concentrated in urban areas. The development of this vector depends significantly on local environmental temperature, humidity, precipitation and vegetation. In this regard, multispectral EO data can provide globally consistent and scalable sources to obtain the required environmental variable inputs, and extract significant and consistent monitoring and forecasting models for vector population. The work reported in this thesis about this topic has led to the following results: 1) A method to map vegetation types in urban areas at high spatial resolution using Sentinel2 multispectral EO data. The results show an improvement in the quality of the resulting vegetation maps with respect to what is available by means of state-­of-­the-­art techniques. 2) A method that combines EO-­based spectral indices, temperature layers, and precipitation measurement to model the temporal evolution of the local mean Ae. aegypti population. The approach leverages the random forest (RF) machine learning (ML) technique and its embedded nonlinear features importance ranking (mean decrease impurity, MDI) to rank the effects of environmental variables and explain the resulting model. 3) A weighted generalized linear modeling (GLM) technique to predict Ae. aegypti population using multispectral EO data covariate inputs. GLMs are generally simple to implement and explain, but do not provide the same level of prediction quality as ML methods. The proposed weighted GLM compares well with ML techniques in quality, and provides capability for more explicitly interpretation of the results. 4) A recurrent neural network (RNN) technique for spatio­‐temporal modeling of Ae. Aegypti population at the urban block level using multispectral EO data as inputs. This study is needed because spatial models obscure seasonality effects while temporal model are blind to spatial changes in micro-­climates. The proposed technique shows great promise with respect to the use of free multispectral EO data for spatio-­temporal epidemiological modeling. All the proposed techniques have been applied in the Latin American region where the risk of Ae. aegypti vector transmitted diseases are the highest in the world. They were validated thanks to the long term partnership with the University of Alagoas in Maceio (Brazil) and the Brazilian company: ECOVEC.