Modeling the Temporal Population Distribution of Ae. aegypti Mosquito using Big Earth Observation Data

TL;DR

This study develops a machine learning-based model using Earth observation data to predict the temporal distribution of female Aedes aegypti mosquitoes, aiding in vector control efforts for mosquito-borne diseases.

Contribution

It introduces a novel RF-based model integrating EO-derived environmental variables to estimate mosquito populations across different control regimes.

Findings

RF outperforms other models in accuracy

Environmental variables' importance varies by regime

A parsimonious model maintains high effectiveness

Abstract

Over 50% of the world population is at risk of mosquito-borne diseases. Female Ae. aegypti mosquito species transmit Zika, Dengue, and Chikungunya. The spread of these diseases correlate positively with the vector population, and this population depends on biotic and abiotic environmental factors including temperature, vegetation condition, humidity and precipitation. To combat virus outbreaks, information about vector population is required. To this aim, Earth observation (EO) data provide fast, efficient and economically viable means to estimate environmental features of interest. In this work, we present a temporal distribution model for adult female Ae. aegypti mosquitoes based on the joint use of the Normalized Difference Vegetation Index, the Normalized Difference Water Index, the Land Surface Temperature (both at day and night time), along with the precipitation information, extracted from EO data. The model was applied separately to data obtained during three different vector control and field data collection condition regimes, and used to explain the differences in environmental variable contributions across these regimes. To this aim, a random forest (RF) regression technique and its nonlinear features importance ranking based on mean decrease impurity (MDI) were implemented. To prove the robustness of the proposed model, we compared it to other machine learning techniques and statistical models. Our results show that machine learning techniques perform better than linear statistical models for the task at hand, and RF performs best. By ranking the importance of all features based on MDI in RF and selecting the subset comprising the most informative ones, a more parsimonious but equally effective and explainable model can be obtained...