DEEP LEARNING MODELS FOR ANALYZING AND PREDICTING MOSQUITO POPULATIONS

Abstract

Understanding and forecasting mosquito population dynamics is a critical component of managing vector-borne disease risk, particularly in regions where Aedes aegypti mosquitoes are endemic. This dissertation presents a progression of modeling strategies for predicting mosquito abundance, with a central theme of integrating mechanistic and machine learning approaches. The first study evaluates the ability of deep neural networks to learn the dynamics of a high-fidelity mechanistic mosquito population model. The results show that equation-free models trained on synthetic data can successfully replicate spatiotemporal abundance patterns, generalize well across diverse environmental conditions, and provide significant computational speedup. The second study introduces a probabilistic forecasting framework that converts local weather data into weekly predictions of gravid female trap counts. The method demonstrates robustness to imperfect weather forecasts and offers a practical tool for real-time surveillance. The third study presents a hybrid modeling approach using Universal Differential Equations (UDEs), in which a shallow neural network is embedded within a compartmental mosquito life cycle model. This structure enables the model to learn environmental drivers of mosquito trap counts while preserving biological interpretability. The UDE model accurately reproduces synthetic trap count dynamics and highlights the value of combining mechanistic and machine learning approaches. Collectively, these studies demonstrate the potential of scientific machine learning to advance mosquito population forecasting and inform targeted vector surveillance and control strategies.