Mosquitoes are vectors for a number of serious illnesses, such as Dengue, Zika, Malaria, and West Nile Virus. In the United States, West Nile Virus (WNV) is the leading mosquito-borne disease. As there are currently no vaccines to prevent WNV nor medications to cure it, government agencies must sustain financially taxing programs to monitor mosquito populations and WNV infections and share this data across various departments in an effort to prevent WNV outbreaks. In this study, we develop four machine learning models that forecast WNV infections in humans, enabling government and healthcare officials to take proactive action instead of reacting to real-time infection data. Our models take in open-access data describing ecological variables – such as temperature, humidity, wind, air quality index (AQI), and enhanced vegetation index (EVI) — and use that data to predict future WNV infections five weeks in advance. We then perform a comparative analysis of the two types of machine learning architectures – support vector machine (SVM) regressors and random forest (RF) regressors – represented across our four models to evaluate which is best suited for the task. Our results indicate RF regressors are best suited to the task of forecasting WNV infections; however, SVM regressors perform comparably well and even exceed RF regressors when the magnitude of error is unweighted. Additionally, our results contribute a new perspective on the usefulness of AQI and wind speed for predicting mosquito-borne infections. Our RF regressor’s feature importance results indicate that AQI and wind speed were of similar importance as EVI and humidity – ecological variables well-known to influence mosquito population dynamics. Our work provides valuable directions for future research and development of early warning systems for disease prevention efforts as our models’ ability to forecast WNV infections five weeks in advance provides critical lead time for government officials to pursue mosquito containment efforts and healthcare facilities to increase capacity, enabling proactive action in combating WNV.Link to abstract published at AGU's Fall 2022 Session

Accurate land cover data can provide powerful insight into characterizing the effects of climate change. Remote sensing satellites enable state-of-the-art land cover measurements, but data collected on the Earth's surface offers a new perspective on land cover characteristics through its more localized scope. Areas that may be generalized to a single pixel in a remote sensing satellite’s data products can be observed at a more granular level through on-site data collection strategies. Low-cost sensors, such as NASA’s Science and Technology Education for Land/Life Assessment (STELLA), make such granular data collection more cost-effective. Citizen Science programs, like the Global Learning and Observations to Benefit the Environment (GLOBE) Program, provide a blueprint for reliably scaling this type of on-site data collection. STELLA is an open-source platform that allows volunteers, like Citizen Scientists, to measure electromagnetic waves to calculate irradiance and temperature in a specific Area Of Interest (AOI). STELLA is cost-effective, as its kit can be assembled by any user with access to makerspace tools commonly found in educational institutions, like 3D printers and soldering irons. This work presents a comparative analysis of the land cover measurements recorded by the STELLA sensor and remote sensing satellites, such as LANDSAT9. Surface temperatures were recorded hourly using the STELLA sensor on four different types of land cover within a 500-square-meter area in Reno, Nevada. The results indicate a statistically significant discrepancy between measurements recorded by the STELLA sensor and LANDSAT, highlighting an untapped data trove in localized sensor measurements. Additionally, we present a data collection control flow for Citizen Science volunteers to record reliable STELLA sensor data. We demonstrate how such Citizen Science data can provide a valuable alternative perspective when compared to its state-of-the-art counterparts, rendering it a valuable tool for future studies. AGU23 Poster LinkAGU23 Abstract Link