Search for: All records

Machine learning has revolutionized environmental sciences by estimating scarce environmental data, such as air quality, land cover type, wildlife population counts, and disease risk. However, current methods for validating these models often ignore the spatial or temporal structure commonly found in environmental data, leading to inaccurate evaluations of model quality. This paper outlines the problems that can arise from such validation methods and describes how to avoid erroneous assumptions about training data structure. In an example on air quality estimation, we show that a poor model with anr²of 0.09 can falsely appear to achieve anr²value of 0.73 by failing to account for Simpson’s paradox. This same model’sr²can further inflate to 0.82 when improperly splitting data. To ensure high-quality synthetic data for research in environmental science, justice, and health, researchers must use validation procedures that reflect the structure of their training data.

 

Efficiently managing agricultural irrigation is vital for food security today and into the future under climate change. Yet, evaluating agriculture’s hydrological impacts and strategies to reduce them remains challenging due to a lack of field-scale data on crop water consumption. Here, we develop a method to fill this gap using remote sensing and machine learning, and leverage it to assess water saving strategies in California’s Central Valley. We find that switching to lower water intensity crops can reduce consumption by up to 93%, but this requires adopting uncommon crop types. Northern counties have substantially lower irrigation efficiencies than southern counties, suggesting another potential source of water savings. Other practices that do not alter land cover can save up to 11% of water consumption. These results reveal diverse approaches for achieving sustainable water use, emphasizing the potential of sub-field scale crop water consumption maps to guide water management in California and beyond.

 

Mosquito-borne diseases (MBD) threaten over 80% of the world’s population, and are increasing in intensity and shifting in geographical range with land use and climate change. Mitigation hinges on understanding disease-specific risk profiles, but current risk maps are severely limited in spatial resolution. One important determinant of MBD risk is temperature, and though the relationships between temperature and risk have been extensively studied, maps are often created using sparse data that fail to capture microclimatic conditions. Here, we leverage high resolution land surface temperature (LST) measurements, in conjunction with established relationships between air temperature and MBD risk factors like mosquito biting rate and transmission probability, to produce fine resolution (70 m) maps of MBD risk components. We focus our case study on West Nile virus (WNV) in the San Joaquin Valley of California, where temperatures vary widely across the day and the diverse agricultural/urban landscape. We first use field measurements to establish a relationship between LST and air temperature, and apply it to Ecosystem Spaceborne Thermal Radiometer Experiment data (2018–2020) in peak WNV transmission months (June–September). We then use the previously derived equations to estimate spatially explicit mosquito biting and WNV transmission rates. We use these maps to uncover significant differences in risk across land cover types, and identify the times of day which contribute to high risk for different land covers. Additionally, we evaluate the value of high resolution spatial and temporal data in avoiding biased risk estimates due to Jensen’s inequality, and find that using aggregate data leads to significant biases of up to 40.5% in the possible range of risk values. Through this analysis, we show that the synergy between novel remote sensing technology and fundamental principles of disease ecology can unlock new insights into the spatio-temporal dynamics of MBDs.

 

School closures may reduce the size of social networks among children, potentially limiting infectious disease transmission. To estimate the impact of K–12 closures and reopening policies on children's social interactions and COVID-19 incidence in California's Bay Area, we collected data on children's social contacts and assessed implications for transmission using an individual-based model. Elementary and Hispanic children had more contacts during closures than high school and non-Hispanic children, respectively. We estimated that spring 2020 closures of elementary schools averted 2167 cases in the Bay Area (95% CI: −985, 5572), fewer than middle (5884; 95% CI: 1478, 11.550), high school (8650; 95% CI: 3054, 15 940) and workplace (15 813; 95% CI: 9963, 22 617) closures. Under assumptions of moderate community transmission, we estimated that reopening for a four-month semester without any precautions will increase symptomatic illness among high school teachers (an additional 40.7% expected to experience symptomatic infection, 95% CI: 1.9, 61.1), middle school teachers (37.2%, 95% CI: 4.6, 58.1) and elementary school teachers (4.1%, 95% CI: −1.7, 12.0). However, we found that reopening policies for elementary schools that combine universal masking with classroom cohorts could result in few within-school transmissions, while high schools may require masking plus a staggered hybrid schedule. Stronger community interventions (e.g. remote work, social distancing) decreased the risk of within-school transmission across all measures studied, with the influence of community transmission minimized as the effectiveness of the within-school measures increased.