PhD Research
Nutrient pollution and severe algal blooms remain major problems in western Lake Erie despite intergovernmental efforts to regulate sources in the U.S. and Canada. The Maumee River Basin has been the largest nutrient contributor to western Lake Erie. Historically, distributed agricultural areas dominated the nutrient contributions to the rivers, where sources include animal waste and inorganic fertilizer. Prior studies of nutrient source attribution have focused on large watersheds or counties at long time scales; source attribution at finer spatiotemporal scales, which can enable more effective nutrient management, remains a substantial challenge.
Our study addresses this challenge by attributing phosphorus release at the subwatershed scale using a lightweight network model framework. Since phosphorus release is uncertain, we integrated water-quality measurements, excess phosphorus availability over land, and flow dynamics into a probabilistic framework to attribute phosphorus release to different sources. Our model reveals significant spatial and temporal variability in phosphorus release, which is averaged out in the coarse-scale attribution calculated using sparsely deployed water-quality monitors.
Being able to identify such variability can greatly benefit targeted enforcement by enabling prioritization of regions, time periods, and source types with higher pollutant release.
Read more about this work
Bidirectional Flow in Volcanic Conduit and Bubble Dynamics
Persistently degassing volcanoes typically erupt multiple times a day, emitting copious gas and thermal energy but relatively little magma. The ascent of the erupted magma is likely a byproduct of degassing, but it is unclear how exactly eruptive behavior and degassing are related. Without the ability to observe degassing processes at depth, we rely on erupted samples to derive constraints on pre-eruptive flow conditions. Some of these samples seal in magma droplets named melt inclusions, which represent valuable snapshots of volatile concentrations at depth.
Here, we use numerical simulations to study how magma flow in vertical conduit-like structures at depth alters the volatile concentrations in melt inclusions. We demonstrate that volatile-rich, ascending magma mixes with volatile-poor, descending magma that has lost volatiles at the surface. The degree of mixing depends on the physical properties of magma and bubbles. This magma mixing, together with the carbon dioxide fluxing, can significantly shift the water and carbon dioxide concentrations in melt inclusions.
Our study suggests that some degree of magma mixing is almost inevitable in persistently degassing volcanoes, but that mixing may vary considerably with depth. We suggest that melt inclusion data could potentially help tracking the evolving flow conditions in volcanic conduits.
Read more about this work
Formation of Embayments and Melt Inclusions during Crystallization
Embayments and melt inclusions are pockets of melt hosted by magmatic crystals. They are widely used for constraining the melt composition and magmatic processes at depth prior to eruptions. We use numerical modeling to investigate how embayments and melt inclusions form in crystallization, and whether their volatile compositions are representative of the system.
We find that more significant boundary-layer H2O enrichment in surface dislocations leads to more decrease in undercooling, enhancing the formation of embayments and melt inclusions in a positive feedback. Faster background melt flow deepens and widens embayments. Initial shape of surface dislocations determines the formation and controls the geometry of embayments. Before ascent, embayment volatile contents are likely higher than matrix-melt contents, with a diffusion profile.
Reducing Sampling Bias of Landslides by Identifying Unrecorded Events from Satellite Images
Global records of landslides are subjected to incompleteness and sampling bias because of spatially different levels of monitoring and reporting.
In this work, we use a convolutional neural network to detect landslides from globally available satellite images and compare the detection with existing records of landslides. We train the model using an open high-resolution landslide dataset containing labeled landslide and non-landslide satellite images from the Bijie region in China. We test the trained model on low-resolution satellite images of Oregon, USA and Rwenzori Mountains, Uganda taken by Sentinel-2, which provides global coverage. To achieve better generalization of our model, we conduct multiple operations including downsampling the original training set to match the lower resolution of the test images, data augmentation, and distribution matching for the input test images. We visualize the model's decision making process using Grad-CAM. By comparing the detected landslides with the existing record, we find that the landslides in Rwenzori Mountains, Uganda are much more under-reported than in Oregon, USA.
This work can be used to detect potentially unrecorded landslides and understand the landslide sampling bias globally.
Optimizing Contact Tracing Policies for COVID-19
COVID-19 success stories from countries using contact tracing as an intervention tool for the pandemic have motivated US counties to pilot opt-in contact tracing applications. Contact tracing involves identifying individuals who came into physical contact with infected individuals.
We build on an agent-based epidemiological simulator that resolves spatiotemporal dynamics to model San Francisco, CA, USA. Census, OpenStreetMap, SafeGraph, and Bureau of Labor Statistics data inform the agent dynamics and site characteristics in our simulator. We test different agent occupations that create the contact network, e.g. educators, office workers, restaurant workers, and grocery workers. We use Bayesian Optimization to determine transmission rates in San Francisco, which we validate with transmission rate studies that were recently conducted for COVID-19 in restaurants, homes and grocery stores. Our sensitivity analysis of different sights show that the practices that impact the transmission rate at schools have the greatest impact on the infection rate in San Francisco.
Through our research, we are able to identify the occupations, like educators, that are at greatest risk. We use common geophysical data analysis techniques to bring a different set of insights into COVID-19 and policy research.
Read more about this work
Quantifying Groundwater Storage in Central Valley, California Using GRACE Data and Head Measurements
Groundwater resource is crucial for the global water supply and agricultural development. Some of the most important agricultural regions, such as California's Central Valley, depend heavily on groundwater for irrigation. However, groundwater management is challenging because we are unable to measure groundwater storage directly.
We compare the ability of using GRACE data and in situ measurements of wells in quantifying groundwater storage changes. We develop an approach to spatially aggregate the well data and calculate regional water storage in a higher frequency than that of individual wells. In addition, we also explore whether the spatial variations in the changes of groundwater level are indicative of subregional aquifer properties.
