Blue Waters User Portal |

2021-2022 New Frontiers Graduate Fellows

The following are the 2021-2022 fellows and their project plans.

	Lucas Ford North Carolina State University Project title: Data-driven space-time models for generalizing reservoir operations to improve streamflow forecasting over the contiguous U.S. Project abstract: Streamflow impoundment behind dams is one of the largest human-induced modulations of the hydrologic cycle over the land surface, yet the most sophisticated streamflow forecasting models currently rely on overly simplified reservoir representations and their release policies. Reservoir operations are driven by various factors including water demand, flood control, and hydropower generation, thus generalized operations for large river basins are difficult to derive. Using the Blue Waters supercomputer, I will apply a series of data-driven statistical models to develop generic rules for release estimates based on physical variables and readily available reservoir characteristics. These release policies will be distilled into an efficient algorithm that can be easily implemented within land surface models or as a post-processing tool to provide more accurate streamflow forecasts. The performance of this framework will be evaluated using 1500 reservoirs across the contiguous U.S., allowing for a thorough uncertainty quantification with respect to various geographical and climate settings as well as under diverse reservoir characteristics.
	Dale Forrister University of Utah Project title: Seeing the forest and the trees: Drone based Phenological Monitoring in the Amazonian Rain Forest. Project abstract: Tropical rainforests are the most biodiverse and complex ecosystems on Earth due to their intricate webs of interacting species. For example, the interactions between plants and their insect herbivores form the cornerstone of tropical forest food webs. My research aims to understand how the timing and duration of leaf production(phenology) shapes these interactions. Our understanding of plant phenology in tropical forests is limited by the extreme diversity of plants as well as consistent cloud cover that makes satellite-based remote sensing difficult. To overcome these challenges I use drone based imagery taken above the forest canopy in the Ecuadorian Amazon to track new leaf production. I will use the Blue Water petascale computer system to construct high resolution 3D canopy models and orthomosaics. From these models I will identify phenological events such as leaf production and flowering using a combination of time series analysis and machine learning approaches. Ultimately, I will integrate this crown-level phenology data with an extensive study of changes in insect herbivore communities throughout the year to understand how leafing phenology shapes tropical rainforest food webs.
	Rachel Flood Heaton University of Illinois at Urbana-Champaign Project title: A Computational Account of Human Visual Reasoning Project abstract: Nearly half of the human brain is devoted to vision, and yet there is still much we don’t understand about how that half works: How does the human visual system deliver representations that allow us to understand and reason about the visual world? How does it make contact with our daily experiences, such as the quality of the user experience of a coffee maker, what we deem fashionable, the metaphors we see in artwork, and even how we know whether we can sit on something we recognize as a chair? This question also manifests in emergent and risky technologies that we hope will emulate our own behaviors, such as self-driving cars, unmanned drones (UAV’s), and missile and landing guidance systems. I develop computational models of the human visual system that simulate the way we generate mental representations of the objects we see in our environment, and how we compare our prior knowledge to those representations in order to make inferences about those objects. Starting from 2-dimensional images, the models simulate the activity of large numbers of individual neurons in the brain that fire together to represent contours, surfaces, objects, scenes, and ultimately bind both objects and scenes into abstract representations that can be used for reasoning. The goal of my work is to provide information about psychological theories to vision scientists, new computational methods to artificial intelligence researchers, and give all of us more insight into our own nature and experiences.
	Deanna Nash University of California Santa Barbara Project title: Simulating and evaluating atmospheric river-related precipitation in High Mountain Asia Project abstract: Atmospheric Rivers (ARs) are relatively infrequent corridors of water vapor in our atmosphere that contribute both to beneficial water resources and hazardous weather conditions via precipitation across the globe. ARs reach High Mountain Asia (HMA), impact seasonal precipitation totals, and result in extreme weather events such as flooding, landslides, and lightning storms. Yet, the effects of ARs on water resources and hazards in HMA are not well established. This project will employ multiple sources of data, including reanalysis, remotely sensed satellite data, and regional models to investigate the characteristics of both hazardous and beneficial precipitation in HMA that results from ARs. Using the Weather Research and Forecasting model, we will perform multiple high-resolution (e.g. 1 km and half-hourly) downscaling simulations to investigate the atmospheric flow interactions of ARs with the complex terrain of HMA in order to further understand their importance in the rain and snowfall distributions over HMA. We aim to improve understanding of both beneficial and hazardous ARs in HMA and provide information for future work that aims to improve forecasting skill in a region vulnerable to the impacts of climate change.
	Adam Stewart University of Illinois at Urbana-Champaign Project title: TorchGeo: torchvision for geospatial data Project abstract: The torchvision library has revolutionized the field of computer vision, providing data loaders for commonly used datasets, pre-trained models for transfer learning, and built-in transforms for data augmentation. Due to its reliance on the Python Imaging Library for data storage and ImageNet for pre-training, torchvision is not suitable for multi-spectral satellite imagery. I propose the development of TorchGeo, a library designed to bring the convenience of torchvision to the geospatial data domain. This work will benefit the research community in two different ways. First, it will allow computer vision researchers who know little about geospatial data to benchmark their models on common satellite imagery datasets. Second, it will allow remote sensing researchers who know little about computer vision to benefit from advancements in deep learning research.