The ASA’s Section on Statistics in Defense and National Security (SDNS) sponsored the student poster prize session at the Conference on Data Analysis (CoDA), which took place in Santa Fe, New Mexico, March 7–9.
The prize has become a major feature of CoDA and is a great way to introduce students to the section. The conference highlights data-driven problems of interest to the Department of Energy. Talks and posters feature research from the Department of Energy national laboratories, academia, and industry.
There were 28 students presenting posters, and three prizes were awarded to the following students:
First Prize, $400: Lauren M. Foster, Colorado School of Mines
When Does Uncertainty Matter While Modeling Climate Change in Mountain Headwaters? Contrasting Model Resolution and Complexity Under a Changing Climate in an Alpine Catchment
Alpine, snowmelt-dominated catchments are the source of water for more than one-sixth of the world’s population. These catchments are topographically complex, leading to steep temperature gradients and nonlinear relationships between water and energy fluxes. Recent evidence suggests alpine systems are more sensitive to climate warming, but these regions are vastly simplified in climate models and operational water management tools due to computational limitations. Simultaneously, point-scale observations are often extrapolated to larger regions where feedbacks can both exacerbate or mitigate locally observed changes. It is critical to determine whether projected climate impacts are robust to different methodologies, including model resolution and complexity.
Using high-performance computing and a representative headwater catchment modeled in Parflow-CLM, we determined the spread of uncertainty in hydrologic fluxes across 30 projected changes for the Rocky Mountains. These projections were run on multiple model configurations of varying complexity, including 100m and 1km resolution, a traditional land surface modeling single column vertical flow, and subsurface 3D lateral flow.
We find that model complexity alters nonlinear relationships between water and energy fluxes; for example, higher-resolution models predicted larger reductions to snowpack, surface water, and groundwater stores per degree of temperature increase, suggesting that warming signals may be underestimated in simple models. Hydrologic impacts from increases in temperature are robust to variation in model complexity, but impacts from changes in precipitation are within uncertainty bounds of model configurations. This result corroborates previous research showing that mountain systems are significantly more sensitive to temperature changes than to precipitation changes and that increases in winter precipitation are unlikely to compensate for increased evapotranspiration in a higher energy environment.
These experiments help to bracket the range of uncertainty in published literature of climate change impacts on headwater hydrology, characterize the role of precipitation and temperature changes on water supply for snowmelt-dominated downstream basins, and identify changes to climate impacts at different scales of simulation.
Second Prize, $100: Divya Banesh, UC Davis/Los Alamos National Laboratory
Finding Change Points in Time-Dependent Image Sequences Based on Feature Analysis
Time-dependent sequences of images derived from simulated scientific models or imaged from the experimental sciences can range over long periods of time with many time steps. Often, this data can be cumbersome for the scientist to parse, or extract pertinent information specific to their research. We present an analysis technique to extract change points from a time-dependent image sequence using image processing and computer vision tools. We use these tools to extract types of features that would be of interest to a scientist and analyze changes to these features over time. We then apply change detection to these features to extract time steps that may be most interesting to a scientist. To showcase these tools, we use a cinema database of MPAS-Ocean image files.
Honorable Mention: Michael Darling, Sandia National Laboratories, University of New Mexico
Uncertainty Propagation in Multimodal Image Analysis
The continued proliferation of remote sensing systems, particularly with respect to their use in decision-making applications ranging from national security to climate science, has resulted in increasing demand for data-driven analysis methods that quantify the uncertainty in their results. In this work, we explore the propagation of uncertainty through multiple levels of data analysis. In many cases, the ultimate result of a real-world data analysis is the output of the final stage of a pipeline of processing and inference steps. While statistical uncertainty quantification methods exist for individual pipeline stages, methods that systematically incorporate and propagate uncertainty end-to-end remain largely unexplored.
We demonstrate one possible approach to uncertainty propagation on a multi-sensor supervised pixel classification task based on co-registered optical and LiDAR images. The baseline classification scheme uses the raw data as its features and estimates the uncertainty in the assigned classes. We compare the baseline approach to one in the data that is clustered, and the cluster probabilities are used as features for supervised classifiers. We show how uncertainty propagates through the two approaches and compare the resulting classifications and associated uncertainties.
Leave a Reply