DeepMelt: Harnessing Deep Learning to Predict Future Meltwater Runoff from the Greenland Ice Sheet
My research is the first to apply a deep learning framework, grounded in in-situ observations, to predict meltwater runoff from the Greenland Ice Sheet. I am constructing, training, and testing deep learning algorithms to predict Greenland Ice Sheet meltwater runoff from Modèle Atmosphérique Régional (MAR) outputs, including 2m air temperature, ice surface temperature, albedo, and surface energy balance. The deep learning algorithm will be grounded in in-situ observations, generalizable to other areas of the Greenland Ice Sheet, scalable to ice-sheet scale, and applicable to remote sensing and/or climate reanalysis data. Using principles of interpretable AI, I am determining the most important features, searching for biases in training data, and providing insight into physical processes on the ice sheet surface. I am fine-tuning the highest-performing algorithm on in-situ observations of meltwater runoff. This deep learning framework aims to be an alternative and supplement to regional climate models for Greenland Ice Sheet meltwater runoff predictions by learning directly from observational data.
Rapid Convolutional Neural Network to Delineate and Map Ice Wedge Polygons
Using a rapid convolutional neural network with a watershed transformation, I delineated and mapped ice wedge polygons in Utqiagvik, AK from digital elevation models developed from lidar flown on unoccupied aerial systems. This machine learning framework delineates individual ice wedge polygons, measures their relief, and classifies them as high-, low-, or flat-centered—a representation of their state of degradation. Rapidly mapping ice wedge polygons and visualizing their relief provides insight into the stability of these features and the trafficability of warming Arctic terrain.
Machine Learning Models to Predict Active Layer Thickness
My research characterized and quantified spatiotemporal changes and disturbances in Arctic periglacial terrain subject to freeze-thaw cycles. I analyzed relationships between climate and geographic variables and soil active layer thickness (the depth to which soil seasonally freezes and thaws above permafrost) in Utqiagvik, AK. I constructed and tuned Random Forest algorithms to determine the most important predictors of changes in active layer thickness and built multiple linear regression models to predict future active layer thickness. My work is among the first to link decadal timeseries of climate data to annual active layer thickness measurements and develop models to predict active layer thickness on a local scale.
- Maciel-Seidman, M. L., Merrick, T. L., Grossman, S. M., Richards, D. F., Abelev, A., Vermillion, M. S., Liang, R. T. (2024). Analysis of Active Layer Thickness and Climate Data at Utqiagvik, Alaska with Random Forest and Multiple Linear Regression Algorithms, Poster, Annual Meeting of the American Geophysical Union, Washington, D.C., Dec. 9-13, 2024.
Quantifying Residential Carbon Emissions
This research worked towards determining the carbon offset value of energy efficiency retrofits to low-income housing. I developed a new quantitative methodology to estimate a home’s baseline annual carbon emissions and the potential carbon emissions reductions associated with energy retrofits, using open source resources such as the EPA eGRID database and historic energy bills. This study illustrated the potential for funding energy efficiency upgrades to low-income housing by selling a socially-responsible carbon offset generated by the resulting reductions in emissions from these upgrades, on the voluntary carbon market.
- Maciel-Seidman M, Tzankova Z, Ziegler CC, Lele A, Lu S, Yan Y and Muchira JM (2024) Mobilizing carbon offsetting to reduce energy cost burdens: a new approach for calculating and monetizing the offset value of energy efficiency upgrades to low-income housing. Front. Energy Res. 12:1437560. doi:10.3389/fenrg.2024.1437560
