Embedded Files
Global vegetation diversity is inadequately represented in hydrologic models. This contributes to model error in predicting variables of human and natural interest, such as streamflow, evapotranspiration, soil moisture, and wildland fire risk. To address this, my research has two primary aims. First, I develop and test new methods for mapping vegetation diversity and moisture content at large scales using a fusion of process-based models, machine learning methods, and remote sensing data. Second, I develop strategies for better representing this mapped vegetation diversity in global hydrologic models.
Better representing vegetation diversity in a land surface model
In this study, I developed a new method for parameterizing an important stomatal conductance parameter (g1) in an operational land surface model as a function of local environment. This improves g1's spatial variability relative to traditional parameterization methods. Additionally, we find that this new parameterization method improves model performance for evapotranspiration and streamflow over the default model by 11.8% and 20.1%, respectively. Check out a talk I gave a talk on this study here. Additionally, this work was highlighted for its success by the NASA Science Managed Cloud Environment.
Parameterizing g1 as a function of local environment (the light green distribution; denoted here as EF) leads to significant variability in g1 within land cover types. This increased within-land cover type variability is especially stark compared to the traditional modes of parameterization (the purple and black lines; denoted here as PFT-based parameterization) which allow for only a single g1 value for each land cover type.
Validating plant traits retrieved via a combination of processed-based and machine learning models
State of the art "differentiable" hydrologic models now combine process-based modeling and machine learning to improve model parameterization and predictions. Such models should theoretically improve the accuracy of vegetation traits retrieved globally (see schematic below). Despite this potential, the accuracy of such retrievals are yet untested. Therefore, I am currently working on the first project to test the accuracy of plant traits retrieved via differentiable modeling. Confirming the validity of this model would open a new avenue for mapping and understanding vegetation variability globally.
Differentiable models can fuse traditional processed-based modeling with machine learning to retrieve plant hydraulic traits. To do this, such models train neural networks to predict plant traits as a function of local environment. Importantly, differentiable modeling allows these neural networks to be trained via indirection observations (in this example, using live fuel moisture content, or LFMC) since observations of plant hydraulic traits are sparse.
Developing a long-term, deep learning-based dataset of live fuel moisture content
Live fuel moisture content (LFMC) is key predictor of wildfire risk. Previous work has used deep learning to generate LFMC maps across the western United States by combining microwave and optical remote sensing data. However, microwave data is only available for recent years (2016-2021), limiting the temporal extent of LFMC estimates. My research aims to improve upon this dataset by developing a deep learning framework that leverages insights from the microwave-based LFMC dataset while predicting LFMC without relying directly on microwave inputs.
Page updated
Google Sites
Report abuse