While there has been considerable advances in models for gap filling, i.e., predicting missing entries in a data matrix or continuous field, existing models have been found to perform poorly in an important scientific problem: characterizing plant traits, such as specific leaf area (SLA), leaf nitrogen concentration (Leaf N), and leaf phosphorus concentration (Leaf P), towards accurate quantification of terrestrial ecosystem functions, such as agricultural and forest productivity, and regulation of atmospheric CO2 levels. Plant traits are key to understanding, predicting, and ameliorating the response of biodiversity and terrestrial ecosystems to anthropogenic disturbance and expected environmental changes, e.g., land use changes, increasing level of atmospheric CO2 concentration, or climate change. A detailed uncertainty quantified global map of plant traits, e.g., SLA, Leaf N, Leaf P, etc., over space and time, projecting into the future, will help considerably advance Earth system models (ESMs, coupled models of the atmosphere, ocean, sea ice, and terrestrial land surface), which form the basis for future climate projections. The focus of the project was on developing probabilistic models for matrix gap filling (MGF) for data matrices and spatiotemporal gap filling (STGP) for spatial fields, along with applications of such models to gap filling plant trait databases and generating global maps of key plant traits which can be used as input to ESMs.

The project accomplished all of the planned tasks, leading to advances in probabilistic modeling of data matrices and spatial fields of variables, and their applications to terrestrial ecosystem modeling. For MGF, our project developed a new model based on Bayesian hierarchical probabilistic matrix factorization (BHPMF) which uses a domain based hierarchy to generate uncertainty quantified gap-filling of data matrices. The model was used to gap-fill the TRY database, the largest database on plant traits. We have also developed spatiotemporal process models for STGF along with highly scalable inference based on dynamic nearest-neighbor Gaussian processes. We have also developed gradient boosted normalizing flows for flexible probabilistic modeling in high dimensions based on deep generative models. In addition, for efficient computations for approximate inference in complex probabilistic models, we have developed regularized variational inference, conjugate-computation variational inference, and related algorithms which are computationally efficient, scalable, and stable.

Our advances in probabilistic models for spatial gap filling have been used to generate one of the first uncertainty quantified global maps of key plant traits, such as leaf N and leaf P. Further, such uncertainty quantified trait maps have been used as input to ESMs yielding improvements in carbon flux modeling compared to the state-of-the-art approach in land surface modeling based on deterministic plant functional types. These simulations have increased the variation in plant traits by orders of magnitude compared to the current paradigm. Our results on improved carbon flux modeling may have a transformative impact on how land surface modeling is done going forward.

Last Modified: 12/31/2020
Modified by: Arindam Banerjee