Accurate and efficient prediction of fine‐resolution hydrologic and carbon dynamic simulations from coarse‐resolution models

George Shu Heng Pau, Chaopeng Shen, W. J. Riley, Yaning LiuWater 2016 · DOI

TL;DR

This paper introduces a novel method, the Proper Orthogonal Decomposition Mapping Method (PODMM), to accurately predict fine-resolution hydrological and carbon dynamics using coarse-resolution model outputs. This approach significantly reduces computational costs (by ~1000x) while achieving results comparable to computationally expensive fine-resolution simulations, paving the way for more efficient large-scale Earth system modeling.

What Problem Does This Paper Address?

Hydrological and biogeochemical (BGC) models often need to operate at very fine spatial resolutions to capture important processes like "hot spots" of activity. However, running these hyperresolution models is computationally prohibitive, especially for tasks requiring many simulations (e.g., uncertainty analysis, data assimilation). A common workaround is to "upscale" model parameters and variables to coarser resolutions. This upscaling process, however, can introduce significant biases because it often fails to fully capture the complex, nonlinear interactions between different processes that occur at finer scales. This paper addresses the challenge of bridging this gap: how to obtain accurate, fine-resolution simulation results without the immense computational burden of running full hyperresolution models.

Methods

The core of this paper is the application and enhancement of the Proper Orthogonal Decomposition Mapping Method (PODMM). Here's a breakdown:

  • Proper Orthogonal Decomposition (POD): This is a technique used to identify the dominant modes of variability in a dataset. In this context, it's applied to snapshots of fine-resolution (f) and coarse-resolution (g) model outputs. POD helps to reduce the dimensionality of the problem by capturing the most important spatial and temporal patterns.
  • Mapping Method: PODMM uses the dominant POD modes to create a "reduced-order model" (ROM). This ROM then learns a mapping from coarse-resolution solutions (g) to fine-resolution solutions (f). Essentially, it learns how to reconstruct the fine-scale details from the coarser representation.
  • Key Improvements:
    • Joint Estimation & Temporal Variation: The authors improved the ROM by jointly estimating correlated variables and allowing ROM parameters to vary temporally. This reduced approximation errors by up to 20%.
  • Robustness through Multiple ROMs: To improve robustness, they constructed multiple ROMs using different sets of variables and selected the best approximation based on an error estimator.
  • Error Estimator: A crucial component is an error estimator that efficiently quantifies approximation errors, allowing for the selection of the best-performing ROM without needing the true fine-resolution solution.
  • PODM3 (POD Multicomponent Mapping Method): This extension incorporates correlated ancillary variables into the data matrix to potentially improve accuracy by allowing for more POD bases.
  • Monthly Varying Optimal M (PODMM-vm, PODM3-vm): Recognizing that optimal model complexity can vary seasonally, they developed versions that use month-specific optimal numbers of POD bases.
  • PODM∞(€): This represents the best approximation chosen from all considered methods (PODMM, PODM3, PODMM-vm, PODM3-vm) based on the empirical error estimator. This was found to be the most robust and accurate method.

Datasets

The study utilized data from the Clinton River Watershed (CRB) in Michigan.

  • Model Simulations: The PAWS+CLM model was used to generate both fine-resolution (220 m) and coarse-resolution (7 km and 3.5 km) simulation outputs for the period 2002-2009. These simulations served as the training and validation datasets for the ROMs.
  • Input Data: The PAWS+CLM model was constructed using high-resolution elevation, land use, soil data, river hydrography, climate forcing data (precipitation, temperature, humidity, wind), and simulated steady-state carbon and nitrogen states.

The paper does not introduce new benchmark datasets but rather uses existing data to create a complex simulation environment for testing their ROM approach.

Key Insights and Ideas

  • Upscaling Bias is Real and Significant: The paper strongly demonstrates that simply averaging parameters to coarser resolutions introduces biases, particularly in simulating nonlinear processes. The ROM approach directly addresses this by learning the complex relationships.
  • PODMM Captures Heterogeneity: The PODMM method is remarkably effective at reproducing the spatial heterogeneity of fine-resolution simulations, even when starting from coarse-resolution data. This is a major step beyond traditional upscaling.
  • Correlated Variables are Powerful: Incorporating correlated ancillary variables (e.g., soil moisture at different depths for predicting surface soil moisture) significantly improved the ROM's accuracy and robustness. This highlights the value of multivariate approaches.
  • Temporal Dynamics Matter: Allowing ROM parameters to vary monthly improved performance, underscoring the seasonal nature of hydrological and carbon processes.
  • Error Estimation is Key for Practicality: The development of an empirical error estimator that accurately approximates the true error is crucial for selecting the best ROM in real-world applications where the true solution is unknown.

Key Results

  • Computational Speedup: The ROMs achieved an O(1000) reduction in computational cost compared to fine-resolution simulations.
  • Accuracy: The ROMs accurately predicted soil moisture, latent heat flux, and net primary production.
  • Bias Reduction: Biases in upscaled ROM solutions were reduced by up to 80% compared to coarse-resolution solutions.
  • Distribution Similarity: The subgrid probability distributions of the ROM solutions were nearly indistinguishable from those obtained using the fine-resolution model.
  • Error Metrics: For soil moisture, latent heat flux, and NPP, the validation errors (EROM) were generally small (e.g., 80% of snapshots had errors below 0.04 m³ m⁻³, 9.8 W m⁻², and 4.9×10⁻⁶ g Cm⁻² s⁻¹).
  • Improved Methods: PODM3 and PODM∞(€) showed improved accuracy and robustness over basic PODMM, with PODM∞(€) being identified as the best overall method.

Limitations and Future Work

  • Discrete Conservation: The current mapping scheme does not explicitly enforce discrete conservation of quantities (e.g., water). The authors suggest a two-way coupling mechanism with the coarse model to address this.
  • Independent Variables: The method currently models variables independently, without explicitly enforcing physical relationships between them. While the tested variables were weakly correlated and the ROM reproduced these correlations, this is a limitation for strongly coupled processes.
  • Scale Mismatch: Directly downscaling from kilometer to centimeter scales using PODMM might be challenging due to computational demands. A hierarchical approach coupling multiple PODMM-based ROMs at different scales is proposed.
  • Site-Specific vs. General ROMs: While site-specific ROMs are effective, developing ROMs applicable to multiple sites (site-independent ROMs) is a future goal, though it presents challenges in reproducing structured heterogeneities.
  • Computational Demands for ROM Construction: Constructing ROMs can still be computationally demanding, requiring efficient use of supercomputers and potentially off-peak cycles. The authors are developing a software framework (pROME) to facilitate this.

Summary updated

Abstract

Abstract The topography, and the biotic and abiotic parameters are typically upscaled to make watershed‐scale hydrologic‐biogeochemical models computationally tractable. However, upscaling procedure can produce biases when nonlinear interactions between different processes are not fully captured at coarse resolutions. Here we applied the Proper Orthogonal Decomposition Mapping Method (PODMM) to downscale the field solutions from a coarse (7 km) resolution grid to a fine (220 m) resolution grid. PODMM trains a reduced ‐ order model (ROM) with coarse‐resolution and fine‐resolution solutions, here obtained using PAWS+CLM, a quasi‐3‐D watershed processes model that has been validated for many temperate watersheds. Subsequent fine‐resolution solutions were approximated based only on coarse‐resolution solutions and the ROM. The approximation errors were efficiently quantified using an error estimator. By jointly estimating correlated variables and temporally varying the ROM parameters, we further reduced the approximation errors by up to 20%. We also improved the method's robustness by constructing multiple ROMs using different set of variables, and selecting the best approximation based on the error estimator. The ROMs produced accurate downscaling of soil moisture, latent heat flux, and net primary production with O(1000) reduction in computational cost. The subgrid distributions were also nearly indistinguishable from the ones obtained using the fine‐resolution model. Compared to coarse‐resolution solutions, biases in upscaled ROM solutions were reduced by up to 80%. This method has the potential to help address the long‐standing spatial scaling problem in hydrology and enable long‐time integration, parameter estimation, and stochastic uncertainty analysis while accurately representing the heterogeneities.
BibTeX
@inproceedings{pau2016accurate,
  author = {George Shu Heng Pau and Chaopeng Shen and W. J. Riley and Yaning Liu},
  title = {{Accurate and efficient prediction of fine‐resolution hydrologic and carbon dynamic simulations from coarse‐resolution models}},
  booktitle = {Water},
  year = {2016},
  doi = {10.1002/2015wr017782},
}
Figure