Deep Learning Optimizes Data-driven Representation of Soil Organic Carbon in Earth System Model over the Conterminous United States
Abstract
Soil organic carbon (SOC) is a key component of the global carbon cycle yet not well represented in Earth system models to accurately predict global carbon dynamics in response to climate change. This study innovatively integrated deep learning, data assimilation, 25,444 vertical soil profiles and the Community Land Model version 5 (CLM5) to optimize the model representation of SOC, with a specific focus over the conterminous United States. We firstly constrained parameters in CLM5 by observations of vertical profiles of SOC both in a batch mode (i.e., one-batch) with no differentiation of soil layers and sites, and at individual sites (i.e., site-by-site). The estimated parameter values from the site-by-site method were either randomly sampled (i.e., random-sampling) as continentally homogeneous parameters (i.e., one constant value for each parameter across the continent) or optimally extrapolated as spatially heterogeneous parameters (i.e., varying parameter values to match the spatial patterns from the site-by-site method) over the conterminous United States through the deep learning technique (neural networking). We compared modeled spatial distributions of SOC by CLM5 with default, randomly sampled, one-batch optimized, and deep learning derived parameter values to those derived directly from observations and two observation-driven gridded SOC maps. CLM5 with optimized parameter values from random-sampling and one-bath methods substantially corrects the overestimated SOC storage by that with default parameters, yet still has considerable geographical biases. CLM5 with the spatially heterogeneous parameter values derived from the neural networking method shows the least estimation error without much geographical bias across the U.S. continent. Our study indicates that deep learning in combination with data assimilation achieves a more accurate representation of massive soil carbon data by complex land biogeochemical models.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFM.B13H2595T
- Keywords:
-
- 0414 Biogeochemical cycles;
- processes;
- and modeling;
- BIOGEOSCIENCES;
- 0428 Carbon cycling;
- BIOGEOSCIENCES;
- 0480 Remote sensing;
- BIOGEOSCIENCES