Ph.D. student Zhihao Wang, alongside co-authors GEOG Assistant Professor Yiqun Xie, Assistant Research Research Professor Lei Ma and Professor and Associate Chair George Hurtt, in addition to Assistant Professor Xiaowei Jia from the University of Pittsburgh, recently published their paper entitled “High-Fidelity Deep Approximation of Ecosystem Simulation Over Long-Term at Large Scale” in the 31st ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems, which is a top-venue for spatial computing.

Process-based models are frequently utilized by climate and ecology scientists to tackle complex questions related to climate change. However, these models currently operate at very coarse spatial resolution (e.g., >100 km) due to computation cost. Artificial intelligence, especially deep learning, has achieved promising success in accelerating sophisticated and expensive physical simulation models, as seen by examples such as NVIDIA’s FourCastNet and Google DeepMind’s GraphCast, which predict  “weather conditions up to 10 days in advance more accurately and much faster than the weather simulation system led by the European Center”, quoted from Google DeepMind news for their Science publication in 2023. 

Similar computational bottlenecks have also become a limiting factor in ecosystem modeling. In this work, the team focused on the Ecosystem Demography (EDv3) model, which represents the cutting-edge global ecosystem model. ED has been used by NASA Carbon Monitoring System as well as the Global Ecosystem Dynamics Investigation (GEDI) mission, and it has also been deployed to real operational environments (e.g., Maryland).

Designing a deep learning model to approximate ED is a non-trivial task. A major challenge is error accumulation, where small deviations at each small time step (e.g. a day or month) tends to be amplified after longer projections (e.g. 100 years). Moreover, target variables of ED have heterogeneous patterns that may confuse the model during the training process.

This work developed a new deep learning framework, Deep-ED, to explicitly address these challenges, explained Wang. Experiment results in the northeastern U.S. demonstrated high approximation quality and efficiency. 

“This work holds the promise to dramatically increase the speed of the most sophisticated ecosystem modeling, thereby opening new applications previously impossible, while retaining process level understanding”, said Hurtt, who serves as the science team leader of NASA Carbon Monitoring System. 

The team will build on this advancement to further expand the capabilities of the AI-powered model (e.g., high-resolution at global scales).

Image: NVIDIA, Google and UMD models