Abstract
Ecosystem respiration ($R_{eco}$) is a critical component of the global carbon cycle, yet accurately characterizing its dynamics remains a challenge due to the complex interplay between biological processes and climatic drivers. This paper introduces a novel data-driven and physics-aware approach for estimating $R_{eco}$ dynamics using Dynamic Mode Decomposition with control input (DMDc). By representing $R_{eco}$ as a state-space model with autonomous components and exogenous climatic controls, the proposed method allows for intervention studies and a deeper understanding of ecosystem-climate interactions. Experimental results using Fluxnet2015 data demonstrate that DMDc achieves prediction accuracy comparable to state-of-the-art methods while providing unique insights into the low-dimensional behavior of various vegetation ecosystems.
Introduction
Predicting climate change requires precise modeling of $R_{eco}$, the total carbon dioxide released by plants, bacteria, and animals. Traditional modeling approaches often rely on simple exponential equations based on temperature, which frequently fail to capture the non-stationary nature of temperature sensitivity in real-world ecosystems. Conversely, deep learning models, while powerful, often function as "black boxes" that lack physical interpretability and require vast amounts of data. This research seeks a middle ground: a flexible, data-driven model that remains grounded in the physical and dynamical nature of ecosystems.
Methodology
The core of this approach is Dynamic Mode Decomposition with control input (DMDc), an emerging tool for analyzing nonlinear dynamical systems.
State-Space Representation
The model treats $R_{eco}$ as a state-space system where the current state is influenced by its own internal "autonomous" rhythm and external "control inputs," such as air temperature ($T_{air}$) or soil water content (SWC). The DMDc algorithm decomposes complex ecosystem data into linear "modes," each characterized by a specific frequency and growth or decay rate.
Time Delay Embedding (TDE)
To account for unobserved variables that influence an ecosystem's trajectory, the researchers employed Time Delay Embedding (TDE) based on Takens’ Theory. By incorporating a history of past measurements into the current state, the model can capture the influence of latent factors without requiring direct measurements of every possible driver.
Experimental Setup
The study utilized half-hourly eddy covariance measurements from the Fluxnet2015 dataset. The model was tested across diverse vegetation types, including Deciduous Broadleaf (DBF), Evergreen Needleleaf (ENF), Mixed Forests (MF), and Grasslands (GRA). Analysis focused on the peak growing seasons (May to September) to capture the most active periods of ecosystem respiration.
Key Insights and Ecosystem Dynamics
Low-Dimensionality and Dominant Modes
A significant finding of this work is that despite the inherent complexity of ecosystems, their primary respiration behavior is relatively low-dimensional. Analysis of the singular values of the data suggests that most ecosystems are governed by just four to six dominant dynamic modes.
As shown in Figure 2, the rapid decay of singular values across various sites indicates that a small number of modes capture the majority of the system's variance. Sites with a "well-structured spread" of these values tend to be more predictable, whereas noisier data or slower climate dynamics lead to less structured singular value profiles.
Driver Sensitivity and Intervention
The separation of autonomous dynamics from control inputs allows for "virtual intervention" studies. For instance, researchers can swap $T_{air}$ for SWC as the control input to determine which driver more accurately influences respiration at a specific site. The results indicated that SWC is often a superior predictor for $R_{eco}$ in grasslands or high-temperature sites, while $T_{air}$ remains dominant in colder forest ecosystems.
Results and Discussion
The DMDc model achieved Root Mean Square Error (RMSE) levels comparable to, and in some cases exceeding, the standard Nighttime (NT) and Daytime (DT) partitioning methods used in the FLUXNET community.
While simple models perform adequately for short-term "nowcasting," the inclusion of TDE was found to be crucial for maintaining accuracy in longer-term (e.g., two-week) forecasts. This highlights the importance of historical context in modeling the trajectory of biological systems.
Conclusion and Future Work
This study demonstrates that DMDc is a promising, physics-aware tool for analyzing the dynamic behavior of vegetation ecosystems. It provides a robust framework for understanding how different climatic drivers influence carbon loss across multi-temporal scales. Future research will focus on implementing this approach across entire seasons and years to capture long-term shifts in ecosystem dynamics and further refining the integration of historical data for enhanced forecasting.
Summary updated