Data synthesis in the Community Land Model for ecosystem simulation

• A data synthesis model is proposed to generate geosystem data in climate simulation.
• The feature extraction method is customizable for different physical properties in time and frequency domain.
• The physical relationships between geosystem variables are discovered and modeled in different feature space.

Though many ecosystem states are physically observable, the number of measured variables is limited owning to the constraints of practical environments and onsite sensors. It is therefore beneficial to only measure fundamental variables that determine the behavior of the whole ecosystem, and to simulate other variables with the measured ones. This paper proposes an approach to extract fundamental variables from simulated or observed ecosystem data, and to synthesize the other variables using the fundamental variables. Because the relation of variables in the ecosystem depends on sampling time and frequencies, a region of interest (ROI) is determined using a sliding window on time series with a predefined sampling point and frequency. Within each ROI, system variables are clustered in accordance with a group of selective features by a combination of Affinity Propagation and k-Nearest-Neighbor. In each cluster, the unobserved variables are synthesized from selected fundamental variables using a linear fitting model with ARIMA errors. In the experiment, we studied the performance of variable clustering and data synthesis under a community-land-model based simulation platform. The performance of data synthesis is evaluated by data fitting errors in prediction and forecasting, and the change of system dynamics when synthesized data are in the loop. The experiment proves the high accuracy of the proposed approach in time-series analysis and synthesis for ecosystem simulation.