Data Assimilation of Solar-Induced Chlorophyll Fluorescence Improves Gross Primary Production Simulation by a Process-Based VISIT-SIF Model in a Rice Paddy

Abstract

Simulating gross primary production (GPP) is a key objective of terrestrial ecosystem models, and many studies have shown that solar-induced chlorophyll fluorescence (SIF) is a reliable proxy for GPP. This study combines SIF data with a process-based vegetation integrative simulator for trace gases (VISIT-SIF) model to enhance GPP simulations in the Mase rice paddy field in Tsukuba, Japan. Using data assimilation techniques (Bayesian optimization) with both ground-based SIF data and satellite-derived SIF (CSIF) (both from 29 June 2019 to 10 September 2020), we optimized key model parameters and improved the simulation of GPP. Sensitivity analysis via SHapley Additive exPlanations (SHAP) revealed that the maximum rate of carboxylation (Vcamx)-related parameters significantly influence GPP, while the absorbed photosynthetically active radiation (APAR)-related parameters are more critical for SIF modeling. Model optimization resulted in substantial performance improvements, particularly in simulating GPP at half-hourly and daily scales. For half-hourly results, the R² values of SIF improved from 0.37 to 0.60, and the relative error decreased from 124.21% to 63.39%, but the model went from underestimation to overestimation; for GPP, R² values improved from 0.47 to 0.68, relative error decreased from 150.00% to 47.85%, and the model's tendency to underestimate has been mitigated. At the daily scale, model simulations demonstrated higher R² values and lower relative errors than observations. Using the CSIF data set also improved the model but was less effective than densely measured ground SIF. Further, we explored the relationship between SIF and GPP on half-hourly scale, daily scale, and weekly scale and found that the larger the time scale, the stronger the linear relationship of SIF-GPP. Overall, using SIF as a proxy for GPP and optimizing key parameters through data assimilation significantly enhanced the simulation accuracy of the VISIT model. However, challenges remain, such as model biases under cloudy conditions and SIF overestimation during specific stages. This research demonstrates the value of assimilating SIF data into the VISIT model and highlights the potential of satellite-derived SIF for improving GPP estimations, though at a small scale.