Submitted to: Biological Systems Simulation Group Proceedings
Publication Type: Proceedings
Publication Acceptance Date: April 1, 2010
Publication Date: April 10, 2010
Citation: Thorp, K.R., Hunsaker, D.J., French, A.N. 2010. Assimilating Leaf Area Index Estimates from Remote Sensing into the Simulations of a Cropping Systems Model. Biological Systems Simulation Group Proceedings. Volume 1, Pages 30-31. University of Arizona:Maricopa Technical Abstract: Spatial extrapolation of cropping systems models for regional crop growth and water use assessment and farm-level precision management has been limited by the vast model input requirements and the model sensitivity to parameter uncertainty. Remote sensing has been proposed as a viable source of spatial information for guiding model simulations, but techniques for merging remote sensing with cropping systems models have not been rigorously explored. We developed and tested two techniques for assimilation of remotely sensed green leaf area index (GLAI) into the CSM-CROPSIM-CERES-Wheat model: one based on model ‘updating’ and the other based on model ‘forcing’. The ‘updating’ method adjusts the model only on the dates when GLAI observations are available. The ‘forcing’ method adjusts the model on a daily timestep using linear interpolation to compute GLAI between measurement dates. Assimilation of GLAI observations into the model simulation is more complex than simply overwriting the GLAI state variable, because the daily growth rate equations are fundamentally focused at the individual plant level, while GLAI is an area-based variable. After computing daily growth for individual plants, the plant component weights are computed on an area basis using the plant population parameter, and GLAI is computed from the more fundamental plant leaf area (PLA, cm2 plant-1) state variable according to: GLAI = (PLA-SENLA)x PLTPOP x 0.0001 (1) where SENLA (cm2 plant-1) is the total leaf area that has been senesced from the plant and PLTPOP (plants m-2) is the plant population. To drive the model based on remotely sensed GLAI estimates, the model was reprogrammed to complete the following basic steps after it finished daily growth rate calculations: 1. Read a file containing the remotely sensed GLAI observations 2. Compute (PLA-SENLA)sim as simulated by the model 3. Backcalculate (PLA-SENLA)obs by plugging the GLAI observation into equation 1 4. Compute the deficit plant leaf area: DEFICIT = (PLA-SENLA)obs –(PLA-SENLA)sim 5. Adjust the PLA state variable by DEFICIT These steps effectively adjust the PLA state variable in the model, such that GLAI is later computed (eq. 1) as the GLAI observed with remote sensing. Adjustment of the more fundamental PLA state variable insures that the effects of the data assimilation are not lost during the growth rate calculations at the individual plant level on the following timestep. The impact of data assimilation approaches on model performance was rigorously tested using measurements from two wheat irrigation scheduling experiments, conducted at Maricopa, Arizona during the winters of 2003-2004 and 2004-2005. These experiments provided canopy spectral reflectance information and measurements of canopy weight, wheat yield, and evapotranspiration (ET) under varying planting densities and nitrogen rates for testing the ability of the assimilation techniques to improve model simulations. Ground-based radiometric measurements were available over each treatment plot two to four times per week from emergence to harvest. A four-band, hand-held radiometer (model BX-100, Exotech, Inc., Gaithersburg, MD) was used to collect the remote sensing data. The instrument was equipped with 15° field-of-view optics and positioned at a nadir view angle approximately 1.5 to 2.0 m above the soil surface. Data collection occurred in the morning around the time of a 57° solar zenith angle. Frequent radiometric observations of a calibrated, 0.6 m2, 99% Spectralon reflectance panel (Labsphere, Inc., North Sutton, NH) were used to characterize solar irradiance throughout the data collection period. Canopy reflectance factors in the red (RED; 610 to 680 nm) and near-infrared (NIR; 790 to 890 nm) were computed as the ratio of the average canopy radiance over the corresponding time-interpolated value for solar irradiance. The NDVI was computed from reflectance factors using the well-known equation: NDVI = (NIR - red)/(NIR + red) (2) Canopy reflectance measurements were available for computing NDVI on 29 and 31 dates during the 2003-2004 and 2004-2005 wheat growing seasons, respectively. The NDVI measurements were subsequently used to estimate GLAI using an empirical method that incorporates information about the leaf angle distribution of the crop. These GLAI estimates were directly inserted into CSM-CROPSIM-CERES-Wheat as described above. Monte Carlo simulation methods were used to assess the performance of GLAI data assimilation in light of uncertainty in the model parameters that govern the GLAI simulation. Parameters that were varied using Monte Carlo techniques included the potential leaf area (LAVS and LARS), the laminar area to weight ratio of leaves (LAWRS and LAWR2), and the plant population (PLTPOP). When considering uncertainty in these parameters, assimilation of GLAI by ‘updating’ and by ‘forcing’ was able to reduce error between measured and simulated canopy weight and ET by 43.6% and 56.5% and by 45.0% and 51.6%, respectively, as compared to the stand-alone model. The assimilation techniques had greater difficulty improving wheat yield simulations, because simulated yield was more sensitive to parameters other than GLAI, especially in the 2004-2005 growing season. Assimilation of remotely sensed data into cropping systems models has potential to improve simulations of key model outputs, such as canopy weight and ET, but further efforts are warranted to explore and finetune techniques for merging these two technologies. In this work, we implemented a simple direct insertion approach to overwrite the GLAI state variable. Future efforts will test other data assimilation approaches, such as using a radiative transfer model as the mediator between remote sensing observations and crop model simulations. Also, we will explore statistical approaches, such as Kalman filtering, that aim to account for error in both the observed and simulated quantities when merging remote sensing data into the simulations.