Multi-assimilation methods based on AquaCrop model and remote sensing data

Abstract: Data assimilation has been known a promising approach for crop growth processing monitoring and yield estimation. In order to analyze the assimilation accuracy and efficiency of above ground biomass (AGB), canopy cover (CC) and yield of crop using AquaCrop crop growth model, which is an FAO crop model to simulate yield response to water, the field experiments were conducted during the 2012?2013, 2013?2014, and 2014?2015 winter wheat growing seasons at the National Precision Agriculture Demonstration Research Base in Beijing, China, and data were obtained in the jointing stage, flag leaf stage, anthesis stage and filling stage, respectively. Combined with ground remote sensing data, the AGB, CC and yield were simulated using the AquaCrop model under different water and nitrogen treatments with 3 data assimilation algorithms, which included particle swarm optimization (PSO), simulated annealing (SA) and shuffled complex evolution (SCE-UA). Both AGB and CC were used as state variables in the 3 algorithms. The computation efficiency and assimilation results of these algorithms were compared. The results showed as follows: 1) Among the 3 assimilation algorithms, the efficiency of SCE-UA assimilation algorithm was the highest (833 s), while the SA was the lowest (1 433 s). 2) Under different growth stages, the accuracy of AGB reduced with the growth of winter wheat. The simulation values of AGB were underestimated in jointing and flag leaf stages while overestimated in anthesis and filling stages. The simulation values of CC were overestimated in the jointing and flag leaf stages while underestimated in anthesis and filling stages. The ranges of RMSE (root mean square error), consistency index and MBE between the assimilated and measured AGB and CC by SCE-UA method were 0.57-1.92 t/hm2, 0.90-1.00 and -11.8%-18.5%, and 6.6%-12.1%, 0.94-1.00 and -3.7%-9.5%, respectively. 3) Under 3 different water treatments (W0 (rainfed), W1 (normal irrigation) and W2 (over irrigation)), the AGB and yield values were underestimated. The CC values were overestimated under W0 treatment and underestimated under W1 and W2 treatments. The ranges of RMSE, consistency index and MBE between the assimilated and measured AGB, CC and yield by SCE-UA method were 0.93-1.43 t/hm2, 0.98?1.00 and -12.5%?1.3%, and 8.0%?15.5%, 0.89?1.00 and 0.3%?18.4%, and 0.64?0.85 t/hm2, 0.84?0.97 and -12.5%?-5.7%, respectively. 4) Under 4 nitrogen fertilizer treatments (N1 (no fertilization), N2 (half fertilization), N3 (normal fertilization) and N4 (over fertilization)), the AGB values reduced with the increasing of the amount of nitrogen fertilizer, the CC values were overestimated and the yield values were underestimated. The ranges of RMSE, consistency index and MBE between the assimilated and measured AGB, CC and yield by SCE-UA method were 0.98-1.69 t/hm2, 0.98?1.00 and -10.1%?6.0%, 8.9%?11.4%, 0.97-1.00 and -0.8%?8.7%, 0.60?0.96 t/hm2, 0.90?1.00 and -9.6%?-2.6%, respectively. For the 3 assimilation algorithms, the SCE-UA algorithm produced a higher estimation accuracy for the total AGB, CC and yield than the PSO algorithm and SA algorithm. The results indicate that the PSO, SA and SCE-UA can well simulate winter wheat AGB, CC and yield, and amongst them the SCE-UA algorithm performs the best in terms of computation efficiency and accuracy.