Predicting Winter Wheat Yield And Quality By Integrating Of Remote-sensing Data And The Weather Forecast Data Into The DSSAT Model

With the development of economy, society, technology, and the improvement of life quality, attention to food quality and safety as grain yield has increased concern in recent years. Chinese agriculture products face not only enormous competitive pressures at the international level, but also strongly impact from foreign agriculture products at domestic market since accession to the WTO. Low quality of Chinese agriculture products become increasingly. Industrialization production with high quality and effect for winter has become a highlight work in the processing of Chinese crop production. Therefore, early-stage predictive information on crop yield and quality fast and instantaneously at regional and national scales is essential for effective management, harvest and stockpile, and market price. Integration of remote-sensing with instantaneous and spatial continuity and crop-growth simulation models with temporal continuity has became an effective approach for estimating crop grain yield and quality across wide regions. In this study, the objective of this work is to predict wheat yield and grain protein content (GPC), especially the latter, by coupling the DSSAT model and proximal remote sensing. The DSSAT (Decision Support System for Agrotechnology Transfer) model was used to wheat yield and GPC prediction. The main work includes global sensitivity analysis of DSSAT model, parameter calibration, biophysical parameters estimation using remote sensing, data assimilation, and weather forecast. Some important conclusions can be drawn as follows:(1) Parameter calibration is the prerequisite step before local applications of crop growth models, while sensitivity analysis is the primary task to identify the importance of various parameters. The Extended Fourier Amplitude Sensitivity Test (EFAST) approach was used to evaluate the sensitivity of DSSAT output responses of interest to genotype parameters and soil parameters. The outputs for sensitivity analysis included grain yield and quality at maturity, and the critical processed variables [leaf area index (LAI), aboveground biomass (AGB), and aboveground nitrogen accumulation (AGN)] at the whole growth stages. The key results indicated that the sensitivity analysis of some processed variables is performed different and revealed some differences on the time series. The sensitivity parameters of grain yield and grain quality are also different, and interactions between parameters to grain quality are more than that to grain yield. Some sensitive parameters [PHINT, LSPHS, LAIS, SALB, VEFF, TDFAC, P1D and RDGS] to processed variables were not showed sensitivity to yield or GPC. However, these parameters cannot be neglected in parameter calibration.(2) The DSSAT model was calibrated with the generalized likelihood uncertainty estimation (GLUE) method integrating of systematic parameters calibration process and the above sensitive parameters. The results showed that the GLUE method can be used to accurately estimate the genotype parameters of wheat and the simulated LAI, AGB, AGN were close the measured values. The root mean square error (RMSE), normalized RMSE, model efficiency (E), and index of agreement (d) values were 0.42,0.17,0.64 and 0.87 for LAI,1.77 ton ha-1,0.28, 0.78 and 0.92 for AGB,33.04 kg ha-1,0.28,0.41 and 0.79 for AGN, respectively. Compared with simulated LAI and AGN,the simulated AGB performed best. A good consistency was found between predicted and measured yield (RMSE=0.23 ton ha-1, NRMSE=0.05, E=0.77, and d 0.94) and predicted and measured GPC (RMSE=1.91%, NRMSE=0.12, E=-19.85, d =0.28), respectively. The simulation accuracy of yield was better than that of GPC. In general, the DSSAT model was proved to be a useful decision-making and predicting tool for winter wheat production in Beijing region.(3) A data assimilation approach using a particle swarm optimization (PSO) algorithm was developed to integrate remotely sensed data into the DSSAT model with AGN as state variable for estimating yield and GPC of winter wheat. Our results showed that the normalized difference red edge index (NDRE) produced the most accurate selection of spectral indices for estimating AGN, with R2 and RMSE values of 0.663 and 34.05 kg ha-1, respectively. A data assimilation method (R2=0.729 and RMSE=32.02 kg ha-1) performed better than the spectral indices method for estimation of canopy N accumulation. Simulation values by the data assimilation method was very consistent with the measured values, with R2 and RMSE values of 0.711 and 0.63 ton ha-1 for yield, and 0.367 and 1.95%, respectively. Estimating grain protein content by gluten type could improve the estimation accuracy, with R2 and RMSE of 0.519 and 1.53%, respectively. Our study showed that estimating wheat yield, and especially GPC, could be successfully accomplished by assimilating remotely sensed data into the DSSAT model.(4) Two assimilation variables (derived from a hyperspectral sensor), called LAI and AGN, were jointly used to calibrate the sensitive parameters and initial states of the DSSAT model, to improve simulated output of the yield and GPC of winter wheat. The results showed that the modified simple ratio (MSR) and normalized difference red edge (NDRE) better estimated LAI and AGN compared with the other possible vegetation indices, with R2 and RMSE of 0.829 and 0.598, and 0.794 and 37.75 kg ha-1, respectively. The R2 and RMSE values, respectively, of the regression between the simulated (using the two assimilation variables method) and measured LAI were 0.828 and 0.494, and for AGN were 0.808 and 20.26 kg ha-1. These two assimilation variables resulted in yield and GPC estimates of winter wheat with a high precision and R2 and RMSE values of 0.698 and 0.726 ton ha-1, and 0.758 and 1.16%, respectively. This study provides a more robust method for estimating yield and GPC of winter wheat based on the integration of the DSSAT crop model and remote sensing data.(5) Predicting yield and GPC of winter wheat at different time nodes were analyzed, and the best predicted time node for yield and GPC was determined, respectively. The predicted weather includes median weather and three extreme weather conditions, which are extreme light, extreme temperature, and extreme precipitation. The results showed that there are relative large deviations between predicted yield and GPC and simulated yield and GPC with actual weather data since the unknown weather data are too much at the early growth stage, and the predicted yield and GPC fluctuated greatly. The best predicted node for yield in this study was 21 May because of the consistency between predicted and simulated values and little effect on later extreme weather. For the same reason, the best predicted node for GPC in this study was 31 May.(6) Integration of remote-sensing with data assimilation approach and weather forecast into DSSAT modelwas carried out for predicting wheat yield and GPC at regional. The results showed that MSR and green normalized difference vegetation index (GNDVI) better estimated LAI and AGN compared with the other possible vegetation indices, which are based on transformed multispectral data from hyperspectral data calculated by spectral response function. The assimilation efficiency could be improved by optimizing PSO set, upscaling process, and parallel computing method. There was a highly significant level between predicted and measured yield, and a significant level between predicted and measured GPC. A good consistency was found between the predicted yield and measure yield and the predicted GPC and measured GPC. The results indicated that it was feasible to integrate remote-sensing with data assimilation method and weather forecast by DSSAT model for predicting wheat yield and GPC.