Crop Growth And Yield Predictions And Uncertainty Analysis Under Climate Change

The impact of quantified climate factors on crop productivity is a significant scientific issue that has attracted widespread attention both domestically and internationally.In recent years,extreme weather events such as heavy rain,drought,and high temperatures have become more frequent and intense,and the level of climate risk is on the rise.Therefore,timely and accurate predictions of yield changes and evaluations of the impact of quantified climate changes on yield are of great importance for developing adaptation measures and ensuring global food security.Process-based crop models are effective tools for explaining the interactions between local environmental variables,crop genotypes,and management practices.However,the simplification of certain processes and uncertainty surrounding certain parameters in these models may result in inaccurate predictions.Furthermore,crop models often require a large amount of local observation data for model calibration and significant computational resources.Machine learning algorithms based on statistical models are easy to handle and relatively easy to compute.They can capture the nonlinear relationship between environmental variables and yield with high accuracy.However,these algorithms cannot fully consider the crop growth process(such as the CO2 fertilization effect)and have significant limitations in analyzing the impact of future climate change on yield.Therefore,we need to consider using different methods to comprehensively and accurately explain the relationship between climate change and crop productivity,and reasonably quantify the uncertainty of the analysis.This will help us develop more scientific and effective adaptation measures to ensure global food security.In this study,we use different methods to simulate and predict yield changes and the impact of climate change on yield at different scales.Firstly,we use machine learning to predict and simulate yield and determine the impact of different variables on yield.Then,we combine machine learning with grid crop models to improve the accuracy of the model and effectively constrain the model’s uncertainty.Based on this,we propose a new framework to explore yield prediction and uncertainty and study the impact of model combination and the number of models on uncertainty.Finally,we establish a remote correlation between large-scale circulation and yield and analyze the impact of climate variability on food security.This study provides valuable reference for decision-makers to develop scientific management measures and helps to improve understanding of the relationship between climate change and food security.The main conclusions of this paper are as follows:(1)Developing machine learning models to predict wheat yield and reveal their non-linear response.We have developed machine learning-based models to accurately predict wheat yield by incorporating multi-source environmental data,such as soil properties,climate,and vegetation indices.Our research found that the Random Forest(RF)model outperformed the Support Vector Machine(SVM)model in predicting wheat yield.Additionally,we discovered that the RF model using Near-Infrared Vegetation Index(NIRv)had a slightly better prediction accuracy(R~2=0.74;RMSE=758 kg/ha)than using Enhanced Vegetation Index(EVI)(R~2=0.73;RMSE=762 kg/ha)or Normalized Difference Vegetation Index(NDVI)(R~2=0.73;RMSE=770 kg/ha).Our findings also showed that vegetation-based indices had the most significant impact on wheat yields compared to other environmental covariates.Based on our RF＿M5 model,we identified that NIRv during T3was the most critical predictor for determining crop yield.Additionally,we identified that several environmental factors such as Rad(T2 and T3),RHum(T1),SOC,Ws(T1–T4),R5(T1–T2),and Frost(T2)were the main limiting factors for wheat yield.Interestingly,drought had a relatively small contribution to yield change due to the use of irrigation during winter wheat production in our study.We constructed Partial Dependence Plot(PDP)plots to explain how different predictors at each developmental period affected wheat yield based on RF models.We discovered that NIRv showed both linear and nonlinear relationships with wheat yield,while wheat yield had threshold-like responses to other environmental variables.These PDP results can help better understand how factors limit wheat yield.Our findings demonstrate the potential of using NIRv for yield prediction,and our modelling approach can be applied globally in other regions using publicly available data.In the future,incorporating process-based crop models or newly developed vegetation indices could further enhance our yield prediction model.(2)Crop yield forecasting and associated optimum lead time analysis based on multi-source environmental data.We have developed yield forecasting systems for three major crops using a machine learning method that using multi-source environmental variables and field trial data from diverse locations across China.Our findings indicate that our machine learning approach driven by multi-source environmental data can provide satisfactory crop yield forecasts,with performance comparable to that reported in previous studies.For winter wheat,we were able to predict yields with a correlation coefficient(r)of0.81-0.85 and normalized root-mean-square error(n RMSE)of 10.5–11.4%around one to three months prior to harvest.For spring maize,summer maize,early rice,mid rice,and late rice,we were able to predict yields with r ranging from 0.71-0.82 and n RMSE ranging from 7.4-17.9%,one to two months before harvest.Our machine learning model provides useful information for farmers and policymakers to reduce yield loss before the end of the growing season.Furthermore,we identified the key predictors that influence the yields of wheat,maize,and rice.For winter wheat,we found that solar radiation and vegetation indices(especially during the jointing to milk development stages)were the main predictors,while for spring maize,vegetation indices(throughout the growing season)and drought(especially during the emergence to tasseling stages)were the most important predictors.Soil moisture(throughout the growing season)was the dominant predictor for summer maize,late rice,and mid rice,while precipitation(especially during booting to heading stages)was the main predictor for early rice.Our future work will involve the development of a hybrid approach that combines a biophysical model with our machine learning technique to further improve the accuracy of crop yield forecasts.(3)Project crop yield changes and uncertainty analysis under future climate change using multi-model ensemble approaches.Linked climate and crop simulation models are widely used to assess the impact of climate change on agriculture.However,it is unclear how ensemble configurations(model composition and size)influence crop yield projections and uncertainty.Here,we investigate the influences of ensemble configurations on crop yield projections and modelling uncertainty from Global Gridded Crop Models(GGCMs)and Global Climate Models(GCMs)under future climate change.We found that the projections of crop yields and their related uncertainty are subject to variations depending on different ensemble configurations and exhibit regional specificity,particularly for crops such as wheat and soybean that are sensitive to ensemble members.Ignoring this difference could result in an underestimation of the impact of climate change on crop yields.Furthermore,we found that using approximately six GGCMs and 10 GCMs was sufficient to determine the modelling uncertainty across the nine crop models and 32climate models we considered,while a cluster-based selection of 3–4 GGCMs can effectively represent the full ensemble.These results highlight the importance of considering different ensemble configurations to better project crop yields and effectively utilize multiple models for specific applications.(4)Incorporating machine learning and crop model to predict crop yields and constraining the overall uncertainty.This work presents a hybrid model that combines machine learning and crop models to predict maize and soybean yields while considering the impact of extreme climate events and crop pests and diseases(CPD).We found that our approach substantially increases the model performance compared to using only a crop model.The R2 values for maize and soybean were 0.05-0.32 and 0.05-0.32 for the crop model,and 0.37-0.58 and 0.2-0.35 for the hybrid model,respectively.The n RMSE was around 0.2 for both crops in the hybrid model.Our analysis revealed that CPD,heat,and drought are the main predictors for soybean,while cold days,CPD,and drought are the main predictors for maize.Our hybrid model reduced the uncertainty around 48-67%.During T1(2040-2069),the main source of uncertainty for maize was the crop model,while for soybean it was the general circulation model(GCM).During T2(2070-2099),the shared socioeconomic pathway(SSP)became the main source of uncertainty.Our findings suggest that a hybrid approach that combines machine learning and crop models is an effective way to predict crop yields while considering the impact of various factors such as extreme climate events and CPD.This method can provide valuable insights for farmers,policymakers,and other stakeholders in the agricultural industry.(5)Climate drivers impact on global food security under future climate change.In this section,we first identify and compare the primary climate drivers affecting crop yields and their changes under climate change.Secondly,we robustly estimate the global and regional sensitivity of crop yields to climate drivers and quantify yield changes during strong oscillation phases in historical and future periods.Thirdly,we assess the significant areas impacted by strong oscillation phases.Finally,we analyze the contribution of different sources of uncertainty in the impact of climate drivers on global crop yields.For maize and wheat,ENSO is the dominant driver in Europe and Northern Asia,while IOD is the primary driver in Africa.However,NAO is expected to be the main driver influencing crop yields in most areas of the northern hemisphere from 2000 to 2099.Compared to historical data,NAO has a large impact on soybean and rice yields.During strong negative NAO and positive ENSO phases,maize,soybean,and rice yields decrease significantly.In contrast,soybean and rice yields increase during strong negative ENSO phases.During strong positive IOD phases,wheat and soybean yields decrease significantly in most areas.Extreme NAO significantly affects rice and soybean yields.We also analyze the uncertainty in this study and find that GCM-induced variance is higher than SSP and GGCM.However,the uncertainty contribution of NAO to SSP was greater than that of GCM in rice and maize.Our study improves understanding of climate variability and its impact on global breadbaskets under climate change,with potential implications for improving the resilience of the global food system.