Deep Learning Based Corn Yield Estimation Method And Attribution Analysis

Crop yield estimation is one of the essential utilities of agricultural production simulation.Quantitative analysis of crop yield response to weather factors and its spatio-temporal distribution is beneficial for understanding agricultural production systems and dealing with climate change threats to agricultural production.Therefore,it is essential for achieving sustainable agricultural development and enhancing food security.Taking large-scale corn yield estimation as a research mission,this study has developed a deep learning based crop yield estimation model to address the challenge of insufficient extraction of the complex interaction mechanism between crop growth and weather factors in existing studies.The model has been applied to estimate the county-level corn yield in the US Corn Belt and project future climate change impact.This study realized the effective extraction of spatio-temporal characteristics of corn yield at large spatial scales and improved the accuracy and stability of corn yield estimation model.We proposed attribution analysis based on deep learning model to analyze the response of corn yield to weather fluctuations.We also evaluated future climate change impact on corn yield with CMIP6 future climate scenario dataset and deep learning model.This study provides new insights and analysis tools for large-scale crop yield estimation and future climate change impact assessment.The main contents and results of this study are as follows:(1)To address the problem of spatial heterogeneity of corn production environment at large scales,we demonstrated a method based on multi-source agricultural data and unsupervised clustering for large-scale corn planting environment stratification.Based on this method,783 counties in 9 states of the US Corn Belt were divided into multiple assumed homogeneous corn production areas,and linear regression models were constructed to test the yield estimation accuracy.Results showed that the stratification results effectively reflected the regional differences of corn production environments and corn yield in the US Corn Belt.The visualization results based on t-SNE dimensionality reduction showed the regional differences among data from a perspective of feature distribution.The linear models for corn yield estimation with consideration of the spatial stratification showed improvements compared to the model without stratification.The magnitude of accuracy improvements ranged from 3%to 19%.The results showed the effectiveness of stratification to improve the large-scale corn yield estimation and provide spatial information for the later construction of deep learning model for yield estimation.(2)A deep learning model named DeepCropNet(DCN)for crop yield estimation was developed to capture the complex interaction mechanism between corn and weather factors.The model extracted the time-series features during the corn growing period through long short-term memory neural network embedded with attention mechanism.The regional yield differences across space were learned based on multi-task learning.The model was applied to estimate the county-level corn yield in the US Corn Belt from 1981 to 2020.The results showed that DCN model accurately estimated the county-level corn yield in the US Corn Belt both in all years and stressful years with root mean squared error(RMSE)as 0.93 Mg ha-1 and 1.26 Mg ha-1,and R2 as 0.81 and 0.65,respectively.DCN model outperformed LASSO and random forest(RF),with overall accuracy improved by 11-19%,and the improvement was more obvious in stressful years(20-23%).Multi-task learning effectively improved the accuracy of DCN model in all regions,with an average improvement of 3%.The distribution of attention values indicated that DCN model captured the time-series cumulative influence of each weather factor on corn yield,and identified the silking stage as the key stage for corn yield estimation.(3)To address the problem of low interpretability of deep learning model for crop yield estimation,a yield-weather attribution analysis method based on deep learning model was developed to capture the key weather factors affecting corn yield and important stages.The spatio-temporal patterns of relationships between corn yield and weather factors and their changing trends were analyzed.The results showed that the DCN model captured the nonlinear relationships between the corn yield and fluctuation of weather factors and the thresholds of weather stress that cause yield reduction.The DCN model identified the important factors and key growth stages that affect corn yield:PRCP has a negative impact on sowing and emergence,with an average yield reduction of-0.05 Mg ha-1,and both KDD and VPD had high influence on the silking stage,with average yield reductions as-0.15 Mg ha-1 and-0.09Mg ha-1,respectively.The DCN model could effectively extract the influence of different weather stress on corn yield,as well as the spatial differences of corn yield loss caused by each weather factor.The DCN model revealed that the sensitivity of corn yield to weather factors significantly reduced in the US Corn Belt from 1981 to 2020.Compared with 1981-2000,the magnitudes of sensitivity in 2001-2020 were reduced by 50%,35%,and 75%for GDD,KDD,and PRCP,respectively.(4)To address the uncertainty of future climate impact on corn yield,we used DCN model with CMIP6 future climate scenario simulation dataset to analyze the climate change trends and their impact on corn yield in the US Corn Belt from 2021 to 2050.The results showed that the impact of future climate change on corn yield in the US Corn Belt would be mainly negative,with an increasing trend of yield reduction magnitude and influence area.By 2050,GDD in the growing season could increase by 3.6%,KDD could increase by 51.6%,PRCP could increase by only 8.3%,and VPD could increase by 4.3%under the low-carbon emission scenario(ssp126).These climate changes would lead to an average yield reduction of 0.70 Mg ha-1.Under the high-carbon emission scenario(ssp585),GDD could increase by 10.8%,KDD could increase by 157.8%,PRCP could decrease by 4.7%,and VPD could increase by 23.5%.The average corn yield would reduce by 0.85 Mg ha-1.Compared to the historical yield average level,climate change is possible to cause the yield reduction with magnitudes of 8.7%and 10.5%under the low-carbon emission and high-carbon emission scenarios,respectively.