基于多源数据和LSTM模型的县域冬小麦估产

Winter wheat yield estimation at county-scale based on the multi-source data and LSTM model

及时准确地估计区域冬小麦产量对维护国家粮食安全和农业可持续发展具有重要意义。本研究利用中国冬小麦主产区2001—2018年的遥感数据、气象数据和县域产量,构建基于长短期记忆(Long Short-term memory,LSTM)估产模型,并与传统随机森林(Random Forest,RF)、支持向量机(Support Vector Machine,SVM)和决策树(Decision Tree,DT)模型对比,研究不同模型的估产性能,分析不同特征对模型精度的影响,评估模型的提前预测能力。研究结果表明:1)基于全部数据的LSTM模型精度最高,平均R2为0.853,平均NRMSE为7.22%。与DT、RF和SVR模型相比,LSTM模型将R2提高了0.324、0.088和0.028;2)光合作用相关的地面下行长波辐射(R20.737)、近地面气温(R20.747)、地面下行短波辐射(R20.735)和降水率(R20.681)超过了其他单一特征的估产能力,在单一特征的基础上增加特征的数量将进一步提高估产的准确性。气象数据、波段反射率和植被指数对估产的贡献依次降低,当同时使用这三种数据时估产准确性最高(R20.866、NRMSE7.00%)。3)小麦生长周期从10月8日至次年6月10日,8 d一个时相,合计32个时相数据,基于三种数据源的LSTM模型预测产量的能力在1~6时相增加,在7~19时相趋于平稳,在20~29时相再次上升,30~32时相基本保持稳定不再增加。当使用前29个时相的数据时,LSTM模型可以提前24 d获得最大的产量预测精度(R20.873、NRMSE 6.90%)。本研究提出的方法不仅估产精度较高,而且能够实现提前预测产量,可为农业管理和农业经济活动提供高效可靠的大面积冬小麦估产途径。

Timely and accurately estimating regional winter wheat yield is critical for maintaining national food security and sustainable agricultural development.In this study,we used remote sensing,meteorological data,and county yields from 2001 to 2018 in China’s main winter wheat producing areas to build a yield estimation model based on the Long Short Term Memory Networks(LSTM),compare the yield estimation performance of different models with Random Forest(RF),Support Vector Machine(SVR),and Decision Tree(DT)models,analyze the effects of different feature combinations on model accuracy,and evaluate the advance prediction ability of the models.The results show that:1)the model based on all data has the highest accuracy with an average R2 of 0.853 and an average NRMSE of 7.22%.When compared with the DT,RF and SVR models,the LSTM model improves R2 by 0.324,0.088 and 0.028;2)Photosynthesis-related surface downward longwave radiation(lrad)(R2,0.737),nstantaneous near surface air temperature(temp)(R2,0.747),surface downward shortwave radiation(srad)(R2,0.735)and precipitation rate(prec)(R2,0.681)surpass other single features in yield estimation.Adding more features to a single feature would increase yield estimation accuracy.The contribution of meteorological data,band reflectance,and vegetation index to yield estimation decrease in order.Using three data sets,the accuracy of yield estimation is highest(R20.866,NRMSE 7.00%).3)Winter wheat fertility data lasts from October 8 to June 10 of the following year,with one time-phase every eight days and a total of 32 time-phase data.The ability of the LSTM model to predict yield increases in time phases 1~6,plateaus in time phases 7~19,increases again in time phases 20~29 and remains steady without further increase in time phases 30~32.The LSTM model based on three data sources achieves the highest yield forecast accuracy(R20.873,NRMSE 6.90%)24 days earlier using data from the first 29 time phases.The method in this study has high yield estimation accuracy and can achieve early yield prediction,which can provide an efficient and reliable way to estimate winter wheat yield in large areas for agricultural management and economic activities.