欢迎访问《麦类作物学报》编辑部网站

为探究利用遥感技术和深度学习方法实现大范围内冬小麦产量早期准确预测的可行性，以中国豫北地区为研究区域，将2013—2022年冬小麦返青到成熟期内的归一化差值植被指数、气温、日照时数和降水量作为特征参数，结合县域单产数据构造旬尺度的冬小麦估产数据集，利用卷积神经网络（convolutional neural networks，CNN）-长短期记忆（long short-term memory，LSTM）构建可提前估产的冬小麦混合神经网络（CNN-LSTM）模型，并通过对不同生育时期CNN-LSTM估产模型逐个训练，比较其早期估产能力，以筛选最优估产模型。结果表明，返青—成熟期CNN-LSTM模型具有较强的鲁棒性，五折交叉验证的平均决定系数（R²）和平均均方根误差（RMSE）分别为0.86和402.76 kg·hm^-2，较对应的LSTM模型估产精度大幅提升，平均R²和RMSE分别升高0.18和降低188.05 kg·hm^-2。返青-成熟期CNN-LSTM估产模型稳定性和泛化能力均较优，2022年测试集估产的R²、RMSE和平均绝对百分比误差（MAPE）分别为0.91、337.25 kg·hm^-2和4.25%。在豫北32个县（区）中，75%的县（区）产量预测相对误差小于5%，21.875%的县（区）产量预测相对误差在5%~10%之间，仅林州市一个县（区）产量预测相对误差在15%~20%之间。CNN-LSTM估产模型具有早期估产能力，抽穗-扬花期CNN-LSTM估产模型R²和RMSE分别为0.83和431.98 kg·hm^-2，该模型有能力提前1个月实现冬小麦产量的预测。因此，利用CNN-LSTM模型可建立稳定且准确的冬小麦估产模型，从而为农业决策提供有力支持。

英文摘要:

To explore the feasibility of using remote sensing technology and deep learning methods to achieve early and accurate prediction of winter wheat yield on a large scale, the northern Henan region of China was taken as the research area. Using the normalized difference vegetation index, temperature, sunshine hours, precipitation and county-level yield during the period from re-greening to maturity of winter wheat from 2013 to 2022 as characteristic parameters. The Long Short-term memory(LSTM)-Convolutional Neural Networks(CNN) was used to estimate the winter wheat yield in the northern Henan region. By training the CNN-LSTM yield estimation models at different growth stages one by one, their early yield estimation ability was compared. The results showed that the CNN-LSTM yield estimation model in re-greening-maturity stages had relatively strong robustness. The average coefficient of determination(R²) in the five-fold cross-validation was 0.86, and the average root mean square error(RMSE) was 402.76 kg·hm^-2. There was a significant improvement in the yield estimation accuracy compared with the corresponding LSTM model, with the average R² increased by 0.18, and the average RMSE decreased by 188.05 kg·hm^-2. The CNN-LSTM yield estimation model in re-greening-maturity stages was a model with high stability and strong generalization ability. The R², RMSE, and mean absolute percentage error(MAPE) for yield estimation of the 2022 test set were 0.91, 337.25 kg·hm^-2, and 4.25%, respectively. Among the 32 counties(districts) in northern Henan, 75% of them had a relative error less than 5% in yield prediction, of which 21.875% had a relative error of 5% to 10% in yield prediction, and only one county(district) in Linzhou City had a relative error of 15% to 20% in yield prediction. The CNN-LSTM yield estimation model had early yield estimation ability. The R² and RMSE of the CNN-LSTM yield estimation model during the heading-anthesis stage were 0.83 and 431.98 kg·hm^-2, respectively. This model had the ability to predict winter wheat yield one month in advance. Therefore, the CNN-LSTM could be used to establish a stable and accurate winter wheat yield estimation model, thus providing strong support for agricultural decision-making.

查看全文查看/发表评论下载PDF阅读器

关闭