欢迎访问《麦类作物学报》编辑部网站

为探讨利用时空建模的深度学习方法提高大区域冬小麦产量预测精度的可行性，从县级冬小麦产量预测角度出发，使用卷积神经网络(convolutional neural networks, CNN)从气候和土壤数据中提取与作物产量密切相关的特征数据，利用注意力机制（attention mechanism）捕捉特征数据之间的相互依赖性，最后将重新加权的特征与长短期记忆网络(long short-term memory network, LSTM)从年产量中捕获到的时间依赖性结合来预测县级冬小麦产量。结果表明，注意力机制模块能够有效地考虑到从CNN中提取的特征之间的相对重要性；模型 RMSE为686.82 kg·hm^-2，相较于支持向量机（SVR）、深度全连接（DFNN）和随机森林（RF）模型分别降低了43%、30%和67%，且R²在0.755以上，MAPE低于14.11%，预测精度均优于传统方法。这说明将注意力机制、CNN和LSTM结合建立的预测模型具有良好的泛化能力和空间平稳性，可用于大区域冬小麦产量预测。

英文摘要:

Utilizing deep learning methods for spatiotemporal modeling can effectively enhance the accuracy of large-scale winter wheat yield predictions. In this study, focusing on winter wheat yield prediction at the county-level, we proposed a county-level model (CNN-LSTM-Attention) that integrates attention mechanisms, convolutional neural networks (CNN), and long short-term memory networks (LSTM). The model emploied a CNN to extract features closely related to crop yield from climatd and soil data, utilized attention mechanisms to capture interdependencies among the extracted features, and combined the reweighted features with the temporal dependencies captured by the LSTM network to predict county-level winter wheat yield. The results indicated that the attention mechanism module effectively could take the relative importance of features extracted from CNN into consideration. The model achieved an RMSE of 686.82 kg·hm^-2, reducing by 43%, 30%, and 67% compared to support vector machines (SVR), deep fully connected neural networks (DFNN), and random forest (RF) models, respectively. Additionally, R² was above 0.755, and MAPE was below 14.11%, demonstrating superior performance compared to traditional methods. The model exhibited robust generalization capabilities and spatial stability.

查看全文查看/发表评论下载PDF阅读器

关闭