融合迁移学习的土壤湿度预测时空模型

Spatio-temporal model of soil moisture prediction integrated with transfer learning

针对土壤湿度观测数据量过少导致模型出现过拟合而影响预测精度的问题,本文提出了融合迁移学习的土壤湿度预测时空模型。首先,将ERA5-land数据集作为源域。然后,通过三维卷积层提取土壤湿度滞后时刻的空间特征,并融入长短期记忆网络提取其时间特征,对网络模型进行预训练。最后,以微调方式在SMAP数据集中调整网络参数,进而预测未来土壤湿度。实验结果表明,本文提出的时空深度学习模型相对于卷积神经网络、长短期记忆网络和PredRNN时空预测模型预测精度更高,同时通过迁移学习方法可以进一步提升模型的预测精度。

Using the deep learning methods can solve the model over-fitting caused by less observation data, and improve the prediction accuracy. This paper proposes spatio-temporal model of soil moisture prediction integrated with transfer learning. Firstly, the EAR5-land dataset is used as the source model.Then three-dimensional layer convolution is used to extract the spatial characteristics of the lag time of the soil moisture, and the long short-time memory network is integrated to extract the temporal characteristics.Third, the network model is pre-trained. Finally, the fine-tune method is applied to adjust the network parameters in the SMAP dataset for soil moisture prediction. The experimental results show that the proposed model has the better prediction results than the convolutional neural network, long short-term memory network and PredRNN. Meanwhile the method of transfer learning can improve the prediction accuracy.