基于Transformer和多序列特征的土壤含水率预测

Prediction of Soil Moisture Content Based on Transformer and Multi Sequence Feature

摘要:

土壤含水率（Soil moisture content，SMC）的准确预测在农业生产中至关重要。多时间序列多源遥感数据可提供多种特征的时间变化信息，但多时相多序列信息往往在SMC反演中得不到有效利用。Transformer网络在处理多序列特征中表现优越，本研究基于Transformer结构构建了提取SMC的深度回归模型，并将其与卷积神经网络回归（Convolutional neural network regression，CNNR）、长短期记忆网络（Long short-term memory，LSTM）回归、门控循环单元（Gated recurrent unit，GRU）回归进行比较。实验结果表明，使用长时间序列特征数据更有利于SMC预测；在利用5d的历史数据预测5d后的SMC时，Transformer回归相较于CNNR、LSTM和GRU，决定系数平均提升0.0953、0.0324、0.0336，均方根误差平均降低0.014、0.0026、0.0030cm3/cm3。对特征影响和中间隐藏特征的变化分析显示，为不同时刻特征分配合适的注意力更有利于预测SMC。

Abstract:

Accurate prediction of soil moisture content (SMC) is very important in agricultural production. Multi time series and multi-source remote sensing data can provide time change information with multiple characteristics, but multi temporal and multi sequence information is often not effectively used in SMC inversion. We hope to use multiple time series features to predict the change trend of SMC. Transformer network performed well in processing multiple sequence features. A deep regression model for SMC extraction was constructed based on Transformer structure, and it was compared with convolutional neural network regression (CNNR), long short-term memory (LSTM) regression, and gated current unit (GRU) regression. Multi source heterogeneous remote sensing data, including Sentinel 1, soil moisture active passive (SMAP), etc. were used as model inputs, and field measurement data were used as SMC reference values. The experimental results showed that the use of long time series feature data was more conducive to SMC prediction. When using the historical data of 5 days to predict SMC after 5 days, compared with CNNR, LSTM and GRU, the determination coefficient of Transformer regression was increased by 0.0953, 0.0324 and 0.0336 on average, and the root mean square error was decreased by 0.014cm3/cm3, 0.0026cm3/cm3 and 0.0030cm3/cm3 on average. The feature extraction and regression mechanism of the model were analyzed by quantifying the impact of input features on regression, the sequence changes of hidden features in the middle, and the output performance. The analysis of feature influence and the change of hidden features in the middle showed that allocating appropriate attention to features at different times was more conducive to predicting SMC.