Abstract: A county-level rice yield prediction model （CNN-BiLSTM-RA） was proposed, which integrated convolutional neural network （CNN）, bidirectional long short term memory network （BiLSTM）, and residual attention （RA） mechanism, key spatial features were effectively extracted from county-level rice meteorological data through CNN layers, the dynamic changes of time series data were deeply analyzed using BiLSTM layers, and RA mechanism was introduced to enhance the recognition and capture of key features in meteorological data. Using historical rice yield and meteorological data from 81 counties in Guangxi from 2015 to 2017 as samples, the prediction accuracy and effectiveness of the CNN-BiLSTM-RA model were compared with CNN, TRANSFORMER, BiLSTM, CNN-BiLSTM, and BiLSTM-RA models. The results showed that the R², MAE, RMSE, and MAPE of the CNN-BiLSTM-RA model were 0.986 1, 0.121 9, 0.224 8, and 0.864 8, respectively, indicating a high degree of fit between the predicted and actual values of the model. The CNN-BiLSTM-RA model fully utilized the spatial feature extraction ability of CNN, the time series data analysis advantages of BiLSTM, and the RA mechanism’s ability to enhance key feature capture. It was a new method suitable for high-precision prediction of rice yield in counties.

Key words: rice yield prediction, convolutional neural network, bidirectional long short term memory network, residual attention