Abstract: To address the issues including high dimensionality of hyperspectral data and the influence of dimensionality reduction on prediction accuracy, an Attention mechanism Embed Bidirectional Recurrent Neural Networks (AE-BiRNN) model embedded with a dot product attention mechanism was developed to predict moisture content in tobacco strips. The AE-BiRNN model associates the spectral information of the front and back bands of the spectral sequence of moisture content in tobacco strips through the Bidirectional Gate Recurrent Unit (BiGRU), uses the dot product structure-based attention mechanism to assign different weight parameters to the bands according to the similarity information, and plays the role of characteristic wavelength selection. Training and verification were performed on data from 155 hyperspectral tobacco strip images 640 px×640 px in size. The results showed that: 1) The AE-BiRNN model could achieve satisfactory prediction accuracy in the absence of characteristic wavelength selection, and the introduction of the dot product attention mechanism significantly increased the convergence speed of the model. 2) The determination coefficient R_v² of AE-BiRNN for the regression prediction of the verification set was 0.940 4 with the root mean square error RMSEV of 0.014 71, which were obviously better than those of Partial Least Squares Regression (PLSR) and Random Forest (RF), Convolutional Neural Network LeNet and Multilayer Perceptron (MLP) methods. The method supports real-time monitoring of moisture content in tobacco strips and timely regulation of ordering cylinder parameters.