LAUSR

This study aims to explore the application of agricultural information classification combined with the Internet of Things technology in agricultural production and economic management in the context of deep learning (DL). The agricultural information classification system is built based on DL. In terms of experimental methods, qualitative and quantitative methods are used to compare the extractive forward and the generative reverse generation algorithm in the abstract generation method. Typical extraction methods that directly extract important sentences are relatively ineffective. The model parameter tuning and training of generative (regenerating new sentences) are difficult, and the effect does not reach the level of agricultural scientific research. Qualitative and quantitative evaluation practice proves that the effect is better. In the data training effect of Rice Digest, after nine days of iterative training for about 1400 steps, the loss value is 1.33, which is equal to the standard data set. In the data training effect of Wheat Digest, after nine days of iterative training for about 1400 steps, the loss value is 1.496, which is equal to the standard data set. Later, the existing original model is adjusted to fit the model of the agricultural dataset scale. Among them, the minimum learning rate is reduced from 0.01 to 0.003 to expand the learning rate drop rate. In the visual training comparison, when training to about 1400 steps, the cross-entropy loss function values are about 1.43 and 1.29, respectively. The curve smoothing factor is set to 0.97 to observe the overall change in target loss value. The results show that when the parameter settings remain unchanged, the loss value of the model starts to increase linearly at about 1400 steps, and the effect reaches the expected value. This study improves the relevance, completeness, and accuracy of information acquisition in the field of agricultural science and technology information and improves the utilization of agricultural information.