Rice yield prediction based on LSTM and GRU

Rice yield prediction is a vital problem in the national agriculture and economy. The development of deep learning overcomes the obstacles of traditional machine learning and shows superior performance in solving complicated problems. Especially for natural language processing (NLP) models such as LSTM and GRU, these models outperform the time series data, thus having great potential for complex agricultural spatiotemporal data with high dimensionality and nonlinearity. However, there is little discussion about performance of these two models in rice yield prediction. In this article, we adopted two popular NLP models to build and test 12 different model frameworks based on optimal hyperparameter configurations. And we compared model depth as well as bidirectional setting on the rice yield prediction by observing the performance of MSE losses throughout the training process. The results illustrated that both simple and complex models had outstanding fitting for small-sample training, and the depth and direction of the models did not significantly impact the performance of the experiment. But the complex model notably increases the training cost and decreases the convergence rate, implying that it’s not necessarily suitable for time-series problems with small-sample data. Further, the results could provide insights into a deep learning framework construction and hyperparameter selection for subsequent studies with comparable characteristics.