Prediction of maize growth stages based on deep learning

Abstract

An accurate forecast of daily meteorological factors throughout the year is not only of great significance for the study of climate trends in a certain area but also enables the prediction of crop growth stages. Moreover, the prediction of crop growth stages is related to the scheduling of planting and tillage, the determination of machine harvest time, and the prediction of crop yield. However, highly complex dynamics cause large volatility in meteorological factors, so it is very challenging to predict the crop growth stage accurately, based on weather data. To solve this problem, we propose a data-driven encoder-decoder model, using long short-term memory (LSTM) and convolutional LSTM (ConvLSTM), which can be applied to forecast daily sunshine duration, cumulative precipitation, and average temperature for the coming year. To further test the performance of the ConvLSTM-based model, it is compared with the conventional LSTM encoder-decoder model and the convolutional neural network (CNN)-LSTM encoder-decoder model. The results demonstrate that, the ConvLSTM-based model is more accurate than the others for forecasting temperature (MAE = 2.602 °C, RMSE = 3.456 °C), precipitation (MAE = 3.878 mm, RMSE = 10.503 mm), andsunshine hours (MAE = 3.445 h, RMSE = 4.172 h) in 2014–2016. Furthermore, precise forecasting of meteorological factors allows us to develop a hybrid model and a data-driven model for the prediction of each growth stage separately. The hybrid model combines the ConvLSTM encoder-decoder model with empirical models, whereas the data-driven model comprises the ConvLSTM encoder-decoder model and traditional neural network structures. Finally, we compared the two types of models on a real-world dataset from Dandong, and concluded that the data-driven model is more accurate than the hybrid model for prediction of maize growth stages, with R2 in the range of 0.755–0.883, MAE 0.588–2.205 days, and RMSE 0.978–2.729 days. In the future, these models can also be used to predict the growth stages of other crops.