Long, ZH; Qin, QM; Zhang, TY; Xu, W (2020). Prediction of Continuous Time Series Leaf Area Index Based on Long Short-Term Memory Network : a Case Study of Winter Wheat. SPECTROSCOPY AND SPECTRAL ANALYSIS, 40(3), 898-904.

The continuous time series of Leaf Area Index (LAI) can reflect the growth of winter wheat, and the prediction of future LAI is important for guiding agricultural production. The crop growth models, such as the World Food Studies (WO-FOST), can predict the future LAI by simulating the growth and development of winter wheat. But the simulation depends on numerous input parameters, such as future meteorological data, which is difficult to obtain. Due to the continuity and regularity of LAI variations of winter wheat, the future LAI can be predicted with historical LAI through deep learning methods. However, deep learning methods require a large number of samples with labels to build training dataset. The scarcity of training dataset limits the application of deep learning methods in practice. To solve the above problems, we used data assimilation framework to combine remote sensing data with WOFOST model and constructed 15-year time series dataset of winter wheat LAI in Hebei province. Shuffled Complex Evolution (SCE) algorithm was applied to minimize difference between corrected MODIS LAI and simulated LAI for optimizing initial parameters of WOFOST. Based on the dataset, multiple LAI prediction models with different input lengths of historical LAI were established by using the Long Short-Term Memory (LSTM). The abilities of different prediction models to delineate LAI variations of winter wheat were evaluated. Results showed that the LSTM-based models can predict the future LAI of winter wheat effectively. The prediction model with an input length of 20 days achieved the highest accuracy. and RMSE of the prediction model were 0. 986 5 and 0. 183 6 after winter wheat returned green. For different stages of winter wheat growth, the accuracy was higher before winter wheat bloomed and reduced slightly after winter wheat bloomed. Therefore, it could be concluded that the method of constructing training dataset proposed in this study could be a reference for the application of deep learning methods in similar problems. The prediction models built in this study also verified the effectiveness of the LSTM, which provided a helpful way for predicting the future LAI of crops.