Application of Deep Learning Method to Drought Prediction

Abstract

Drought has brought great potential threat to agriculture, ecology, economy, society and available water resources of China, while accurate drought prediction can help risk management and development of early warning system, and it can reduce the destructive impact of drought. Among many prediction methods, data-driven model is a suitable tool with small data demand and fast development speed. With the development of machine learning, especially deep neural network (DNN), deep learning method shows great ability in drought prediction, and is reported to outperform traditional time series model (e.g. integrated moving average autoregressive model, ARIMA). However, its use needs to be widely estimated, further developed and adjusted for geoscience analysis. The standardized precipitation evapotranspiration index (SPEI) is reported to meet the needs of agricultural drought monitoring and early warning under the background of climate warming. SPEI at 1-, 3- and 6-month time scale are selected as the quantitative description of agricultural drought, and DNN model driven by meteorological and circulation variable is presented to explore the ability of SPEI prediction at the lead time of 1-3 months. The traditional long short-term memory neural network (TLSTM) has been used in drought prediction, which is limited by the quality of prediction factors and noise. Therefore, an improved TLSTM model (ILSTM) is proposed. With highlight of large-scale climate characteristics, a convolution neural network (CNN) module is combined with the ILSMT model. This newly-developed model (CLSTM) can extract circulation information that contributes to the regional drought change, as well as other outputs of prediction model. Evaluation of the drought prediction capabilities in different models is based on the Pearson correlation coefficient, the root mean square error, and the mean absolute error. Results indicate that overall prediction ability of DNN models outperforms the ARIMA model. And the comparative evaluation results among DNN models show that the architecture of the model has an important impact on the prediction performance. The ILSTM model can extract comprehensive information that contributes to future drought change by nonlinear coding of input variables through the full connected layer. When the correlation coefficient can be raised by 0.04-0.25, the root mean square error can be reduced by 0.07-0.32 and the mean absolute error can be reduced by 0.06-0.27 at the validation stage with different lead time comparing with the TLSTM model. Taking advantage of the circulation information as extra inputs to the ILSTM model, the CLSTM model outperform the ILSTM model, when the correlation coefficient can be raised by 0.03-0.44, the root mean square error can be reduced by 0.09-0.33 and the mean absolute error can be reduced by 0.05-0.26. Both results show that deep learning method has great ability in short-term regional climatic drought prediction.

Keywords

• drought prediction
• standardized precipitation evapotranspiration index
• long and short-term memory network
• convolutional neural network
• deep learning