Abstract
Drought is an extreme climate phenomenon that has a great impact on the economy, tourism, agriculture, and water resources. Drought prediction can provide an early warning of the occurrence of drought and reduce losses. In this article, the standard precipitation evapotranspiration index (SPEI) on four time scales: SPEI-3, SPEI-6, SPEI-9, and SPEI-12 are used to measure and predict drought. Unlike the general methods of directly modeling the SPEI index, time-series imaging and feature-based transfer learning are used to extract the features of the SPEI sequence and use the extracted features for prediction. First, we use Gramian Angular Summation/Difference Field (GASF/GADF), Markov Transition Field (MTF), and Recurrence Plot (RP) as the time series imaging techniques to encode SPEI sequences into images. Secondly, we utilize imaging data sets and convolutional neural networks (CNNs) such as residual network (ResNet) and VGG to train the feature extraction network. Finally, the following four regressors: Random Forest (RF), Long and Short-Term Memory network (LSTM), Wavelet Neural Network (WNN), Support Vector Regression (SVR) are used to model the extracted features and drought prediction. To verify the effectiveness of the method proposed in this article, we use the SPEI of four time scales at eight stations in the Haihe River Basin for prediction. Compared with the existing methods, the prediction results of different time scales and stations are improved. For example, after feature extraction, LSTM can reach MAPE = 0.5400, SMAPE = 0.4452, MAE = 0.2150, MSE = 0.0853 and R <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> = 0.8960 in the SPEI-12 prediction of the Beijing site, and other results show that the proposed method is not sensitive to the time scale of drought prediction.
Highlights
Drought refers to the periodical precipitation on land that is lower than normal for many months or years [8]
To fix problems of existing models, we propose a novel feature extraction method based on time series imaging [14] and feature-based transfer learning that can effectively extract drought data features, and its prediction performance does not depend on the study area
In this part, we will give the results based on the model Random Forest (RF), Long and Short-Term Memory network (LSTM), Wavelet Neural Network (WNN), and Support Vector Regression (SVR) to predict the four time scales standard precipitation evapotranspiration index (SPEI): SPEI-3, SPEI-6, SPEI-9, and SPEI-12 for eight sites
Summary
Drought refers to the periodical precipitation on land that is lower than normal for many months or years [8]. [24] proposed the use of machine learning-based methods, including RF, SVR, and boosted regression trees (BRT) for drought prediction in the United States. The prediction performance of most of the models proposed above is strongly dependent on the study area, observation sites, drought evaluation indicators, and time scales, and most models directly predict the original data and cannot extract important features from the data. To fix problems of existing models, we propose a novel feature extraction method based on time series imaging [14] and feature-based transfer learning that can effectively extract drought data features, and its prediction performance does not depend on the study area.
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
