Drought Prediction Based on Feature-Based Transfer Learning and Time Series Imaging

Drought is an extreme climate phenomenon that has a great impact on the economy, tourism, agriculture, and water resources. prediction can provide early warning of occurrence drought reduce losses. In this article, standard precipitation evapotranspiration index (SPEI) four time scales: SPEI-3, SPEI-6, SPEI-9, SPEI-12 are used to measure predict drought. Unlike general methods directly modeling SPEI index, time-series imaging feature-based transfer learning extract features sequence use extracted for prediction. First, we Gramian Angular Summation/Difference Field (GASF/GADF), Markov Transition (MTF), Recurrence Plot (RP) as series techniques encode sequences into images. Secondly, utilize data sets convolutional neural networks (CNNs) such residual network (ResNet) VGG train feature extraction network. Finally, following regressors: Random Forest (RF), Long Short-Term Memory (LSTM), Wavelet Neural Network (WNN), Support Vector Regression (SVR) model To verify effectiveness method proposed in scales at eight stations Haihe River Basin Compared with existing methods, results different improved. For example, after extraction, LSTM reach MAPE = 0.5400, SMAPE 0.4452, MAE 0.2150, MSE 0.0853 R 2 0.8960 Beijing site, other show not sensitive scale