A new deep learning method for meteorological drought estimation based-on standard precipitation evapotranspiration index

Abstract

In this study, we studied drought parameter estimation based on the standard precipitation evapotranspiration index (SPEI), for four stations located at different places in Türkiye. We proposed a new hybrid deep structure of convolutional neural network (CNN) and long short-term memory (LSTM) methods and compared it with several other existing methods. Our extensive experimental results obtained in different cases of the station, time scale, and estimator show that better estimation results are often obtained with 12-month time scale for all methods. For instance using SPEI-9 values for Şanlıurfa station, the CNN method achieves 0.57 RMSE, 0.24 MADE, 0.11 MAE, and 99.38% R2, whereas the proposed adaptive CNN-LSTM obtains values of 0.42, 0.18, 0.04, and 99.69% for the same metrics, respectively. Also, using SPEI-6 values for Mardin station, the bidirectional-LSTM method achieves 0.72 RMSE, 0.35 MADE, 0.22 MAE, and 98.94% R2, while the proposed adaptive CNN-LSTM obtains values of 0.32, 0.12, 0.02, and 99.53%, respectively In this means, the proposed method is superior to existing methods in all cases in terms of almost all performance metrics. It is also robust to changes in given input time series data coming from different stations. We also investigated the percentage occurrence of drought categories at different time scales. The results of this analysis also show that the highest percentages are seen in the wet category for all stations. The percentages occurrence of drought wet category are 48.3, 50.07, 50.14, and 50.79 at SPEI-3, SPEI-6, SPEI-9, and SPEI-12 time scales, respectively.