Data-driven forecasting of meteorological drought for the Upper Blue Nile Basin, Ethiopia

In the era of high climate variability, the needs to manage water resources have become of high importance. Most especially in areas that frequently suffer from water shortage, water managers and decision-makers make efforts to assess the trends of hydrological processes and develop models to reduce risks and develop mitigation plans. In view of this, data-driven modeling techniques are getting popularity in their easiness and simplicity to accurately predict time series hydrological parameters. This study presents a data-driven prediction of the Standardized Precipitation Evapotranspiration Index (SPEI) drought index based on large-scale climate indices, sea surface temperatures, hydro-meteorological variables, and topographic attributes by using the multilayer perceptron network. The main objective was to analyze the sensitivity of input parameters that trigger droughts and measure their predictive ability by comparing the predicted values with the observed ones.

In this paper, the resilient back-propagation algorithm is used to train the Artificial Neural Network (ANN) predictive models that incorporate hydro-meteorological variables, climate, sea surface temperatures and topographic attributes as covariates to predict the SPEI at 3-, 6-, and 12-month timescales for seven stations in the Upper Blue Nile basin of Ethiopia, during 1986-2015. For the best architecture, the coefficient of determination ranges from 0.820 to 0.949, the root mean square error varies from 0 .263 to 0.428, and mean absolute error ranges from 0.008 to 0.064. Pertaining to the statistical achievement, we conclude that the ANN model can be used for forecasting the SPEI Index. The results also revealed that adding climate indices improves the accuracy of predictions and the model performance is independent of the timescales. However, careful selection of the input variables is required to enhance the accuracy of predictions.