A deep learning based approach for long-term drought prediction

Drought is a natural disaster that comes with high hazardous impacts on the society. Its effects are mostly manifested as hydrological drought. Identifying past droughts and predicting future ones is very vital in limiting their effects. However, the random and nonlinear nature of drought variables makes accurate drought prediction remain a challenging scientific problem. Neural networks have shown great promise over the last two decades in modeling nonlinear time series. But the issue of nonconvex optimization ensues when two or more hidden layers are required for highly complex phenomena. This research looks into the drought prediction problem using deep learning algorithms. We propose a Deep Belief Network consisting of two Restricted Boltzmann Machines for long-term drought prediction using lagged values of Standardized Streamflow Index (SSI) as inputs. The proposed model is applied to predict different time scale drought indices across the Gunnison River Basin located in the Upper Colorado River Basin. The study compares the efficiency of the proposed model to that of traditional approaches such as Multilayer Perceptron (MLP) and Support Vector Regression (SVR) for predicting the different time scale drought conditions. The proposed model shows an edge in performance over the traditional methods using Root Mean Square Error and Mean Absolute Error as metrics.