Development of a Long-Range Hydrological Drought Prediction Framework Using Deep Learning

Long-range (1 to 6 months in advance) prediction of droughts is challenging due to its inherent complexity. In this study, we developed a Long-Range Hydrological Drought Prediction Framework (HDPF), empowered by a Deep Learning (DL) approach. Starting with two state-of-the-art approaches, namely Long Short-Term Memory (LSTM), and one-dimensional Convolutional neural networks (Conv1D), we picked out Conv1D to develop the HDPF, being its relatively better performance. The devised HDPF leverages a comprehensive set of eight meteorological precursors, harnessing their collective potential to offer predictions of reasonable accuracy (> 70%). The developed HDPF is able to extract the hidden information from the pool of meteorological precursors along with its evolution over time and influence on the upcoming drought status. Additionally, while comparing the performance of the Conv1D against LSTM, it is noticed that the performance of LSTM is at par with that of Conv1D. However, considering the model parsimony and computational time we advocate the usage of Conv1D. Moreover, comparison against other popular machine learning models, such as Support Vector Regression (SVR) and Feedforward Neural Network (FNN) further affirms the superiority as well as benefits of Conv1D. The developed HDPF can also be useful to other basins in a different climate regime, subject to its recalibration with the location-specific datasets. Overall, this study advances drought prediction methodologies by demonstrating the potential of DL techniques while underscoring the utility and adaptability of the proposed Conv1D-based HDPF.