Drought modeling and forecasting using emerging deep learning techniques

Drought is a significant challenge to world water security and ecosystem resilience, defined by long dry periods leading to water scarcity. Because of the stochastic recurrence and severe socioeconomic impacts of droughts, precise drought modeling and forecasting are required for effective water resource management. Hence, exploring the efficiency of shallow and deep machine learning techniques to enhance predictive accuracy of drought forecasting models is necessary. This article investigates and compares the efficiency of a shallow multilayer perceptron (MLP) model with two deep models: a two-hidden layer MLP (hereafter TMLP) and Long Short-Term Memory (LSTM). To this end, long-term (1950 to 2024) grid-based monthly Standardized Precipitation Evapotranspiration Index (SPEI) datasets near the holy city of Karbala were retrieved and used to train and test all the models. The nearby grid data sets were averaged arithmetically to represent the temporal variation of drought across this data-scarce city. The representative dataset was then separated into training and testing datasets. The optimum predictors for one-month ahead SPEI forecasting scenarios were determined via mutual information between SPEI and its lagged vectors. The research initially used a single hidden layer MLP model, varying the number of neurons from 1 to 10, and found that performance improved (lower RMSE and greater NSE) as the number of neurons increased, with the best results at 5 neurons. When the range of neurons was expanded to 20, maximum efficiency was attained with 14 neurons. Adding a second hidden layer, 9 neurons in the first layer and 6 in the second produced the greatest results. This emphasizes the need for careful hyperparameter adjustment to prevent overfitting and maximize performance, which depends on balancing the number of neurons and model complexity. In a comparative analysis, the research utilized LSTM models to capture sequential dependencies in the data. The LSTM model, with a single layer of 50 units followed by a dense layer, demonstrated robust performance in predicting sequential data, achieving an RMSE of 0.3631 and an NSE of 0.8427 for the training set and an RMSE of 0.3915 and an NSE of 0.7881 for the testing set. However, the results showed that the shallow MLP with 14 hidden neurons slightly outperformed its deep TMLP and LSTM counterparts. Thus, the use of complex deep learning models is not suggested for SPEI modeling in the studied area, given the data length and characteristics. Our research underscores that while deep learning methodologies like LSTM models can effectively capture long-term dependencies and temporal correlations, simpler models like the shallow MLP can provide better performance in certain scenarios. The findings highlight the importance of selecting suitable model architectures and hyperparameters for optimal prediction performance. Understandings from this study will help stakeholders better predict and mitigate the effects of water shortages on ecosystems, agriculture, and society, contributing to proactive drought management and adaptation measures.