Accurate prediction of any type of natural hazard is a challenging task. Of all the various hazards, drought prediction is challenging as it lacks a universal definition and is getting adverse with climate change impacting drought events both spatially and temporally. The problem becomes more complex as drought occurrence is dependent on a multitude of factors ranging from hydro-meteorological to climatic variables. A paradigm shift happened in this field when it was found that the inclusion of climatic variables in the data-driven prediction model improves the accuracy. However, this understanding has been primarily using statistical metrics used to measure the model accuracy. The present work tries to explore this finding using an explainable artificial intelligence (XAI) model. The explainable deep learning model development and comparative analysis were performed using known understandings drawn from physical-based models. The work also tries to explore how the model achieves specific results at different spatio-temporal intervals, enabling us to understand the local interactions among the predictors for different drought conditions and drought periods. The drought index used in the study is Standard Precipitation Index (SPI) at 12 month scales applied for five different regions in New South Wales, Australia, with the explainable algorithm being SHapley Additive exPlanations (SHAP). The conclusions drawn from SHAP plots depict the importance of climatic variables at a monthly scale and varying ranges of annual scale. We observe that the results obtained from SHAP align with the physical model interpretations, thus suggesting the need to add climatic variables as predictors in the prediction model.
Droughts are slow-moving natural hazards that gradually spread over large areas and capable of extending to continental scales, leading to severe socio-economic damage. A key challenge is developing accurate drought forecast model and understanding a models' capability to examine different drought characteristics. Traditionally, forecasting techniques have used various time-series approaches and machine learning models. However, the use of deep learning methods have not been tested extensively despite its potential to improve our understanding of drought characteristics. The present study uses a deep learning approach, specifically the Long Short-Term Memory (LSTM) to predict a commonly used drought measure, the Standard Precipitation Evaporation Index (SPEI) at two different time scales (SPEI 1, SPEI 3). The model was compared with other common machine learning method, Random Forests, Artificial Neural Networks and applied over the New South Wales (NSW) region of Australia, using hydro-meteorological variables as predictors. The drought index and predictor data were collected from the Climatic Research Unit (CRU) dataset spanning from 1901 to 2018. We analysed the LSTM forecasted results in terms of several drought characteristics (drought intensity, drought category, or spatial variation) to better understand how drought forecasting was improved. Evaluation of the drought intensity forecasting capabilities of the model were based on three different statistical metrics, Coefficient of Determination (R2), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE). The model achieved R2 value of more than 0.99 for both SPEI 1 and SPEI 3 cases. The variation in drought category forecasted results were studied using a multi-class Receiver Operating Characteristic based Area under Curves (ROC-AUC) approach. The analysis revealed an AUC value of 0.83 and 0.82 for SPEI 1 and SPEI 3 respectively. The spatial variation between observed and forecasted values were analysed for the summer months of 2016-2018. The findings from the study show an improvement relative to machine learning models for a lead time of 1 month in terms of different drought characteristics. The results from this work can be used for drought mitigation purposes and different models need to be tested to further enhance our capabilities.
Drought forecasting with a long lead time is essential for early warning systems and risk management strategies. The use of machine learning algorithms has been proven to be beneficial in forecasting droughts. However, forecasting at long lead times remains a challenge due to the effects of climate change and the complexities involved in drought assessment. The rise of deep learning techniques can solve this issue, and the present work aims to use a stacked long short-term memory (LSTM) architecture to forecast a commonly used drought measure, namely, the Standard Precipitation Evaporation Index. The model was then applied to the New South Wales region of Australia, with hydrometeorological and climatic variables as predictors. The multivariate interpolated grid of the Climatic Research Unit was used to compute the index at monthly scales, with meteorological variables as predictors. The architecture was trained using data from the period of 1901-2000 and tested on data from the period of 2001-2018. The results were then forecasted at lead times ranging from 1 month to 12 months. The forecasted results were analysed in terms of drought characteristics, such as drought intensity, drought onset, spatial extent and number of drought months, to elucidate how these characteristics improve the understanding of drought forecasting. The drought intensity forecasting capability of the model used two statistical metrics, namely, the coefficient of determination (R2) and root-mean-square error. The variation in the number of drought months was examined using the threat score technique. The results of this study showed that the stacked LSTM model can forecast effectively at short-term and long-term lead times. Such findings will be essential for government agencies and can be further tested to understand the forecasting capability of the presented architecture at shorter temporal scales, which can range from days to weeks.