Drought Forecasting: Application of Ensemble and Advanced Machine Learning Approaches

Abstract

Depending on the severity and spatial-temporal variability, droughts can have a wide range of impacts such as crop failure, water shortages, and food insecurity. Accurate and timely forecasting is necessary to mitigate the hazards of extreme weather events, such as droughts, brought on by climate change. A district like Chitradurga in India, which typically receives around 450–600 mm of annual rainfall, will require advanced drought mitigation strategies and plans before the onset of the drought. This research focuses on 1-step lead time forecasting of meteorological drought episodes making use of the 6-month Standardised Precipitation Index (SPI-6) as indicator. The fine resolution rainfall data (0.25 $^{\circ }\,\,\times0.25^{\circ }$ ) obtained from the Indian Meteorological Department was used to derive the 6-month SPI data of 23 grid stations. The 1-step lead time SPI-6 time series was forecast considering the antecedent SPI-6 time series data as model input. The Mutual Information was used to determine the most relevant input features for drought forecasting. The standard Artificial Neural Network, an advanced machine learning framework - Multivariate Adaptive Regression Splines, and the ensemble learning-based CatBoost Regression and Gradient Tree Boosting paradigms were employed to forecast drought episodes. Error and efficiency metrics were employed for performance evaluation of the simulated models. The multivariate adaptive regression splines and gradient tree boosting forecasts had slightly higher accuracy and lower error rates than the artificial neural network model, which suggests that they may be more reliable for drought forecasting. The root mean square error and normalized Nash-Sutcliffe efficiency ranges of the multivariate adaptive regression splines model (during test phase) were 0.37–0.54 and 0.78–0.87, respectively. The thematic maps that were created using spatial interpolation of model forecasts from all the stations also confirmed that the district as a whole experienced drought in April 2019.

Keywords

• CatBoost regression
• droughts
• ensemble learning
• gradient tree boosting
• multivariate adaptive regression splines
• mutual information gain