EGU General Assembly 2022
© Author(s) 2022. This work is distributed under
the Creative Commons Attribution 4.0 License.

A spatiotemporal deep learning forecasting model for long-term drought prediction

Athanasios Loukas1 and Lampros Vasiliades2
Athanasios Loukas and Lampros Vasiliades
  • 1Aristotle University of Thessaloniki-Dept. Rural and Surveying Engineering, University Campus, 541 24 Thessaloniki, Greece (
  • 2University of Thessaly, Department of Civil Engineering, , Pedion Areos 38334 Volos, Greece (

Droughts are slow-moving natural hazards that comes with high hazardous impacts on the society. Early and timelines forecasting of a drought event can help to take proactive measures and set out drought mitigation strategies to alleviate the impacts of drought. Traditionally, forecasting techniques have used various time-series and/or machine learning methods. However, the use of deep learning methods has not been tested extensively despite its potential to improve our understanding of drought characteristics. This study develops a hybrid spatiotemporal scheme for integrated spatial and temporal forecasting. Temporal forecasting is achieved using a deep feed-forward neural network (DFFN and the temporal forecasts are extended to the spatial dimension using a deep learning approach the Long Short-Term Memory (LSTM) to forecast an operational meteorological drought index the Standardized Precipitation Index (SPI) calculated at multiple timescales. The temporal input variable determination is achieved with the use of the Gamma test that estimates the minimum mean square error (MSE) that can be achieved when modelling the unseen data using any continuous non-linear models. 48 precipitation stations and 18 independent precipitation stations, located at Pinios river basin in Thessaly region, Greece, are used for the development and spatiotemporal validation of the hybrid deep learning forecasting model. Several drought characteristics (drought severity and duration, drought category and spatial extent) are analysed to better understand how drought forecasting was improved. Several quantitative temporal and spatial statistical indices are considered for the performance evaluation of the models. Furthermore, qualitative statistical criteria based on contingency tables between observed and forecasted drought episodes are calculated. The results show that the lead time of forecasting for operational use depends on the SPI timescale. The hybrid spatiotemporal deep learning forecasting model could be operationally used for predicting up to three months ahead for SPI short timescales (e.g. 3-6 months) up to six months ahead for large SPI timescales (e.g. 12-24 months). The above findings could be useful in developing a drought preparedness plan in the region and for drought mitigation purposes.


Key words: deep learning, drought, Standardized Precipitation Index, drought forecasting, spatiotemporal droughts, DFNN, LSTM.

How to cite: Loukas, A. and Vasiliades, L.: A spatiotemporal deep learning forecasting model for long-term drought prediction, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-11509,, 2022.


Display file

Comments on the display

to access the discussion