Spatial meteorological drought forecasting using deep learning for New South Wales, Australia

Abstract:

Droughts are one of the major natural hazards causing severe economic and social losses. One of the major challenges in drought management is accurate drought forecasting under the influence of various climatic variables. Over the past decade, researchers have aimed to forecast drought using either single-based or hybrid-based machine learning approaches, which have been proven successful under certain conditions. However, with the advent of higher computational capability and the availability of long temporal climatic variable datasets, the use of deep learning techniques is potentially the next logical step. Therefore, the present work aims to develop a deep learning model based on Stacked Long Short Term Memory (S-LSTM) to forecast monthly Standardized Precipitation Index (SPI) at multi-month lead times using lagged climatic variables as predictors. Our study area is the New South Wales region of Australia and covers recent severe droughts. The model uses data collected from the Scientific Information for Land Owners (SILO) database to train the model from the years 1901-2000. Next, the model was tested from 2001-2018, and the spatial variation across the state was analysed. The forecasted results were analysed in terms of different drought characteristics e.g.drought intensity, drought duration and drought categories. Lastly, the results were compared with popular machine learning models such as Artificial Neural Networks (ANN) and Random Forest (RF) to understand the capabilities of the S-LSTM model. The analysis shows that deep learning models perform better than the ML models and a better understanding of drought characteristics is developed, especially at long lead times when the S-LSTM model is used.