Spatio-temporal modeling of climate change impacts on drought forecast using Generative Adversarial Network: A case study in Africa

Drought is an extreme weather event, affecting the ecological conditions of vegetation and agricultural productivity, poses challenges for millions of people in Africa, and its long-term prediction is definitely important. Accurate drought forecasting is a challenging subject due to its dependence on different climatic variables, and its spatio-temporal, nonstationary and non-linear characteristics. In particular, Deep Learning technologies have achieved excellent results in long-term time series forecasting. Thus, this study proposes a Generative Adversarial Networks (GAN) model which combines Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) network for drought forecasting in Africa. This approach focuses on how the future spatio-temporal variations of drought will vary under climate change effects using multivariate remote sensing data over Africa from 1999-2022. We considered hydrological, meteorological and vegetation spectral factors for GAN as model input variables. The study assessed agricultural drought using the soil moisture index (SMI) as a response parameter. Experimental results confirmed the reliability of the proposed model for forecasting agricultural drought. Compared to existing deep learning models, the proposed GAN based CNN-LSTM model achieved the lowest RMSE, MAPE, and MAE values of 1.008, 0.009, and 0.739, respectively. The findings demonstrate that the proposed model can be used as a reliable forecasting method that helps to estimate drought in arid and semi-arid regions.