Droughts are one of the most complex, common, and catastrophic natural disasters, causing severe damage to agriculture and the economy. However, drought susceptibility must be measured and predicted in a systematic way, especially in light of potential climate change scenarios. This study aimed to predict current and future drought susceptibility in Bangladesh using historical climate data (1991–2020) and coupled model intercomparison project 6 data for three seasons: pre-monsoon, monsoon, and post-monsoon. We applied an advanced machine-learning algorithm of artificial neural network (ANN) with a genetic algorithm (GA) optimizer to predict drought-prone areas. Nine hydrological parameters–rainfall, temperature, humidity, cloud coverage, wind speed, sunshine, potential evapotranspiration, and solar radiation–were used to develop drought susceptibility maps. Receiver operating characteristic curves and statistical metrics were used to validate the models. The results of a multilayer perceptron ANN coupled with a GA-based optimizer showed that the relevant statistical measures for training and testing datasets were the root mean square error (RMSE = 0.127 and 0.160) and coefficient of determination (R2 = 0.967 and 0.949) for the pre-monsoon season, monsoon season (RMSE = 0.023 and 0.035; R2 = 0.998 and 0.997), and post-monsoon season (RMSE = 0.083 and 0.142; R2 = 0.986 and 0.959), respectively. Further, drought-prone areas in the baseline drought period of 2020 for pre-monsoon season represented 23.86%, 14.24%, 12.85%, 29.92%, and 19.13% of the total area, respectively; similarly, for monsoon corresponding values were 1.83%, 44.18%, 4.99%, 8.76%, and 40.24%; and for post-monsoon drought they were 24.43%, 20.94%, 16.04%, 37.79%, and 0.80% of the total landmass of Bangladesh. These results can help reduce future drought impacts and be of value in assisting policy responses in the country.
Read full abstract