Optimization of echo state networks for drought prediction based on remote sensing data

In this paper, we used echo state networks - a class of recurrent neural networks - for prediction of drought based on remote sensing data. To this end, the drought index was obtained for a number of stations in different clime zones of Iran. For each station, we also extracted the corresponding vegetation indices based on satellite imagery. Our model takes the satellitebased features as input and outputs the severity of drought. One of the major challenges of echo state networks is optimization of the reservoir parameters. Here we used a method based on Kronecker product in order to substantially reduce the parameter space to be optimized. We then used various optimization techniques including genetic algorithms, simulated annealing and differential evolution. Our results show that the method based on differential evolution results in the best performance as compared to others.