Drought forecast model based on Artificial Neural Networks for Brazilian municipalities

The increase in the frequency of droughts, driven by climate change, implies the need to understand and mitigate these extreme events. In Brazil, there are technical-scientific gaps in relation to climate disaster warnings. The integration of an inventory of droughts that caused losses with remote sensing data, hydrometeorological and climate indexes, using artificial neural networks (ANN) can contribute to a drought forecast. In this study, we developed a monthly forecast model for droughts in Brazilian municipalities using ANN. Precipitation and temperature indexes, in addition to municipal descriptors, for example, the region of the country, the biome, and distance from the oceans and the Amazon, were used as predictor variables in the model. We used an inventory of droughts that caused losses by municipalities (2013–2022) from the Brazilian Integrated Disaster Information System. After model training, we tested the ANN for drought forecasts for lead times of 1–4 months, using seasonal forecast data from the European Center for Medium-Range Weather Forecasts (ECMWF). The overall accuracy of the ANN model for drought simulation was 0.931. The forecast accuracy ranged from 0.922 for a 1-month lead time to 0.757 for 4 months. Remarkably, the model reproduced the spatial pattern of droughts, especially when the output is interpreted as a continuous index of drought risk. We conclude that the trained model is efficient and the results indicate strong potential for drought forecasting and warning, using ANN, remote sensing data, hydrometeorological and climate indexes.