Abstract
The Northern Mesoregion of Minas Gerais (NMMG) lies within the brazilian semi-arid region and is subject to recurrent and prolonged droughts. Despite the advances to drought forecast using machine learning techniques, evaluating the extent to which predictive models trained on specific locations can generalize their estimates to regional scales remain necessary. Thus, we proposed a spatial transferability experiment using Long Short-Term Memory (LSTM) models to predict the Standardized Precipitation Evapotranspiration Index (SPEI) with sliding and tumbling window techniques in five selected municipalities and their bordering districts in the NMMG. To evaluate the efficacy of this transferability, the results were contrasted with the locally trained model’s performance. Although this approach yielded robust performance in 71.8% of bordering districts with the tumbling window, it resulted in limited generalization capability in 28.2% due to the size of geographical area, differences in the rainfall spatial distribution and the presence of physiographic outliers.

This publication has 15 references indexed in Scilit: