Estimativa e previsão da propagação da seca nos biomas brasileiros

Abstract

Due to the risks that drought brings to water, food, and energy security from a specific population, extreme drought can cause not only huge economic losses but also endanger human and animal lives. For countries like Brazil, where hydroelectricity is the primary energy source and agriculture is the leading contributor to the Brazilian economy, intense droughts can cause harm in several ways. Therefore, to mitigate further damage to life quality and economy, understanding drought events behavior and being able to predict future periods of aridity with accuracy can be valuable strategies. A manner of counting drought events and their intensity is the use of standardized indexes, which utilize statistical analysis of a time series to produce indicators of wet and dry periods. The challenge is in collecting high-quality meteorological and, especially in Brazil, hydrological data, making it difficult to obtain estimates regarding drought to generate diagnoses and forecasts. Considering the difficulties of collecting hydrological data, the goal of this work was to analyze drought conditions across Brazilian biomes through hydrometeorological data from a time series that spans from 1980 to 2010, utilizing information from 735 catchments distributed throughout the country. To characterize such drought conditions, it was used standardized indexes regarding meteorological (SPI and SPEI) and hydrological events (SSI), considering precipitation (P), reference evapotranspiration (ETo), and discharge (Q), respectively, in different time scales, varying between 1, 3, 6, 12, and 24 months. Drought events were counted throughout the time series, and trend analysis was conducted on the hydrometeorological variables and the drought indexes. Additionally, models using machine learning (ML) were tested to aid in predicting hydrological drought indexes based on precipitation and reference evapotranspiration, considering a lag between the meteorological indexes and the hydrological one, which was evaluated through cross-correlation analysis. The ML methods employed were support vector machine (SVM), gene expression programming (GEP), and artificial neural networks (ANN). The trend analysis of the micrometeorological data, discharge, and drought indexes variables indicated variations that led to an increase in both hydrological and meteorological drought events during the time series across much of the country, especially in Pantanal, Cerrado, and Amazon rainforest. Considering the total number of drought events, it is noticed that a higher number of events were detected with SPI and SPEI than with SSI, notably in Pampa and Caatinga. It was possible to observe that in Pantanal there is a well-defined lag between meteorological and hydrological drought; however, in the Amazon rainforest and Cerrado, the obtained results indicate that factors beyond P and ETo influence SSI estimates. Among the ML models tested, SVM and GEP provided the best estimates, producing smaller errors than the results generated by ANN. Generally, the events in longer time scales, such as 12 and 24 months, showed estimates with smaller errors, although, in Pantanal, ML models produced satisfactory results in all time scales as long as an ideal lag is considered between the variables. These tools can provide a better understanding of drought behavior in Brazil and the possibilities of promoting severe and extreme hydrological drought events forecast, particularly in areas lacking measurements or consistent time series of discharge.