Use of Machine Learning in flood forecasting in the Amazon: a case of study of the Acre River flooding in Rio Branco, Acre, Brazil

Flooding is a type of natural disaster in which the rising river level during the high-water season comes into contact with society, causing damage. In the Amazon, floods are among the most frequent natural disasters in the region. Aiming to minimize the impacts caused by flooding, this research sought to apply different statistical hydroclimatological modeling methods and analyze their effectiveness in forecasting the monthly maximum river levels of an Amazonian river at a given control point, four months in advance. To this end, a case study was conducted on the 2012 flooding of the Acre River in the city of Rio Branco, capital of the Brazilian state of Acre, where the river reached a level of 17.64 meters—its third highest recorded level since measurements began. Four statistical modeling methods were used: one based on Multiple Linear Regression (MLR) and three using different Artificial Neural Network (ANN) algorithms. For model simulations, the input data consisted of 40 years (1971 to 2010) of monthly maximum levels of the Acre River in Rio Branco, monthly average Sea Surface Temperature (SST) data from the tropical regions of the Pacific and Atlantic Oceans, and monthly average atmospheric pressure data from Darwin, Australia, and Tahiti, French Polynesia, over the same period. The simulation results from each model were compared with observed data at the monitoring station, and model accuracy was evaluated using performance indices. In the analysis of the models tested to simulate monthly maximum levels in a continuous historical time series, all showed acceptable predictions with satisfactory performance indices, with the MLR-based model standing out. However, when comparing only the maximum level observed in 2012, the ANN-based models were more accurate, missing the observed value by only 27 cm four months in advance, proving more efficient at capturing the climatic patterns that cause the Acre River to reach exceptional levels—unlike the MLR method, which underestimated the peak by 323 cm. The performance results found in this study endorse these statistical hydroclimatic models as operational tools for environmental agencies, serving as indispensable instruments for water management and disaster prevention months in advance, thereby helping to mitigate the frequent flooding of the Acre River.