Modeling freshwater inflow to a tropical estuary using stochastic and neural-network approaches under contrasting hydroclimatic conditions in Brazil

Freshwater inflow regulates estuarine circulation, salinity gradients, sediment transport, and biogeochemical fluxes, yet long-term discharge records remain scarce in many tropical coastal basins. We compared linear regression, stochastic time-series models (ARMA, ARIMA, and SARIMA), and multilayer perceptron (MLP) neural networks for reconstructing monthly discharge at the Ituberá gauging station, upstream of the Serinhaém estuary (southern Bahia, Brazil), using records from 1966 to 2019. No model family was uniformly superior across rolling-origin selection and independent validation. Predictor-matched comparisons showed no general advantage of MLPs over linear formulations, and extending calibration through 2011 did not consistently improve post-2012 performance. Mean monthly discharge was approximately 15% lower after 2012, while changes in variance and the general seasonal pattern were not detected and evidence for altered temporal dependence remained inconclusive. One-step reconstructions transferred more consistently to held-out data than recursive and longer-horizon predictions, which showed error accumulation. The reconstructed series provides an upstream freshwater boundary indicator for future estuarine modeling, while ecological consequences require explicit hydrodynamic and biogeochemical simulation.