The research team, led by Luísa Maria de Souza Viana of the Universidade Estadual do Norte Fluminense Darcy Ribeiro, compared multiple linear regression, autoregressive moving average (ARMA), autoregressive integrated moving average (ARIMA), and seasonal ARIMA (SARIMA) models against feedforward multilayer perceptron (MLP) neural networks. Their goal was to reconstruct monthly discharge at the Ituberá gauging station, located on the Cachoeira Grande River approximately 21.2 kilometers upstream from the estuary mouth. The station drains a gauged area of 263 square kilometers within a basin of roughly 473.68 square kilometers, situated in a tropical humid climate with mean annual precipitation of about 2,095 millimeters and no marked dry season.

The timing of the study gave it an unusual natural experiment. Northeast Brazil has endured a severe and persistent drought beginning in 2012, characterized by widespread precipitation deficits and sustained hydrological stress. The researchers designed two temporal configurations: T1, which calibrated models on data through 2005 and validated on the subsequent record, and T2, which extended calibration through 2011 and focused validation on the post-2012 drought period. This allowed them to test whether models trained on longer, more recent records would generalize better under hydroclimatically contrasting conditions—a question of growing urgency as climate extremes intensify worldwide.

The methodological rigor of the study stands out. Rather than relying on a single validation split, the team employed rolling-origin cross-validation, an expanding-window procedure that preserved temporal ordering and evaluated candidate models at multiple forecast origins within the calibration period. Candidate MLP configurations varied in predictor set, hidden units, and weight-decay regularization, with three fixed-seed random initializations per candidate per fold. The final selected networks were fitted twenty times and averaged into ensemble predictions, reducing dependence on any single initialization. A seasonal-naive benchmark—simply repeating the discharge observed in the same calendar month of the previous year—provided a minimum-performance reference that all fitted models had to beat.

The headline finding is striking in its simplicity: no model family was uniformly superior. Under T1, the univariate MLP using three discharge lags achieved the lowest cross-validation errors of all model families, with a root mean squared error of 1.713 cubic meters per second. Under T2, however, the SARIMA model dominated the selection stage with a cross-validation RMSE of 1.309 cubic meters per second—roughly 30 percent lower than its ARMA and ARIMA competitors. Yet when these same models faced independent validation on held-out data, the rankings reversed. The T2 SARIMA’s selection-stage advantage evaporated, producing validation errors similar to simpler models, while the univariate MLP edged ahead of the stochastic models under T2 but fell behind them under T1.

Statistical testing drove the point home. Paired Diebold–Mariano tests, which compare predictive accuracy using absolute-error loss differentials over matched dates, detected no significant differences between the univariate MLP and any of the stochastic models under either temporal configuration after correction for multiple testing. Even more revealing were the predictor-matched comparisons, in which linear regression and MLP models were given identical input variables. Under T1 with discharge-plus-precipitation predictors, linear regression significantly outperformed the MLP—the only statistically significant result among the matched contrasts. The nonlinear transformation, in other words, added flexibility without adding accuracy.

The study also delivered a sobering hydroclimatic signal. Mean monthly discharge dropped from 5.52 cubic meters per second before 2012 to 4.71 cubic meters per second afterward—a reduction of approximately 14.8 percent. Thirteen post-2012 months exhibited discharge below the 10th percentile of the pre-2012 distribution, and both Welch and Wilcoxon tests confirmed the reduction was significant after adjustment. Yet the Brown–Forsythe test found no significant change in variance, the monthly climatology retained its general shape, and formal change-point tests pointed to candidate breakpoints in 1981 or 1985 rather than 2012. The post-2012 period, the authors conclude, represents a hydroclimatically contrasting interval of reduced discharge magnitude rather than a formally demonstrated regime shift.

Perhaps the most technically instructive results concern error propagation. One-step-ahead predictions, which used observed antecedent discharge as input, transferred reasonably well to held-out data. But completely recursive trajectories—where model-generated estimates were fed back into the predictor history—accumulated errors rapidly, with several configurations producing negative Nash–Sutcliffe efficiency values. The univariate MLP, for instance, produced recursive RMSE values of 3.810 and 3.401 cubic meters per second under T1 and T2 respectively, far worse than its one-step performance. Multi-horizon predictions at 1, 3, 6, 12, and 24 months showed a general decline in skill as distance from the last observation increased. The lesson: satisfactory one-step generalization cannot be extrapolated to long uninterrupted reconstructions.

The team quantified predictive uncertainty using a nonparametric moving-block residual bootstrap restricted entirely to calibration information. Pseudo-out-of-sample residuals were organized into overlapping 12-month blocks to preserve short-term dependence and annual structure, then resampled with replacement across 2,000 realizations. The resulting 95% predictive intervals achieved empirical coverage between 96.2 and 97.4 percent—slightly conservative but broadly stable across configurations. These intervals, the authors emphasize, remain conditional on the selected models, predictors, and calibration procedure; they do not incorporate uncertainty from model selection, alternative architectures, or structural hydrological choices.

For the Serinhaém estuary, the implications are conceptual but consequential. River discharge controls estuarine circulation, salinity gradients, stratification, residence time, and the delivery of nutrients and contaminants. Persistent reductions in freshwater inflow may favor landward salinity penetration and alter the time available for sedimentation and biogeochemical processing—changes relevant to an estuary within the protected Pratigi Environmental Protection Area. But the authors are careful: the approximately 15 percent reduction cannot be translated into specific salinity displacement or residence-time changes without calibrated hydrodynamic modeling, and the reconstructed series serves as an upstream boundary indicator rather than a complete estuarine water budget. The broader message for the field is equally clear: in data-scarce tropical coastal basins, rigorous temporal validation, predictor-matched benchmarking, and explicit uncertainty characterization matter more than assuming that more complex nonlinear models are inherently better.

Subject of Research: Environmental Science

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

Article References: Viana, L. M. D. S., Tostes, E. C. L., Constantino, W. D., Luze, F. H. R., Salomão, M. S. M. D. B., de Jesus, T. B., & de Carvalho, C. E. V. (2026). Modeling freshwater inflow to a tropical estuary using stochastic and neural-network approaches under contrasting hydroclimatic conditions in Brazil. Environmental Monitoring and Assessment, 198(11), Article 1211. https://doi.org/10.1007/s10661-026-16034-1

Image Credits: AI Generated

DOI: 10.1007/s10661-026-16034-1

Keywords: hydrology, neural networks, time-series modeling, estuary, Brazil, drought, machine learning, streamflow reconstruction, Modeling, freshwater, inflow, tropical