The study, led by Hugo Pimentel Tavares of the Universidade do Estado do Rio de Janeiro together with Lucas Lamin de Souza Silva, Nilo Antônio de Souza Sampaio and Carin von Mühlen, addresses a problem that has long undermined environmental regulation in developing economies: the fragmentation of water quality data across multiple agencies with heterogeneous parameter nomenclatures, sampling protocols and reporting formats. In south-eastern Brazil, four bodies—the state environmental agency CETESB in São Paulo, INEA in Rio de Janeiro, the federal ANA-HidroWeb system, and IGAM in Minas Gerais—monitor the same river basins, yet their records cannot be directly compared. The authors set out to resolve that heterogeneity with a reproducible pipeline they call the Optimised Water Quality Monitoring Framework, or OWQMF, a four-block architecture that harmonizes multi-source data, computes dual water quality indices with objectively derived weights, produces machine-learning forecasts and selects the best forecasting model through a formal multi-criteria consensus procedure.

The first block tackles data harmonization, a formidable engineering challenge given that the raw archives contained 524 distinct parameter strings in two languages and multiple encodings. The team constructed an ontological map that collapses these into 14 canonical identifiers covering the nine parameters of the Brazilian water quality index, IQA-CETESB, plus five additional analytes. Two mappings demanded particular chemical care. Ammonia nitrogen, used by CETESB and INEA, and total nitrogen, recorded by ANA-HidroWeb, were deliberately kept as separate identifiers because the two species are chemically non-equivalent: ammonia captures the recently discharged, biologically labile fraction that responds acutely to sewage inputs, while total nitrogen integrates all oxidation states regardless of availability. Thermotolerant coliforms and Escherichia coli, by contrast, were mapped together because Brazilian legislation treats them as functionally equivalent. From 219,508 raw observations, the pipeline retained 108,270 harmonized records, and an Isolation Forest algorithm was applied to flag statistical anomalies—crucially, flagged records were retained rather than discarded, since an extreme dissolved-oxygen depression or a coliform surge from a sewage bypass is precisely the evidence a monitoring network exists to capture.

The framework’s second block computes both the regulatory IQA-CETESB index, which uses fixed weights mandated in the national schedule, and a new Optimised Water Quality Index whose weights are derived from the data itself through a game-theory aggregation of three objective weighting methods: CRITIC, Shannon entropy and MEREC. The result is a systematic redistribution of weight away from parameters that vary little and toward those that discriminate between monitoring conditions. Dissolved oxygen rises from a regulatory weight of 0.170 to 0.222, while pH falls from 0.120 to 0.048 and total solids from 0.080 to 0.038. This direction of redistribution held in every basin and under every weight-derivation variant tested, although the identity of the single highest-weighted parameter proved basin-dependent—dissolved oxygen leads where organic loading drives oxygen sag, turbidity where episodic sediment loading dominates. The authors draw a pointed conclusion from this: a single fixed national weight schedule cannot express the difference between measuring regulatory compliance and ranking ecological condition.

Perhaps the study’s most striking finding concerns how much the choice of classification scale alone can change a river’s apparent health. Because CETESB, INEA and the NSF-WQI convention used by ANA-HidroWeb draw their quality category boundaries at different points on the same 0-to-100 scale, between 18 and 31 percent of station-months receive discordant quality labels depending solely on which agency’s scale is applied to an identical index score. In the critical IQA range of 37 to 52, CETESB communicates “Fair”—acceptable—while the NSF convention communicates “Bad,” for the same underlying water. Layering on the ammonia-total nitrogen mismatch, which produces index differences of 8 to 15 points at reservoir and sewage-impacted stations, and the objective re-weighting, the combined inter-index divergence reaches 23 to 31 percent of station-months across the three basins.

The regulatory consequences become stark in the external validation. Using an extended CETESB archive covering 3,671 station-months from 1978 to 2026 across 22 stations in the Paraíba do Sul basin, the researchers compared composite IQA scores against per-parameter thresholds under CONAMA Resolution 357/2005, Brazil’s foundational water regulation. Among station-months officially designated as Classe 2—the baseline regulatory class—74.1 percent concealed at least one per-parameter violation invisible to the composite index. Worse, among station-months the IQA labeled as “Good” (scores of 52 to 78), 95.5 percent were simultaneously non-compliant with at least one parameter-level threshold. The multiplicative geometry of the index allows high sub-index values for some parameters to compensate for failures in others, and dissolved-oxygen violations in particular escaped the composite score at a rate of 92.7 percent. Because CONAMA 357 specifies minimum dissolved-oxygen concentrations as absolute limits that no other parameter can offset, a monitoring strategy based solely on composite indices structurally underserves the legal framework it is meant to support.

The framework’s third and fourth blocks bring machine learning and formal decision science to forecasting. Monthly IQA time series at seven stations were modeled with three gradient-boosting learners—XGBoost, LightGBM and random forest—using autoregressive lags, cyclical month encoding and rolling statistics, validated with walk-forward cross-validation to prevent data leakage. Errors ranged from 2.3 to 4.5 IQA units at clean-catchment and main-channel stations, though the heavily sewage-loaded Queimados stations defeated all models, with percentage errors above 20 percent—a failure mode consistent with the episodic discharge literature. Model selection was then performed not by picking a single algorithm but through a Borda Count consensus across three multi-criteria methods, TOPSIS, VIKOR and MARCOS, with criterion weights derived objectively by CRITIC. Directional accuracy dominated the weighting at 0.809, a result the authors show is structurally driven by correlation: error-magnitude metrics are nearly collinear with one another, while directional accuracy measures something genuinely non-redundant. Bootstrap and leave-one-station-out diagnostics confirmed the ranking’s stability. Finally, Adaptive Conformal Prediction wrapped the winning model’s forecasts in intervals carrying a nominal 90 percent coverage guarantee, achieved at the two stations with the longest records—though shorter records of roughly 80 months yielded only 83.3 percent coverage, an instructive demonstration that conformal guarantees require sufficient calibration data.

To test whether the weighting scheme generalizes beyond Brazil, the team applied it, entirely without recalibration, to three external datasets spanning different regulatory worlds: the Ohio River Basin in the United States with 13,230 station-months, Italy with 49,380, and Ireland with 13,380, the European cases benchmarked against the EU Water Framework Directive. The IQA-equivalent median scores clustered in a remarkably narrow band from 81.9 to 86.1—suggesting superficially similar composite quality on three continents. Yet full regulatory compliance, defined as satisfying all applicable parameter-level thresholds simultaneously, ranged from 12.7 percent in Brazil to 74.7 percent in the Ohio basin, 79.7 percent in Italy and 97.1 percent in Ireland. The dominant failure modes differed structurally: thermotolerant coliforms drove non-compliance in Brazil, E. coli from combined sewer overflows in Ohio, and BOD exceedances from intensive agriculture in Italy. The authors are careful to note that because the sub-index response curves were calibrated for tropical Brazilian rivers and were not regionally recalibrated, the external scores should be read comparatively rather than as regulatory assessments; what the transfer establishes is the portability of the weighting scheme, and the within-context finding that near-identical composite scores can conceal compliance rates differing by a factor of seven.

The study’s implications reach into Brazil’s ongoing “enquadramento” process, the legal classification of water-body segments under CONAMA 357. If classification results differ by agency without any change in the underlying physical data, compliance assessments across classified segments will be inconsistent, undermining the regulatory function the process is designed to fulfil. The authors recommend a harmonized nitrogen-species protocol, alignment of monitoring stations with classified river segments, and the integration of per-parameter compliance tracking as a non-compensatory criterion in future forecast-model selection. They also acknowledge limitations: the framework does not yet cover trace metals or emerging contaminants, which are recorded preferentially at the most polluted stations and lie outside the index’s parameter set, and climate-driven non-stationarity is only partially mitigated by the adaptive conformal layer. Still, as an argument that water quality regulation requires per-parameter assessment rather than compensatory composite scores—and as a demonstration that a reproducible, open-source pipeline can deliver it across four agencies, three basins and three continents—the work marks a substantial step toward monitoring networks that reveal, rather than conceal, the state of the rivers they watch.