AI Forecasts Ecological Risk From Heavy Metals in New York–New Jersey Harbor

A neural network can now predict how dangerous heavy metal pollution is for aquatic life in one of America’s most contaminated urban rivers with remarkable precision, according to new research on the Lower Passaic River in New Jersey. The study, published in Environmental Monitoring and Assessment, paired machine learning algorithms with a widely used ecological risk index to transform what has long been a static, retrospective exercise into a dynamic forecasting tool. The artificial neural network at the heart of the work achieved a testing coefficient of determination of 0.981 and a mean absolute error of just 2.25, meaning it could anticipate ecological risk scores from water chemistry data almost perfectly across hundreds of samples collected over three seasons.

The research team, led by Md Shahnul Islam of Montclair State University together with Sara Naveed, Huan Feng, and Tapos Kumar Chakraborty, focused on the upper portion of the Lower Passaic River, a 17-mile tidal system stretching from the Dundee Dam to Newark Bay. This waterway sits within the New York–New Jersey Harbor system, an estuary shaped by intense maritime commerce, dense urban infrastructure, and a legacy of industrialization. The lower section of the river is listed on the U.S. EPA Superfund National Priorities List and carries historical contamination from polychlorinated biphenyls, dioxins, and heavy metals released by chemical manufacturing and industrial waste disposal in Newark, Passaic, and Paterson.

To build their dataset, the researchers drew on the Passaic River Digital Database, a collaborative monitoring effort maintained by the U.S. Environmental Protection Agency, the U.S. Army Corps of Engineers, the New Jersey Department of Environmental Protection, and the U.S. Fish and Wildlife Service. Five sampling stations were established between River Mile 8.4 and the Dundee Dam, and water was collected during 11 sampling events across three seasonal campaigns: December 2019 representing fall, June through August 2020 representing summer, and March through April 2021 representing spring. At most stations, samples were taken at two depths, roughly 0.9 meters below the surface and 0.6 meters above the river bottom, yielding approximately 18 samples per event and a final dataset of 198 water samples.

The team concentrated on three heavy metals chosen for their toxicity, regulatory relevance, and documented presence in the river: copper, lead, and mercury. Copper concentrations ranged from 3.1 to 42.6 micrograms per liter, lead from 1.8 to 25.4 micrograms per liter, and mercury from 0.12 to 1.36 micrograms per liter. Samples were acidified to below pH 2 with ultrapure nitric acid immediately after collection and stored at 4 degrees Celsius following EPA Method 1669. Copper and lead were measured by inductively coupled plasma mass spectrometry under EPA Method 6020A, while mercury was analyzed by cold vapor atomic fluorescence spectrometry under EPA Method 1631 Revision E. Recovery rates for all three metals fell between 92 and 106 percent, and method detection limits were as low as 0.002 micrograms per liter for mercury.

From these measurements, the researchers calculated a suite of pollution indices, including the contamination factor, ecological risk index, contamination degree, heavy metal evaluation index, heavy metal pollution index, and heavy metal toxicity load. Their target variable was the potential ecological risk index, or PERI, which integrates both the level of contamination and the toxic response of each metal into a single ecological risk score. PERI values across the dataset spanned from 85.7 to 672.3, corresponding to moderate through very high ecological risk levels. The choice of PERI as the machine learning target was deliberate: because it weighs toxicity as well as concentration, it offers a more ecologically meaningful picture than raw metal concentrations alone.

The seasonal and spatial patterns that emerged were striking. Copper showed statistically significant seasonal variation, with concentrations peaking in fall and lowest in spring, a pattern the Kruskal–Wallis test confirmed with a p-value of 1.8E-07. Lead and mercury, by contrast, showed no significant seasonal variation, suggesting their presence in the water column is governed less by short-term runoff events and more by long-term processes such as the slow release of legacy contaminants from bed sediments and the continuous resuspension of previously deposited material. Downstream stations at River Miles 8.4 and 10.2 consistently recorded the highest contamination and risk levels, with contamination degree values reaching 11.27 and 9.296 respectively during fall, and PERI peaking at a serious level of 102 at River Mile 8.4.

Correlation and principal component analyses revealed why suspended particles matter so much in this system. All three metals correlated positively and significantly with suspended sediment concentration, with coefficients between 0.58 and 0.67, and with particulate organic carbon, with coefficients between 0.49 and 0.60. Lead and mercury were also strongly correlated with each other at r equals 0.83, hinting at shared sources or similar environmental behavior. The first principal component, explaining 47.04 percent of total variance, loaded heavily on suspended sediment concentration, copper, lead, and mercury, confirming that particulate matter is a dominant transport pathway. Chlorophyll-a, a proxy for algal productivity, correlated more strongly with lead and mercury than with copper, suggesting a biological dimension to metal cycling as well.

With the environmental groundwork laid, the team turned four machine learning algorithms loose on the problem of predicting PERI: random forest, support vector machine, extreme gradient boosting, and artificial neural network. The dataset was split 70:30 into training and testing subsets with a fixed random seed, and hyperparameters were tuned exclusively on the training data using fivefold cross-validation to prevent data leakage. The neural network, built with two hidden layers of five and three neurons using hyperbolic tangent activation and resilient backpropagation, outperformed everything else. During training it achieved a mean squared error of 3.59 and an R² of 0.999, and on the independent testing set it delivered an R² of 0.981, a Nash–Sutcliffe efficiency of 0.98, and a mean absolute percentage error of 4.24 percent. Random forest ranked second, support vector machine third, and XGBoost trailed with the highest errors in both phases, consistent with prior findings that gradient boosting can underperform on relatively small datasets.

Perhaps the most consequential finding came from variable importance analysis. Across all four models, mercury emerged as the dominant driver of ecological risk, with relative importance values of 76.99 percent for random forest, 64.24 percent for XGBoost, and 53.67 percent for the neural network. This is notable because mercury had the lowest contamination factor of the three metals; its outsized influence on PERI stems from its high toxic response factor and its capacity for bioaccumulation in aquatic food webs. The authors point out that mercury’s persistence and bioaccumulative nature, combined with legacy industrial emissions discharged into the river over decades, make it the primary target for any remediation strategy aimed at reducing overall ecological risk. Lead, despite its higher concentrations, consistently showed the lowest predictive significance.

The study’s broader significance lies in its reframing of ecological risk assessment from a static snapshot into a forward-looking forecast. Traditional index-based approaches can tell managers where contamination stands today, but they cannot anticipate how risk will evolve as hydrology, sediment dynamics, and biological activity shift. By coupling machine learning with PERI, the researchers demonstrated a transferable framework that converts new concentration measurements into reliable ecological risk predictions without recalculating the full index from scratch, an operational advantage for data-scarce or rapidly changing monitoring contexts. The authors recommend that water authorities integrate such predictive systems into routine monitoring to identify contamination hotspots in real time, institutionalize continuous multi-season sampling, and enforce stricter limits on mercury discharges. They also acknowledge limitations, including the restricted number of sampling campaigns and the exclusion of other contaminants such as cadmium and arsenic, and they call for future work incorporating hydrodynamic modeling and real-time prediction systems to strengthen early warning and adaptive management of urban estuaries worldwide.

Subject of Research: Machine learning prediction of heavy metal pollution and ecological risk in the Lower Passaic River within the New York–New Jersey Harbor system

Article Title: Predictive modeling of heavy metal pollution and ecological risk for sustainable water quality management in the NY–NJ harbor system

Article References: Islam, M. S., Naveed, S., Feng, H., & Chakraborty, T. K. (2026). Predictive modeling of heavy metal pollution and ecological risk for sustainable water quality management in the NY–NJ harbor system. Environmental Monitoring and Assessment, 198(9), Article 1013. https://doi.org/10.1007/s10661-026-15832-x

Image Credits: AI Generated

DOI: 10.1007/s10661-026-15832-x

Keywords: heavy metal pollution, machine learning, ecological risk, Lower Passaic River, New York–New Jersey Harbor, artificial neural network, water quality, mercury contamination, potential ecological risk index, Superfund site, estuarine pollution, environmental monitoring