AI Reveals Shifting Nutrient Rules Behind Toxic Lake Algal Blooms

Few environmental problems are as visually dramatic or as stubbornly difficult to manage as a freshwater algal bloom. Every summer, lakes across the world turn green as phytoplankton explode in numbers, clogging intakes, fouling shorelines, and in the worst cases releasing toxins that threaten drinking water for millions. For decades, scientists have wrestled with a deceptively simple question: what actually drives these blooms? A new study of Shahu Lake in northwest China offers an unusually precise answer, and in doing so demonstrates how interpretable artificial intelligence can turn the black box of machine learning into a practical instrument for ecological decision-making.

The research, published in Environmental Earth Sciences, was led by Yong Li and Zhongyao Liang of Xiamen University together with colleagues at the Tianjin Municipal Engineering Design and Research Institute and the Nanjing Institute of Geography and Limnology. The team built a multi-scale interpretable machine-learning framework and applied it to nearly four years of daily water-quality monitoring at Shahu Lake, a shallow, 45-square-kilometre water body lying at 1,093 to 1,102 metres above sea level where the foothills of the Helan Mountains meet the alluvial plain of the Yellow River. Like many lakes in arid and semi-arid China, Shahu Lake experiences pronounced swings in its eutrophication status, driven by artificial water replenishment and a warming, variable climate.

At the heart of the study is a model built on the XGBoost algorithm, a gradient-boosted tree method prized for its efficiency with noisy, multivariate environmental data. The researchers fed the model daily measurements of water temperature, chemical oxygen demand measured by the manganese method, ammonia nitrogen, total phosphorus, total nitrogen, turbidity, and the nitrogen-to-phosphorus ratio, all drawn from the China National Environmental Monitoring Centre between November 2021 and August 2025. After screening out eighteen anomalous readings, just over one percent of the record, the team retained a complete set of 1,370 daily observations, splitting them eighty-twenty into training and independent testing sets. Hyperparameters were tuned with Bayesian optimization combined with five-fold cross-validation.

The resulting model performed remarkably well. It achieved a coefficient of determination of 0.889 on the training data and 0.810 on the unseen testing data, with root-mean-square errors of 3.78 and 5.13 micrograms per litre respectively and Kling-Gupta efficiencies above 0.81 in both cases. Those figures indicate strong predictive accuracy and, crucially, good generalisation, meaning the model had not simply memorised the training record but had learned relationships that hold for data it had never seen. Yet prediction alone was never the goal. The real innovation lies in how the researchers interrogated the model, using the TreeSHAP recursive algorithm, an exact method for attributing each prediction to individual input variables, to peel apart the drivers of chlorophyll-a, the standard proxy for phytoplankton biomass.

What makes the study distinctive is that the interpretation is deliberately multi-scale. Most previous applications of SHAP in aquatic science have averaged feature importance across an entire dataset, a global view that can mask drivers that matter only during particular seasons or events. Li and colleagues computed global importance across all 1,370 days, regional importance within the summer bloom window of July to September, and local importance for the ten most extreme chlorophyll-a events on record. At the regional scale they normalised SHAP values within each feature to remove scale differences, a methodological refinement designed to make season-specific comparisons meaningful.

The global picture largely confirms classical eutrophication theory. Water temperature emerged as the single most influential driver, with the model showing chlorophyll-a responses turning sharply positive as waters warm from about 15 to 24 degrees Celsius before plateauing at higher temperatures. Chemical oxygen demand, a proxy for organic matter and its mineralisation, ranked second, pointing to the role of internal nutrient recycling in shallow lakes, where decaying organic matter on the lakebed releases nitrogen and phosphorus back into the water column. Among nutrients, phosphorus consistently outweighed nitrogen in long-term importance, echoing the foundational whole-lake experiments of David Schindler and supporting the long-standing paradigm that phosphorus control is the backbone of lasting eutrophication management.

But the regional analysis revealed something the global view had concealed. During the July-to-September bloom period, the hierarchy of drivers shifted markedly. Nitrogen-related variables, particularly ammonia nitrogen and the nitrogen-to-phosphorus ratio, rose in importance, while the system effectively transitioned from chronic phosphorus limitation to transient nitrogen limitation under elevated thermal conditions. The response to water temperature peaked narrowly between 23 and 24 degrees Celsius, matching the optimal thermal niche for rapid cyanobacterial proliferation. The authors interpret this seasonal nitrogen stress not as an independent cause of blooms but as a consequence of them: as phosphorus fuels rapid growth, phytoplankton draw down available nitrogen until it becomes the immediate bottleneck, constraining bloom persistence and even influencing how much Microcystis biomass sinks out of the water column.

The local-scale analysis added a further layer of nuance. When the researchers examined the ten highest chlorophyll-a events individually, they found that each followed a distinct decision pathway through the model, even though the samples shared broadly similar environmental backgrounds. Water temperature contributed a strong, consistent 15.6 to 17.5 micrograms per litre in every case, functioning as the essential prerequisite for extreme blooms, with chemical oxygen demand adding a stable 3.65 to 5.08 micrograms per litre. Nutrient contributions, however, diverged sharply from event to event. Under persistently high phosphorus, the influence of total nitrogen versus the nitrogen-to-phosphorus ratio shifted depending on immediate nitrogen availability, and in a few instances nitrogen variables actually registered negative contributions, which the authors attribute to Liebig-style co-limitation by light, temperature, or phosphorus rather than any evidence that nitrogen depletion suppresses blooms under natural conditions.

These findings carry direct management implications. The framework argues for a dual-tiered strategy: long-term recovery still demands sustained reduction of both external and internal phosphorus loads, but during the summer bloom window managers may need adaptive controls on short-term nitrogen fluxes as well. Ignoring the transient nitrogen limitation, the authors warn, could foster nitrogen-depleted conditions that favour cyanobacterial dominance, a concern consistent with recent dual-nutrient reduction studies on Lake Erie and other systems. More broadly, the study positions interpretable machine learning as a state-dependent decision-support tool, capable of telling lake managers not just what matters on average but what matters now, in this season, under these conditions.

The authors are careful to note the limits of their approach. Because the analysis rests on observational data from a single monitoring station, the identified relationships are statistical associations rather than strict causal mechanisms, and the specific feature attributions may not extrapolate to lakes with different data distributions. Nonetheless, the framework’s central lesson is likely to resonate far beyond Shahu Lake: eutrophication is not governed by a single rule but by rules that change with time and state, and only analytical tools that respect those shifts, combining high-frequency monitoring, gradient-boosted models, and exact multi-scale attribution, can hope to keep pace with blooms that are becoming more frequent and more intense in a warming world.

Subject of Research: Multi-scale interpretable machine learning analysis of chlorophyll-a dynamics and eutrophication drivers in Shahu Lake

Article Title: Unraveling multi-scale drivers of Chl-a dynamics via interpretable machine learning: a case study of Shahu Lake

Article References: Li, Y., Huang, Y., Lin, L., Kong, X., & Liang, Z. (2026). Unraveling multi-scale drivers of Chl-a dynamics via interpretable machine learning: a case study of Shahu Lake. Environmental Earth Sciences, 85(15), Article 398. https://doi.org/10.1007/s12665-026-13133-7

Image Credits: AI Generated

DOI: 10.1007/s12665-026-13133-7

Keywords: chlorophyll-a, algal blooms, eutrophication, Shahu Lake, interpretable machine learning, XGBoost, SHAP, phosphorus limitation, nitrogen limitation, water quality monitoring, cyanobacteria, lake management