Surface water quality prediction via an MLA-Mamba hybrid neural network with GRPO optimization – Scientific Reports

Forecasting the health of rivers and lakes is notoriously difficult: pollutant levels swing with weather, land use, and upstream activity, and the signals recorded by monitoring stations are tangled across time and space. A new study reported in Scientific Reports introduces MLA-Mamba, a hybrid deep learning framework that tackles these complexities head-on, delivering more reliable predictions for key surface water indicators and offering uncertainty estimates to support risk-aware decisions.

Why it matters

Water managers need dependable forecasts to trigger early warnings, schedule sampling, and protect ecosystems. Yet the dynamics behind water quality are highly nonlinear, driven by long-range temporal patterns and localized spatial interactions between stations. Traditional statistical models and standard machine learning approaches often fall short because they struggle to capture both the extended temporal dependencies and the fine-grained station-to-station relationships that shape real-world outcomes.

What’s new: the MLA-Mamba architecture

The proposed MLA-Mamba framework blends three key ideas—advanced sequence modeling, local attention for spatial context, and adaptive optimization—to better reflect how surface water systems behave.

• Mamba-based sequence modeling: At the core is an improved Mamba module, a state-space modeling approach designed for long-range temporal dependencies. Unlike conventional recurrent or convolutional models, Mamba handles extended time horizons more efficiently, helping the network learn seasonalities, delayed effects, and persistent trends in water quality series.
• Multi-Head Local Attention (MLA): To capture spatial interactions, MLA focuses attention on nearby monitoring stations rather than treating all inputs equally. By localizing attention and using multiple heads, the model identifies station-specific patterns and cross-station couplings without being overwhelmed by distant, less relevant signals.
• Multi-task learning across indicators: Instead of building separate models for each variable, MLA-Mamba jointly predicts several core indicators—permanganate index (CODMn), ammonia nitrogen (NH3-N), total phosphorus (TP), and total nitrogen (TN). This shared learning setup exploits inter-variable relationships, allowing improvements in one indicator to inform others.

The training twist: GRPO optimization

Beyond architecture, the study introduces Gradient Reparameterization Optimization (GRPO) to the training loop—one of the first applications of this technique in water quality prediction. GRPO dynamically adjusts learning rates during training, aiming to accelerate convergence and stabilize learning. For datasets marked by noise and heterogeneity, such adaptive control helps the model avoid common pitfalls like oscillation and overfitting, while making better use of limited monitoring data.

How it comes together

In practice, historical time series from multiple stations feed into the Mamba module, which builds a compact representation of long-range temporal structure. The MLA mechanism then layers on localized spatial context, attending to nearby stations and relevant features. The combined representation flows into a multi-task head that outputs simultaneous forecasts for CODMn, NH3-N, TP, and TN. During training, GRPO tunes learning dynamics on the fly, and the entire system is optimized end-to-end.

Evidence from real-world datasets

The researchers evaluated MLA-Mamba on two real-world surface water datasets, comparing it against a suite of baseline models spanning traditional statistics and conventional machine learning. Across multiple error metrics, MLA-Mamba consistently outperformed the alternatives. The gains point to the benefits of pairing state-space sequence modeling with local attention and adaptive optimization—particularly for capturing the intertwined temporal and spatial drivers that define surface water quality.

Confidence you can quantify

Predictions are only as useful as their stated confidence. To that end, the team employed Monte Carlo dropout during inference to quantify predictive uncertainty. By sampling multiple forward passes with dropout enabled, the model estimates confidence intervals around each forecast. These intervals give practitioners a clearer sense of risk—critical for setting conservative thresholds, prioritizing interventions, and communicating uncertainty to stakeholders.

What sets this approach apart

• Long-horizon temporal modeling via Mamba’s state-space formulation, avoiding some of the bottlenecks of traditional RNNs and CNNs.
• Spatial sensitivity through Multi-Head Local Attention, which captures station-neighborhood patterns without diluting focus across the entire network.
• Cross-variable synergy through multi-task learning, leveraging shared dynamics among CODMn, NH3-N, TP, and TN.
• Adaptive training with GRPO, delivering faster convergence and more stable training behavior in complex, noisy domains.
• Actionable uncertainty estimates via Monte Carlo dropout for risk-aware planning.

Implications and next steps

As monitoring networks expand and data volumes grow, models that marry temporal depth, spatial awareness, and robust optimization will be increasingly important. The MLA-Mamba blueprint shows how to combine these ingredients effectively for environmental forecasting. Future directions could include integrating meteorological and land-use covariates, extending the spatial attention to dynamic river network graphs, and exploring transfer learning for regions with sparse labels. With uncertainty quantification baked in, such systems can also support decision support tools that weigh forecast confidence alongside predicted pollutant levels.

The takeaway

MLA-Mamba demonstrates that melding state-space sequence modeling with local attention and GRPO-driven training can lift predictive performance for complex, real-world water quality forecasting. By jointly predicting multiple indicators and quantifying uncertainty, the approach offers both accuracy and transparency—two qualities essential for timely, trustworthy environmental management.