Whales, Waves, and Neural Networks Team Up to Forecast Oxygen in Troubled Waters

Dissolved oxygen is one of the most consequential numbers in any river, lake, or estuary. When it drops, fish suffocate, sediments release stored pollutants, and entire food webs begin to unravel, often before human eyes notice anything wrong. That is why environmental agencies invest heavily in continuous oxygen sensors, and why the data those sensors produce have become a tempting target for machine learning. A new study published in Environmental Monitoring and Assessment by Xuemei Xie of Jiangxi Biotech Vocational College and Jiawei Liang of China Railway Nanchang Group tackles a stubborn problem that has haunted water-quality forecasting for years: the raw signals from monitoring stations are noisy, nonlinear, and stubbornly non-stationary, which has long limited what conventional prediction models can achieve.

The researchers propose a hybrid framework they call WOA-EEMD-CNN-LSTM, a pipeline in which three distinct computational ideas are chained together, each one addressing a specific weakness of the previous stage. The first stage uses ensemble empirical mode decomposition, or EEMD, a noise-assisted signal-processing technique first introduced by Wu and Huang in 2009 as an extension of Huang’s original empirical mode decomposition from 1998. EEMD takes the messy, jagged oxygen time series and breaks it down into a set of intrinsic mode functions, each capturing oscillations at a particular scale, plus a residual trend. Instead of asking one model to learn everything at once, the decomposition hands the downstream learner a family of simpler, better-behaved sub-signals.

But EEMD has a well-known catch: its quality depends critically on parameters, notably the amplitude of the white noise injected into the ensemble and the number of ensemble trials. Set them poorly and the decomposition either fails to cancel out the noise or becomes computationally prohibitive. This is where the whale enters the story. The Whale Optimization Algorithm, introduced by Mirjalili and Lewis in 2016, is a metaheuristic inspired by the bubble-net hunting strategy of humpback whales, in which the animals swim in shrinking spiral patterns to corral prey. In the new framework, WOA is used to search the parameter space and tune the key EEMD settings automatically. The authors describe this as entropy-guided parameter adaptation, meaning the optimization is steered by information-theoretic measures of how well the decomposition separates the signal into meaningful components.

Once the oxygen series has been decomposed into its multi-scale components, the framework deploys a combined convolutional and recurrent neural network to forecast each component individually. The convolutional layer, following the principles established by LeCun and colleagues in their landmark 1998 work on gradient-based learning, acts as a local feature detector, scanning short windows of each sub-series for characteristic patterns. The long short-term memory network, the celebrated 1997 architecture of Hochreiter and Schmidhuber, then models the long-term temporal dependencies that determine how each component evolves. Because the components are largely decoupled from one another after decomposition, each CNN-LSTM pair can specialize, learning the dynamics of a slow trend without being confused by high-frequency jitter, and vice versa. Finally, the component predictions are aggregated back together using timestamp alignment, reconstructing a single coherent forecast of dissolved oxygen.

The evaluation is arguably the most rigorous part of the paper. The authors tested the framework on two very different datasets: a four-hour-interval monitoring series from the Xiaojiang estuary in China and a publicly available dataset known as the Refuge dataset, hosted on Kaggle. To keep the comparison honest, they fixed the random seed at 123 in what they call a locked seed evaluation, ensuring that reported numbers reflect the method itself rather than a lucky draw of random initializations. On the Xiaojiang series, the WOA-EEMD-CNN-LSTM model achieved a root mean square error of 0.32867 and a coefficient of determination, or R-squared, of 0.95070, meaning the model explains roughly ninety-five percent of the variance in the oxygen signal. On the Refuge dataset, it recorded an RMSE of 0.28874 and an R-squared of 0.93429.

Those figures place the framework in respectable territory for dissolved-oxygen forecasting, a field where earlier machine learning efforts have produced a wide spread of results. The literature the authors survey reads like a catalogue of the field’s evolution: artificial neural networks and grey systems models for river classification, k-nearest-neighbor approaches paired with particle swarm-optimized LSTMs, variational mode decomposition combined with sparrow search algorithms and support vector machines on the Yangtze, and EEMD paired with LSTM networks for the Xiaofu River. Hybrid decomposition-plus-deep-learning pipelines have also proven themselves far beyond water quality, including hybrid EMD-LSTM models for non-stationary wave prediction in offshore China and EEMD-CNN-LSTM hybrids for crude oil futures. What distinguishes the new work is the explicit optimization of the decomposition stage itself, rather than leaving those parameters to trial and error.

Beyond raw accuracy, the authors emphasize a set of design principles that make the workflow unusually transparent for a deep learning study. The framework provides entropy-guided parameter adaptation so that the decomposition is not a black box bolted onto the front of the network; a causal multi-scale representation so that no information from the future leaks into the past; component-wise learning so that the network’s burden is divided into tractable sub-problems; and timestamp-aligned aggregation so that the reconstructed forecast lines up exactly with the physical timeline of the measurements. Perhaps most notably, they highlight an auditable, leakage-controlled evaluation process, an issue that has dogged time-series machine learning generally, where subtle data leakage between training and test sets can inflate reported accuracy dramatically.

That emphasis on reproducibility matters because dissolved-oxygen forecasting is not an academic exercise. Oxygen levels function as a primary indicator for classifying river health and assessing urban impacts, as established in classic environmental monitoring work dating back to 2007. Machine learning models of oxygen concentration have been applied with dimensionality reduction techniques and with increasingly sophisticated architectures, but deployment decisions, such as issuing an aeration alert at an aquaculture facility or triggering an investigation into an illegal discharge, require forecasts that regulators can trust and verify. A framework that documents its random seeds, controls for data leakage, and decomposes its predictions into interpretable components offers an audit trail that a single opaque network trained on raw sensor data simply cannot provide.

The authors are careful about the limits of what they have demonstrated. Their evaluation rests on two sites, one estuarine and one from a public repository, and they explicitly position the current locked-seed results as a basis for later multi-seed and multi-site validation rather than a final verdict. This matters because time-series models can be sensitive to the specific regime they are trained on: an estuary dominated by tidal mixing behaves differently from a reservoir stratified by season. The Xiaojiang dataset is available from the corresponding author on reasonable request, while the Refuge dataset is fully public, which at least gives independent researchers a path to test the pipeline for themselves. The authors also report that no specific financial support was received for the study and declare no competing interests.

Still, the direction of travel is clear. As climate change intensifies the swings between flood-driven nutrient pulses and drought-driven stagnation, the oxygen dynamics of rivers and estuaries are becoming harder to anticipate with simple statistical models, and the cost of a missed hypoxic event keeps rising. Hybrid pipelines that combine physics-aware signal decomposition, metaheuristic optimization, and deep temporal learning represent a pragmatic middle path: they exploit the strengths of modern artificial intelligence while keeping the signal-processing foundations visible and tunable. If follow-up studies confirm that WOA-optimized decomposition holds up across many sites and seasons, the framework described by Xie and Liang could become a template for the next generation of early-warning systems, giving water managers a few precious hours of foresight before the oxygen disappears and the fish follow.

Subject of Research: Machine learning prediction of dissolved oxygen in water-quality monitoring time series

Article Title: A hybrid framework for water-quality prediction using WOA-optimized EEMD and CNN-LSTM

Article References: A hybrid framework for water-quality prediction using WOA-optimized EEMD and CNN-LSTM. (n.d.). https://doi.org/10.1007/s10661-026-16043-0

Image Credits: AI Generated

DOI: 10.1007/s10661-026-16043-0

Keywords: dissolved oxygen, water quality, deep learning, CNN-LSTM, ensemble empirical mode decomposition, whale optimization algorithm, time-series forecasting, signal decomposition, environmental monitoring, metaheuristics, estuary, machine learning