Evolution-based Feature Selection for Predicting Dissolved Oxygen Concentrations in Lakes

Runlong Yu, Robert Ladwig, Xiang Xu, Peijun Zhu, Paul C. Hanson, Yiqun Xie, Xiaowei Jia

TL;DR

Predicting dissolved oxygen (DO) levels in lakes is vital for monitoring ecosystem health, but it is notoriously difficult due to complex environmental interactions and sparse real-world data. This paper introduces Multi-population Cognitive Evolutionary Search (MCES), a novel evolutionary algorithm that adaptively selects the most relevant feature interactions for different lake types and prediction tasks. By combining process-based physical simulations with machine learning, MCES achieves state-of-the-art prediction accuracy and generates interpretable "gene maps" that reveal distinct ecological patterns.

What Problem Does This Paper Address?

Dissolved oxygen (DO) is a critical indicator of water quality and aquatic biodiversity. However, accurately predicting DO concentrations is highly challenging. A lake's DO levels are driven by a complex web of internal and external factors, including weather, land use, morphometry (depth and area), and metabolic processes like photosynthesis and respiration.

Traditional process-based models rely on physics and ecology principles but suffer from inherent biases due to incomplete physical knowledge. Conversely, purely data-driven deep learning models struggle because real-world DO measurements are incredibly sparse and expensive to collect. Furthermore, different types of lakes (e.g., deep, massive lakes vs. shallow, small lakes) exhibit entirely different phenological behaviors. Existing automated feature selection methods typically apply a single global model, failing to adapt to these diverse ecological settings and often introducing noise by indiscriminately handling all feature interactions.

Methods

The authors propose a two-stage framework called Multi-population Cognitive Evolutionary Search (MCES) that embodies the philosophy of "AI from nature, for nature."

Fig. 1. Overall framework.
  1. Lake Classification: As a preprocessing step, lakes are clustered using a balanced K-means algorithm into four categories based on surface area and volume: small, medium, large, and extra-large.
  2. Feature Selection Stage (Evolutionary Search): The authors establish eight distinct model populations (4 lake types $\times$ 2 target tasks: predicting the upper epilimnion layer vs. the lower hypolimnion layer). Models within these populations use Long Short-Term Memory (LSTM) networks to handle temporal data. Feature interactions are treated as genomes, and mathematical operations (like element-wise products or additions) are treated as genes.
  3. Mutation & Crossover: To prevent the population from homogenizing (a problem known as "reproductive isolation"), MCES introduces both intra-population crossover (sharing the best traits within a lake type) and inter-population crossover (probabilistically swapping advantageous genetic traits across different lake types or tasks).
  4. Sparsity Optimization: Regularized Dual Averaging (RDA) is used to drive irrelevant feature interactions to a weight of zero, leaving only the most impactful relationships.
  5. Model Refinement Stage: Because real-world data is scarce, the models are first trained on abundant data simulated by a metabolic process-based model. In the refinement stage, the discovered sparse feature structures are frozen, and the model weights are fine-tuned using real-world observed labels via a custom weighted imputation loss function.

Datasets

The model was evaluated on a massive, long-term ecological dataset:

  • Scope: Over 41 years of daily ecological observations (starting in 1979) from 375 lakes in the Midwestern USA.
  • Data Volume: Approximately 5.58 million daily records.
  • Features: 39 fields of phenological features spanning morphometry, weather conditions, trophic states, land use proportions, and ecosystem fluxes.
  • Ground Truth: Real-world observed DO concentrations sourced from the Water Quality Portal (WQP).

Key Insights and Ideas

  • AI Mirroring Nature: By structuring the machine learning architecture to mimic distinct species adapting to distinct habitats (habitats being lake types/tasks), the models naturally specialize without losing generalizable algorithmic strength.
  • Overcoming Reproductive Isolation: Standard evolutionary feature selection often gets stuck in local optima because similar models dominate a single population. The inter-population crossover mechanism acts like migration in biology, introducing fresh, successful "genetic material" from other lake types.
  • Interpretable Gene Maps: By visualizing the final sparse matrices as symmetric "gene maps," the authors could ecologically validate the AI. For instance, the maps revealed that DO dynamics in large lakes are predominantly driven by sediment oxygen demand and atmospheric exchange, whereas small lakes are highly sensitive to local land use and immediate meteorological factors.

Key Results

MCES consistently outperformed all baseline models—including standard LSTMs, physics-informed models (EA-LSTM, KGSSL), and state-of-the-art feature interaction methods (AutoInt, AutoGroup, AutoFeature)—across all lake categories and water layers.

  • Quantitative Improvements: In terms of Root Mean Square Error (RMSE), MCES achieved substantial error reductions. For example, in large lakes' epilimnion prediction, MCES achieved an RMSE of 0.988, compared to 2.298 for a standard LSTM and 1.405 for AutoFeature.
  • Ablation Validation: Removing the inter-population crossover mechanism (MCES (-inter)) or training a single global population instead of multi-populations (MCES (-multi)) caused a noticeable drop in accuracy, proving the value of the ecological population structure.

Limitations and Future Work

The paper notes a clear trade-off between feature sparsity and accuracy; pushing the RDA optimizer to select an excessively sparse set of feature interactions eventually degrades predictive capabilities, forcing the model to rely only on individual features. Additionally, the framework's initial stage relies heavily on the quality of simulated labels from a process-based model; if the underlying physical simulation contains massive systematic errors, it could misguide the evolutionary feature selection stage. Future work could explore expanding this bio-inspired architecture to other complex environmental forecasting tasks, such as harmful algal bloom predictions or river temperature modeling.

Summary updated

Abstract

Accurate prediction of dissolved oxygen (DO) concentrations in lakes requires a comprehensive study of phenological patterns across ecosystems, highlighting the need for precise selection of interactions amongst external factors and internal physical-chemical-biological variables. This paper presents the Multi-population Cognitive Evolutionary Search (MCES), a novel evolutionary algorithm for complex feature interaction selection problems. MCES allows models within every population to evolve adaptively, selecting relevant feature interactions for different lake types and tasks. Evaluated on diverse lakes in the Midwestern USA, MCES not only consistently produces accurate predictions with few observed labels but also, through gene maps of models, reveals sophisticated phenological patterns of different lake types, embodying the innovative concept of "AI from nature, for nature".
BibTeX
@inproceedings{yu2024evolution,
  author = {Runlong Yu and Robert Ladwig and Xiang Xu and Peijun Zhu and Paul C. Hanson and Yiqun Xie and Xiaowei Jia},
  title = {{Evolution-based Feature Selection for Predicting Dissolved Oxygen Concentrations in Lakes}},
  booktitle = {PPSN},
  year = {2024},
  doi = {10.1007/978-3-031-70085-9\_25},
}
Figure