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

This paper introduces Multi-population Cognitive Evolutionary Search (MCES), a novel evolutionary algorithm for selecting complex feature interactions to predict dissolved oxygen (DO) in lakes. MCES adapts to different lake types and prediction tasks by using multiple populations and a "gene map" approach, offering more accurate DO predictions and insights into lake ecosystem dynamics.

What Problem Does This Paper Address?

Accurately predicting dissolved oxygen (DO) concentrations in lakes is crucial for understanding water quality and the health of freshwater ecosystems. DO levels are influenced by a complex interplay of factors including morphometry, weather, trophic state, and land use. While process-based models exist, they often struggle with inherent biases due to incomplete knowledge. Data-driven methods, particularly deep learning, show promise but face challenges in selecting the most relevant feature interactions from a vast number of possibilities. Existing feature selection methods often lack flexibility, failing to adapt to the diverse characteristics of different lake types and prediction tasks. Furthermore, the sparse nature of DO data makes model training difficult.

Methods

The paper proposes Multi-population Cognitive Evolutionary Search (MCES), an evolutionary algorithm designed to tackle the complex feature interaction selection problem for DO prediction. MCES builds upon the Cognitive Evolutionary Search (CELS) framework by introducing multiple populations, each tailored to specific lake types and prediction tasks (e.g., predicting DO in the epilimnion of large lakes).

Key methodological components include:

  • Multi-population Architecture: Eight distinct populations are established, each specializing in a combination of lake type (small, medium, large, extra-large) and prediction task (epilimnion or hypolimnion). This allows for adaptive learning tailored to specific ecological contexts.
  • Feature Interaction Representation: Feature interactions are treated as "genomes," and models as "organisms" within an ecosystem. Relevance parameters (α and β) are introduced to dynamically adjust the importance of features and their interactions.
  • Evolutionary Mechanisms: MCES employs mutation and crossover mechanisms. Mutation helps generate new models by altering operations associated with irrelevant interactions. Crossover occurs both intra-population (within a population) and inter-population (between populations), promoting diversity and the exchange of beneficial genetic information.
  • Model Refinement: After the feature selection stage using simulated labels, MCES refines the selected models using real-world observed DO data. A novel loss function is introduced to integrate both simulated and observed data, addressing the scarcity of real labels.
  • "Gene Maps": A visualization technique that represents the learned feature interactions, allowing for interpretation of the model's decision-making process and revealing sophisticated phenological patterns.

Datasets

The study utilizes a comprehensive dataset of ecological observations from 375 lakes in the Midwestern USA, spanning over 41 years (starting in 1979). This dataset comprises approximately 5.58 million daily records, including:

  • Phenological Features: Morphometric data (area, depth, shape), flux data (ecosystem, sedimentation), weather conditions (wind speed, temperature), trophic states, and land use proportions.
  • Observed DO Data: Sourced from the Water Quality Portal (WQP).

The dataset was partitioned for training, validation, and testing based on lake size and data availability, with specific splits for large/extra-large lakes and small/medium lakes.

Key Insights and Ideas

  • "AI from nature, for nature": The core concept of MCES is to mimic natural evolutionary and ecological processes to create AI models that are not only accurate but also interpretable and beneficial for environmental science.
  • Adaptive Specialization: The multi-population approach allows models to specialize for distinct lake types and tasks, overcoming the limitations of generic models. This is akin to how species adapt to different ecological niches.
  • Feature Interactions as Genomes: Representing feature interactions as genomes provides a powerful framework for evolutionary selection and allows for the discovery of complex, non-linear relationships that individual features might miss.
  • Interpretability through Gene Maps: The visualization of learned feature interactions as "gene maps" offers a unique way to understand why the model makes certain predictions, revealing ecological patterns and the importance of specific interactions.
  • Synergy of Simulated and Real Data: The paper demonstrates the effectiveness of combining simulated data (generated from process-based models) with scarce real-world observations for robust model training.

Key Results

  • Superior Prediction Accuracy: MCES consistently outperformed all baseline models, including process-based models and other advanced machine learning approaches (LSTM, EA-LSTM, KGSSL, AutoInt, AutoGroup, AutoFeature), across various lake types and tasks.
  • Ablation Study Validation: Ablation studies showed that the multi-population architecture, inter-population crossover, and the model refinement stage (using real labels) all contribute significantly to MCES's performance.
  • Revealing Ecological Patterns: The gene maps provided insights into how DO dynamics are influenced by different factors in various lake types. For instance, larger lakes are more influenced by sediment oxygen demand and atmospheric exchange, while smaller lakes are more sensitive to local land use and meteorological factors. Temperature-related interactions were found to be significant across the board.

Limitations and Future Work

  • Computational Cost: Evolutionary algorithms, especially multi-population ones, can be computationally intensive. While not explicitly stated as a limitation, this is an inherent characteristic.
  • Generalizability: While evaluated on Midwestern lakes, further validation on lakes from different geographical regions with diverse characteristics would be beneficial.
  • Future Directions: The authors encourage further exploration and development of new algorithms inspired by MCES, aiming to benefit environmental science and freshwater management. The concept of "AI from nature, for nature" suggests potential for broader applications in ecological 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