Transferability of Machine Learning Models for Geogenic Contaminated Groundwaters

Area of Science:

• Environmental Science
• Hydrogeology
• Machine Learning Applications

Background:

• Machine learning (ML) models are effective for identifying geogenic contaminated groundwaters.
• Developing ML models in data-scarce regions is challenging due to the requirement for extensive training datasets.
• Model transferability offers a potential solution for applying ML in regions with limited available data.

Purpose of the Study:

• To investigate the transferability of high fluoride groundwater ML models between different basins within the Shanxi Rift System.
• To evaluate the influence of six factors: modeling methods, predictor types, data size, sample/predictor ratio (SPR), predictor range, and data informing on model transferability.
• To identify key factors that enhance or hinder the successful transfer of ML models for groundwater contamination assessment.

Main Methods:

• Exploration of ML model transferability using hydrochemical and surface parameters as predictors.
• Assessment of the impact of data informing (adding data from target regions to training sets) on transferability.
• Statistical analysis (stepwise regression) to determine significant factors influencing transferability.
• Application of the t-distributed Stochastic Neighbor Embedding (t-SNE) algorithm for data visualization and comparison across basins.

Main Results:

• Successful transferability of high fluoride groundwater models was achieved only when predictors were based on hydrochemical parameters, not surface parameters.
• Data informing, by incorporating samples from challenging regions, significantly enhanced model transferability.
• Stepwise regression confirmed hydrochemical predictors and data informing as significant positive factors, while data size, SPR, and predictor range showed no significant effect.
• Advanced ML models (random forests, artificial neural networks) did not consistently outperform logistic regression in terms of transferability.
• t-SNE analysis highlighted the critical role of predictor types in enabling effective data representation and model transfer across basins.

Conclusions:

• Hydrochemical parameters are crucial for the successful transferability of machine learning models in groundwater contamination studies.
• Data informing is a vital strategy for improving the performance of transferred models in data-limited or challenging hydrogeological settings.
• The choice of predictor type is more critical for model transferability than the complexity of the machine learning algorithm used.