A machine learning-based proxy framework for 2-MIB variation under low-frequency monitoring conditions

To address the limited monitoring frequency and sample availability of 2-MIB in drinking water source areas, this study developed a proxy characterization framework integrating small-sample modeling, high-frequency routine water quality data, and relative-level assessment. Measured 2-MIB data from 2024 to February 2026 and routine water quality data from national control sections in 2025 were used. Linear regression (LR), partial least squares regression (PLS), support vector regression (SVR), and random forest regression (RF) were compared using grid search and five-fold cross-validation to select a model for 2-MIB proxy characterization. A conventional pollution pressure index (CPPI) was constructed from routine water quality indicators for auxiliary interpretation. Measured 2-MIB, 2-MIB proxy estimates, and CPPI were then classified into relative level grades I–III for relative-level comparison. The results showed that RF achieved relatively better characterization performance among the candidate models, while the uncertainty associated with small-sample resampling was further considered through repeated cross-validation. Under a representative censored-data treatment, RF achieved a test-set R 2 of 0.524 and RMSE of 12.317 ng L −1 . The RF-generated 2-MIB proxy estimates were not regarded as equivalent to measured concentrations. The same-grade agreement rate between measured 2-MIB and 2-MIB proxy estimates was 0.764, while that between 2-MIB proxy estimates and CPPI was 0.736. Large-sample application indicated that both 2-MIB proxy estimates and CPPI generally increased in summer and autumn and decreased in winter. These findings provide a reference for relative grading and auxiliary interpretation of routine water quality pressure associated with odor-causing compounds under low-frequency monitoring conditions.