Data and Core Code for Temporal Evaluation of Decision-Calibrated Machine Learning for River Water-Quality Risk Screening in Open Monitoring Networks

This dataset contains processed data, a model-ready table, tabulated analysis outputs, and core Python scripts. The files support the article titled “Temporal Evaluation of Decision-Calibrated Machine Learning for River Water-Quality Risk Screening in Open Monitoring Networks.” The model-ready table contains 67,518 station-date-target records from 300 monitoring locations in four Mid-Atlantic and Chesapeake-related HUC4 basins (0205-0208), covering 1 January 2021 to 2 April 2026. Each record represents one monitoring station, sampling date, and water-quality endpoint after aggregation of duplicate observations by the median. The five endpoints are dissolved oxygen, pH, turbidity, nitrate + nitrite, and phosphorus. Public source records come from the Water Quality Portal, NASA POWER, the USGS National Water Information System, EPA StreamCat, and EPA ecoregion resources. Endpoint-specific empirical thresholds calculated from 2021-2023 observations define the binary risk labels. The chronological split contains 50,203 training records from 2021-2023, 13,725 validation records from 2024, and 3,590 latest-period test records from 2025-2026. Predictors describe monitoring-site and catchment attributes, calendar timing, prior measurements within each station-target series, and antecedent meteorology. The data archive includes the modeling table, risk thresholds, station inventory, processed source-data derivatives, and tabulated analysis outputs. These outputs cover model comparison, probability calibration, threshold sensitivity, error stratification, grouped permutation importance, fixed-budget screening, bootstrap uncertainty, HUC4 performance, hydrometeorological context, and station concentration. The code archive contains two scripts that build the modeling table and train and evaluate a seasonal station-month baseline, regularized logistic regression, random forest, and histogram-based gradient boosting models with isotonic calibration. The scripts require Python with pandas, NumPy, and scikit-learn. Raw source records remain available from the original public providers. The code archive covers the central modeling chain and excludes data-download, figure-generation, high-performance-computing submission, and manuscript-formatting utilities.