Processed Data, Trained Models, and Analysis Code Supporting "Machine Learning Reconstruction of Glider Dissolved Oxygen: Accuracy Across Missions, Depths, and Years"

This Zenodo record contains the processed data, trained models, and analysis code supporting the manuscript “Machine Learning Reconstruction of Glider Dissolved Oxygen: Accuracy Across Missions, Depths, and Years.” The files support the model evaluations, dissolved oxygen reconstructions, interannual analyses, and numerical results reported in the article and its Supporting Information. They were prepared specifically for this study and do not replace the original observational archives. The data include 41 northern Gulf of Mexico glider missions during 2013–2025, including M51 in 2025, together with ship-towed observations from the Mechanisms Controlling Hypoxia (MCH) project and shipboard profiles from the Southeast Area Monitoring and Assessment Program (SEAMAP). Twelve glider missions contributed usable dissolved oxygen (DO) measurements. The other 29 missions provided predictor observations for DO reconstruction. Processed variables include DO, temperature, salinity, pressure, in situ seawater density, optical backscatter, fluorescent dissolved organic matter (fDOM), chlorophyll-a, position, sampling time, bottom depth, and an upper-water-column stability predictor. Screening flags, missing-value indicators, and sample identifiers document observation quality, predictor availability, and membership in training and validation datasets. Separate fields identify measured DO and model estimates. This version includes training-only month–pressure climatology and ridge regression references, annual-anomaly skill relative to zero anomaly, numerical MLP selection histories, and a separate SAITS training-epoch-ceiling sensitivity test. It also includes mission-level error comparisons and diagnostics examining the influence of M34 and M13 on pooled errors. The original four-model fits and reconstruction fields, including M51, remain unchanged. The uploaded files are: 01_processed_data.zip: Quality-screened observations, model input tables, sample identifiers, prepared sequence inputs, scaling parameters, and missing-predictor filling parameters. Mission inventories, glider tracks, MCH transects, and SEAMAP station information describe observational coverage. Filling and scaling parameters were estimated from the training observations for each experiment. 02_final_trained_models.zip: Four final trained models—a multilayer perceptron (MLP), an adapted Self-Attention-based Imputation for Time Series network (SAITS), random forest (RF), and histogram gradient boosting (HGB)—together with final climatology and ridge regression reference models fitted using all available training observations. The four main models correspond to the first of five repeated fits, which supplies the reconstruction fields displayed in the article. The MLP uses a softplus output to constrain DO concentrations to be nonnegative and was trained with randomly hidden optical inputs to improve reconstruction when those inputs are missing at prediction time. The archive includes joblib and PyTorch model files, filling and scaling parameters, training settings, and software-environment information. Intermediate validation-model checkpoints are not included. 03_predictions_and_reconstruction.zip: Validation predictions and one-minute glider reconstruction tables for all five repeated fits, together with validation predictions for the climatology and ridge regression references and the separate SAITS epoch-ceiling sensitivity test. The final-third test used the first two-thirds of observations, ordered by time, within each mission, section, or annual survey for training and its final third for validation. The mission-exclusion test withheld entire glider missions from training. Sample identifiers allow predictions from the two tests to be compared on the same observations. Reconstruction tables retain measured DO where available and provide estimates from each model where DO is missing, including M51 in 2025. The additional SAITS sensitivity fit does not replace the original five-fit comparison. 04_article_statistics.zip: Validation statistics by observation source, mission, region, and pressure; monthly climatologies; annual DO and temperature anomalies; errors in reconstructed annual DO anomalies; interannual variability and trend estimates; model-specific lower-layer definitions; missing-optical-input sensitivity tests; and summaries across the five repeated fits. The mlc_additions directory contains reference-model comparisons, zero-anomaly skill scores, model-selection histories, mission-level contributions to pooled squared error, and post-hoc omission sensitivities for M34 and M13. It also includes M34 predictor-support and optical-input diagnostics, seasonal-coverage checks, and observed oxygen saturation calculations. 05_analysis_code.zip: Article-specific code for loading and applying the trained models, repeating model fits, handling missing predictors, evaluating reconstruction errors, calculating monthly climatologies and annual anomalies, estimating trends, and generating the article figures. The package includes the custom nonnegative MLP and adapted SAITS implementations, optical-input hiding methods, reference-model analyses, mission-level diagnostics, dependency lists, and instructions for using the deposited data and models. MLC_REPRODUCTION.txt describes the added analyses and maps legacy figure and table numbers to the current manuscript. Software requirements distinguish the original training environment from the analysis environment. Development logs, cluster submission scripts, and superseded analyses are not included. README.txt: File organization, principal variables, units, model-loading requirements, original data sources, and interpretation limits. The code package also contains instructions for running the included analyses. Units and annual calculations DO concentrations in the processed model tables are stored in micromoles per kilogram (µmol kg⁻¹). Concentrations and errors reported in millilitres per litre (mL L⁻¹) were calculated sample by sample using in situ seawater density. Practical salinity is unitless, pressure is in dbar, and temperature is in degrees Celsius. Reconstruction tables retain screening flags; the article uses rows marked input_qc_pass = True. Annual anomalies were calculated by subtracting the corresponding monthly climatological means and accounting for differences in mission and spatial coverage. Available monthly anomalies contributed equally to each annual estimate. Each annual series was then centered by subtracting its mean across years. Missing months were not filled, and observational coverage differed among years. Annual anomalies can be negative even though the MLP’s predicted DO concentrations are nonnegative. Annual-anomaly skill is evaluated relative to predicting zero anomaly for the same observed years. This reference is distinct from comparing errors with the observed sample standard deviation and from estimating absolute DO using a climatology fitted only to training observations. Interpretation and limitations Reconstructed DO values are model estimates, not additional oxygen measurements. Their uncertainty depends on the model, depth, mission, and availability of predictor observations. M51’s fDOM channel is treated as missing, and M51 has no measured DO for direct validation. No usable measured glider DO is available after 2021 to validate the reconstructions in 2022–2025. The main models’ settings were adjusted in preliminary experiments on the same archive, so their reported errors may be optimistic. The documented internal-selection steps used training subsets, but the comparisons are not equal-budget searches for optimally tuned model families. The M34 and M13 diagnostics retain the original fitted models and all missions in the primary validation and reconstruction; they do not independently verify sensor calibration or establish a sensor fault. The files should be interpreted using the screening criteria, model settings, validation design, and uncertainty limitations described in the article and its Supporting Information. This record does not include publication figures, intermediate model checkpoints, or complete copies of the original observational archives. Information on the original data sources and their availability is provided in README.txt.