A 1-km Gridded Decadal (2016–2024) Inland Water Quality Dataset for South Korea

Daily 1 km grids of six inland water quality parameters over South Korea, 2016–2024. Retrieved from Harmonized Landsat Sentinel-2 (HLS) S30 surface reflectance combined with ERA5-Land reanalysis, terrain and land cover, and validated against the Water Environment Information System (WEIS) monitoring network. This record holds the dataset, the code that produced it, and the trained models. It accompanies a Data Descriptor submitted to Scientific Data. Contents The product — six archives, one per parameter, each holding 1,012 daily rasters: kr_wq_wt_1km_2016-2024.zip — water temperature (°C) kr_wq_ss_1km_2016-2024.zip — suspended solids (mg L⁻¹) kr_wq_chla_1km_2016-2024.zip — chlorophyll-a (µg L⁻¹) kr_wq_do_1km_2016-2024.zip — dissolved oxygen (mg L⁻¹) kr_wq_tn_1km_2016-2024.zip — total nitrogen (mg L⁻¹) kr_wq_tp_1km_2016-2024.zip — total phosphorus (mg L⁻¹) Each unpacks to a folder named after the parameter, e.g. chla/kr_wq_chla_1km_20160106.tif. Zenodo stores files flat, so the per-parameter folders are delivered as ZIP archives. 6,072 rasters in total: six parameters on each of 1,012 dates, 517 MB unpacked. The code — code_v1.0.zip: the full pipeline (thirteen notebooks plus 2_extract_7x7.py, extract_dem_lulc.py and remote_sensing/HLS.py), together with the static inputs it requires (the negative-buffered national boundary raster, the five-basin study-area polygon, and the 63-feature order) and an environment specification (requirements.txt) whose scikit-learn version is pinned, so the cross-validation folds are reproduced identically. The models are provided in two archives that serve different purposes: deploy_models.zip — the six deployment models, one per WQP, that generated the released grids. It holds six trained checkpoints, six fitted input scalers, feature_meta.json, and deploy_summary.csv. Each checkpoint is self-describing, carrying its feature order, architecture, target transform and uncertainty calibration factor alongside the weights. deploy_summary.csv records, per WQP, the early-stopping epoch at which the deployment model was refit, the smearing factor applied when back-transforming the log-trained targets, the uncertainty calibration factor c, the training R², and the mean predictive sigma. These are the values that reproduce the released mean and uncertainty fields. These models are fit on all matchups, so every monitoring point is in their training set; they are what to use to reproduce or extend the product, but not what to measure skill with. mlp_folds.zip — the 30 cross-validation checkpoints (six WQPs × five folds) from the station-held-out StratifiedGroupKFold that produced the validation figures in the Data Descriptor. Files are named mlp_<target>_fold<1-5>.pt; each carries its features order, fold index, log_target flag and state_dict. Because each fold was trained with a fifth of the stations held out, these are the models behind the reported skill: use them, not the deployment models, to reproduce the out-of-fold predictions, since the deployment models have seen every station. The explicit station-to-fold assignment for these checkpoints ships flat as station_fold_map.csv, so the out-of-fold split can be checked without rerunning the pipeline. The evaluation — oof_predictions.csv: the out-of-fold predictions every skill figure in the Data Descriptor is computed from. One row per monitoring point, date and parameter, 281,598 rows in all (46,933 matchups × six WQPs), holding ptNo, obs_date, st_type, parameter, val_fold, the signed offset delta_days, the observation y_true, the point prediction y_pred, and the Monte Carlo dropout mean mu_pred with both its uncalibrated sigma_pred and its calibrated sigma_cal. Every number in Tables 6, 8, 9 and 10 is a groupby on this one file with no model loaded: within-point R² by removing each point's own mean from the observation and prediction columns before scoring, conditional coverage as the fraction of rows inside mu_pred ± 1.96 × sigma_cal, and window sensitivity by binning on abs(delta_days). Use sigma_cal for interval work, not sigma_pred — the calibration factor is already applied in that column, and scoring with the raw one returns marginal coverage near 0.46–0.75 rather than the reported ~0.95. The same shape ships for the other comparisons. baseline_oof_predictions.csv carries five sets of out-of-fold predictions in a model column: RF and XGB, the tree baselines of Table 5; static and climatology, the two deployable satellite-free baselines of Table 7; and all_features_ref, the ablation harness's own all-feature control, which is not the released MLP (the harness refits under its own settings, so its scores run above the released figures). ablation_oof_predictions.csv holds the six feature configurations behind the feature-group analysis. kr_wq_1km_metadata.json, station_fold_map.csv (the station-to-fold assignment behind the reported skill), oof_predictions.csv (see below) and README_zenodo.md sit flat in the record. Read the README before using the grids; the limitations below are not optional reading. The grid Period: 6 January 2016 – 29 December 2024 (1,012 dates) Grid: 470 rows × 510 columns, EPSG:4326, 0.00996462647316037° Extent: 124.589°–129.671°E, 33.941°–38.625°N Format: GeoTIFF, Float32, LZW + predictor 3, internally tiled 256×256 with overviews; NoData = NaN Bands: 1 = mu_pred (predictive mean), 2 = sigma_cal (calibrated predictive uncertainty) File naming: kr_wq_<param>_1km_<yyyymmdd>.tif, <param> ∈ {wt, ss, chla, do, tn, tp} Of the 239,700 grid cells, 99,950 fall inside the negative-buffered national boundary and 13,469 of those contain permanent water the pipeline can predict. The grid is 1 km, but only water cells ever carry a value. The domain is the five major river basins of the mainland (Han, Nakdong, Geum, Seomjin, Yeongsan). Because the grid begins at 33.941°N, Jeju Island lies entirely south of it and is absent from every raster: Jeju drains independently of those five basins and its water bodies were excluded at the point of station selection, so this is a documented scope limit rather than a gap. A date carries a median of 1,182 cells. A further 24 dates inside the period produced no qualifying cell and are absent from the archive rather than present and empty — a missing file is a real gap, not a packaging error. The six layers differ in kind, not only in accuracy Four independent lines of evidence — feature ablation, SHAP ranking, skill decomposition and within-cell dispersion — agree on the same grouping: Optical retrieval (SS, Chl-a, TP): the satellite reflectance carries the dominant signal. Catchment context (TN): estimated from catchment and land-cover context rather than from the appearance of the water. Meteorologically driven (WT, DO): driven principally by the ERA5-Land forcing. A TN grid is not a measurement of color, and reading it as one will mislead. Match each layer to applications appropriate for what it actually is. Uncertainty Band 2 comes from 100 Monte Carlo dropout passes, rescaled per parameter so that mu_pred ± 1.96 × sigma_cal is an approximate 95 % interval. For SS, Chl-a, TN and TP the symmetric lower limit falls below zero and should be truncated there. For SS, Chl-a, TN and TP a properly asymmetric 95 % interval can be built from the same two bands without rerunning anything. kr_wq_1km_metadata.json carries per-parameter multipliers (uncertainty_asymmetric, keys q_lower/q_upper) that apply directly to sigma_cal: lower = mu_pred + q_lower × sigma_cal (floored at 0), upper = mu_pred + q_upper × sigma_cal. They are the lower (2.5 %) and upper (97.5 %) standardized-residual quantiles expressed per unit of sigma_cal (the raw quantiles and the calibration factor they were divided by are kept under _provenance). Validated on all 1,529,676 released pixels, this removes the below-zero lower limit for essentially every SS, Chl-a, TN and TP pixel while holding 95 % empirical coverage, and the band is narrower than the symmetric form (upper-to-lower width ratio ≈ 2.2 for TN to 2.9 for Chl-a). WT and DO stay effectively symmetric (ratio ≈ 0.97) and are not floored. The symmetric mu_pred ± 1.96 × sigma_cal is kept as the default because it needs no extra constant. Within a 1 km cell, sigma_cal is the plain mean of the contributing 30 m uncertainties. It is not divided by √n: the 30 m predictions inside one cell share a reanalysis cell, terrain and land cover, so they are nowhere near independent and that reduction would understate the uncertainty badly. Known limitations The upper range is compressed. The grids locate elevated conditions well but understate how elevated they were: at the top of the observed range the predictions recover 0.83 of the departure from typical conditions for WT but only 0.36–0.58 for the other five. Use a cell as a screening result and take event magnitude from the in-situ record. Streams narrower than about 60 m are absent. At 30 m native resolution such a body is a mixed pixel throughout, and the one-pixel shoreline erosion that suppresses land leakage removes it entirely. This is a sensor limit, not a pipeline defect. Monitoring points are not always on a predicted cell. Of the 1,194 points behind the training set, 312 (26.1 %) fall on a 30 m cell carrying no prediction, rising to 45.5 % for river points. Sample a small neighborhood rather than the single cell holding a coordinate. Summer is under-represented. The monsoon puts cloud over the peninsula exactly when blooms occur; July has the sparsest coverage of any month. A date is an overpass, not a day. The series is irregular. Do not resample it onto a daily index without deciding what the gaps mean for your application. The model was trained on inland water and is not tuned for optically complex estuarine water; mask the southwestern coastal/estuarine area for freshwater-only applications. License The data and models are released under CC BY 4.0. The code inside code_v1.0.zip is released under the MIT license; see the LICENSE file in that archive. CC BY rather than CC0 because the product derives from inputs that themselves require attribution. If you redistribute or build on these grids, carry the acknowledgements the upstream providers require: Contains modified Copernicus Climate Change Service information (ERA5-Land) and Copernicus DEM data. Contains information from the Water Environment Information System and the Environmental Geographic Information Service of the Republic of Korea. Produced using data from the Harmonized Landsat Sentinel-2 project (NASA/USGS). Input data None of these are redistributed here. Obtain them from the providers: HLS S30 surface reflectance — NASA Earthdata, search.earthdata.nasa.gov ERA5-Land hourly reanalysis — Copernicus Climate Data Store, cds.climate.copernicus.eu WEIS in-situ water quality — National Institute of Environmental Research, water.nier.go.kr Copernicus GLO-30 DEM — OpenTopography, doi.org/10.5069/G9028PQB Land use / land cover — Environmental Geographic Information Service, egis.me.go.kr Global Surface Water Occurrence — JRC (Pekel et al. 2016), global-surface-water.appspot.com Citation Please cite both the Data Descriptor and this dataset, so that corrections propagate through the citation graph.