AquaCorr-Global: A Multi-Sensor, Multi-Decadal Water Quality Matchup Dataset for Inland and Coastal Waters

AquaCorr-Global pairs 12,504,399 in situ water quality records from 372,639 monitoring sites worldwide with 1,984,387 same-calendar-day satellite matchups, drawn from six sensors (Landsat 5/7/8/9, Sentinel-2A/B) over 1984 to 2024. All satellite reflectance is processed through a single ACOLITE Dark Spectrum Fitting framework, producing aquatic-optimized remote sensing reflectance (Rrs, sr⁻¹) rather than land-optimized surface reflectance. In situ data are harmonized from six sources (AquaMatch, ChemLotUS, GLORIA, SeaBASS, AquaSat v1, Caravan-Qual) across seven parameters (chlorophyll a, total suspended solids, Secchi disk depth, CDOM absorption, turbidity, and total and dissolved organic carbon). Validation against co-located GLORIA in situ hyperspectral reflectance shows that the ACOLITE-corrected reflectance closely reproduces field measurements across visible bands, in contrast to land-optimized surface reflectance products. Files aquacorr_matchups.parquet: main matchup table (GeoParquet, EPSG:4326, cloud-optimized: sorted by site, 16 row groups, ZSTD compression). aquacorr_matchups.csv.gz: gzip-compressed CSV export of the main matchup table. aquacorr_rrs_spectra.zarr.zip: Zarr tensor archive (site × observation × channel) with harmonized six-band Rrs, additional ACOLITE-derived bands, water-quality values, and quality flags; unzip to a .zarr directory and open with xarray or zarr. aquacorr_gloria_validation.parquet: GLORIA radiometric validation subset (in situ hyperspectral Rrs, ACOLITE Rrs, and land-optimized surface reflectance at co-located points). aquacorr_metadata.json: machine-readable dataset metadata and file checksums. Quality flags (three systems) Pipeline flags (satellite side): flag_negative_rrs, flag_high_cv, flag_insufficient_pixels, flag_high_sza, flag_acolite, and the composite passed_qc (935,740 matchups pass all five). Source flags (in situ side): one bit-packed integer per parameter, {param}_srcflag, following the Landsat QA_PIXEL convention. Bits encode method class (extraction/lab vs. in vivo/sensor), quality tier, sample fraction, depth type, outlier status, and a recommended bit. Decode with bitwise operations (see below). Coordinate provenance: coord_centroid = 1 marks the 98,165 matchups whose in situ coordinate is a lake centroid (LAGOS-NE via AquaSat v1) rather than a sampling location. Decoding source flags (Python) method_class = chla_srcflag & 0b11 # 0 unknown, 1 extraction/lab, 2 in vivo/sensor, 3 curated quality_tier = (chla_srcflag >> 2) & 0b11 # 0 n/a, 1 restrictive, 2 narrowed, 3 inclusive recommended = (chla_srcflag >> 9) & 1 # 0 no, 1 yes Notes on use Chlorophyll a analytical method is resolved for a subset of records. Extraction-based (HPLC, laboratory spectrofluorometry) and in vivo fluorometry are not interchangeable; users developing or validating retrieval algorithms should prefer the extraction-based (recommended) subset. Over narrow rivers and streams, the 3 × 3 extraction window may include shoreline pixels; screen with a shortwave-infrared water-purity criterion for optical-band analyses. All matchups are retained rather than pre-filtered, so users can apply quality thresholds appropriate to their application.