Quality-controlled and gap-filled daily agro-meteorological dataset from ten automatic weather stations in north-eastern Morocco (2018–2026)
Description
This dataset provides quality-controlled, gap-free daily weather records from ten solar-powered automatic weather stations in agricultural areas of the Oriental region of north-eastern Morocco: Berkane, Berkane Slimania, Bou Areg, Lamrija (Guercif), Maatarka, Labsara, Laouinate, Mestigmer, CCA Layoun and CCA Figuig. It covers 5 December 2018 to 3 March 2026, 15,924 station-days in total, at elevations of 20 to 1204 m. Raw logger exports were harmonised into one schema of 22 variables in SI units. Ten core variables were quality-controlled value by value: air temperature (mean, maximum, minimum), relative humidity (mean, maximum, minimum), solar irradiance, wind speed (mean, maximum) and precipitation. The tests were range, solar-geometry, internal consistency, step, persistence, spatial consistency, Isolation Forest anomaly detection and rain-gauge fault detection against neighbouring stations. Rejected and missing values (3.7 % of core values) were reconstructed with spatial regression, LightGBM and a self-attention (SAITS-type) imputer, benchmarked on 15,314 withheld observations. Each value carries a QC flag (qc), its source (src: observed or imputation method) and an expected error (_unc). Daily FAO-56 Penman–Monteith and Hargreaves–Samani reference evapotranspiration are included. Files: daily_complete/ (one gap-filled CSV per station); ALL_STATIONS_daily_qc_gapfilled.csv (all stations in one table); harmonized_raw/ (raw values before QC); station_metadata.csv; data_dictionary.csv; qc_log.csv (every rejected value and the test that rejected it); qc_summary.csv; precip_month_state.csv; benchmark_metrics.csv; code/ (Python scripts 01–05 reproducing all files and figures); figures/; tables/; README.md. Uses: irrigation scheduling, crop-water and drought modelling, evaluation of reanalysis and satellite products, and benchmarking of quality-control and imputation methods. Reconstructed daily precipitation is a neighbour average, best used at monthly to annual scale; filter prcp_qc ≤ 1 for observed values only. For the other form fields, use the CC BY 4.0 licence and the categories “Meteorology”, “Agricultural Science” and “Machine Learning”. All six authors go in the same order as in the paper.