Daily precipitation station data and code for "Coupling Precipitation Forecasting and Early Warning with Reverse-Martingale Recurrent Neural Networks"
Description
This dataset contains the daily precipitation records and the complete analysis code that reproduce every table and figure in the accompanying Journal of Hydrology article. The observational data are tidy daily-precipitation series for four contrasting climate regions, each provided as a long-format CSV (date, station_id, precip_mm) together with a station_meta.csv of station coordinates: Taiwan (monsoonal, orographic): 23 stations, 1998-2020, from the public Taiwan Historical Meteorological Observations reconstruction (Raingel/historical_weather) of Central Weather Administration CODiS records; an additional 22-station 2012-2022 subset supports the drought-warning study. Texas Hill Country (semi-arid subtropical): 47 stations, 1980-2024, NOAA GHCN-Daily. Germany (temperate maritime): 50 stations, 2005-2022, NOAA GHCN-Daily. Turkey / Anatolia (Mediterranean-continental): 44 stations, 2008-2019, NOAA GHCN-Daily. The code folder provides the Python implementation of the reverse-martingale recurrent network (RMRNN), the sequential change-point (CUSUM/SR) drought-warning detector, the data retrieval and quality-control scripts, and the domain-calibrated stochastic simulators used for the controlled synthetic experiments. The synthetic-testbed outputs are clearly labelled as such; the headline evaluation is on the real four-region station data above. The results folder holds the produced JSON/NPZ outputs, and README.md documents the run order and exact settings (station lists, year ranges, quality-control thresholds, and per-region date windows). Underlying sources and licensing: GHCN-Daily is public domain (NOAA/NCEI); the Taiwan series derive from the openly available Raingel reconstruction of public CODiS records. The tidy derivatives redistributed here are released under CC BY 4.0.
Files
Steps to reproduce
Environment: Python 3.10+ with numpy, scipy, matplotlib, torch. Unzip so that code/, data/, results/ sit side by side. Run all scripts from code/experiments/ (they import code/common/). The retrieved station CSVs are bundled under data/, so experiments run offline without re-downloading. Real forecast skill & path stability: run run_taiwan_precip.py (Taiwan) and run_ghcn_stability.py per region (Texas Hill Country, Germany, Turkey). Each compares the reverse-martingale network (RMRNN) against its lambda=0 ablation and a GRU baseline -> results/*_precip_results.npz. Deep baselines (LSTM, TCN, real data): run run_taiwan_precip_deepbaselines.py -> results/taiwan_precip_deepbaselines.npz. Real-event drought early-warning (drought-lead table + figure): run run_taiwan_drought_warning.py, and rerun per region with the recorded date windows for Texas, Germany, Turkey. These reuse the sequential change-point (CUSUM/SR) detector from run_warning_sr.py with the recorded station lists, train/null/monitor windows, deseasonalisation and ARL0 settings, and ten fixed seeds -> results/*_drought_warning_results.json. Mei-yu onset (wet counterpart): run run_taiwan_meiyu_warning.py over the 2017-2022 plum-rain seasons -> results/taiwan_meiyu_warning_results.json. Controlled synthetic warning study (drought lead, heavy-rain boundary, ARL calibration with known onset times): run run_warning_sr.py -> results/warning_sr_results.json. Gridded synthetic bake-off (vs ConvLSTM/PredRNN/U-Net): run run_gridded_bakeoff.py. Tables and figures: make_realdata_tables.py assembles the cross-landform tables and comparison figure from results/; plot_taiwan_drought_map.py renders the per-station lead map (needs network to fetch public CWA/CODiS coordinates). Exact station lists, year ranges, QC thresholds, bounding boxes, and per-region date windows are documented in README.md and the article's online supplement. To regenerate raw inputs from source instead of the bundled CSVs, use fetch_ghcnd_region.py (NOAA GHCN-Daily) and fetch_taiwan_hydromet.py (public Taiwan CODiS reconstruction).
Institutions
- Institute of Statistical Science, Academia SinicaTaiwan, Taipei