Supplement to: 'AI learns to forecast global flood events from historical news'

Résumé fourni par la source

AI Flood Forecasting Research Repository This repository provides the code, models, and data summaries for the paper: "AI learns to forecast global flood events from historical news". (1) Directory Structure The directory is organized as follows: models/: Contains trained deep learning models exported as tf.SavedModel files for inference. predictions/: Includes model forecasts for the evaluated years (2021–2025). This folder contains: global.zarr: Predictions for the global operational domain. leave_out_{country}.zarr: Zero-shot predictions where the specified country was held out from training. labels/: Ground-truth binary labels extracted from the Groundsource dataset. geometries/: GeoParquet files defining the 0.2∘×0.2∘ prediction grid cropped to a low-resolution landmass map. scripts/: Python notebooks and standalone scripts for reproducing the analysis. (2) Script Functionalities The following scripts are provided to facilitate reproduction and further research: GDACS download.ipynb.ipynb: Fetches flood event metadata (alert levels, dates) and geometries from the Global Disaster Alerts and Coordination System (GDACS) API. Label maker.ipynb: A utility to convert flood event polygons (from Groundsource or GDACS) into binary label matrices. It maps events to target pixels and time bins using configurable overlap thresholds. The input GeoDataFrame must include start_time and end_time (Unix timestamps in seconds) and a geometry column in EPSG:4326. Paper figures.ipynb: The primary visualization notebook. It calculates Precision-Recall (PR) curves, F1 scores, and generates the spatial performance maps shown in the manuscript. (3) Data Requirements and Formats GeoParquet (Grid & Events) pixel_sites.parquet: Requires columns geometry, min_lat, min_lon, pixel_name, and country_codes (ISO alpha-3). Can be generated and cached by Paper figures.ipynb. Zarr / xarray (Forecasts & Labels) Forecasts and label files are stored in Zarr format with the following structure: Coordinates: forecast_time: float64 Unix timestamps (hourly or daily). area_id: string identifiers matching the pixel_name in the grid. Variables: forecast: Continuous values representing flood likelihood scores. label: Boolean (0/1) indicating the occurrence of a newsworthy flood. (4) Benchmark Data Sources To compare model performance against established early warning systems, baseline alerts can be downloaded from: United States (NWS): Warning archives accessible via the Iowa Environmental Mesonet (NWS VTEC). Brazil (CENAD & INMET): Historical alerts from S2ID and thresholds/measurements from BDMEP. United Kingdom (Environment Agency): Historical warnings via the Real-time flood monitoring data archive. Taiwan (NCDR): Historical alert data from the NCDR Open Data History. After downloading, use the Label maker.ipynb script in order to covert it into a zarr that can be used the the Paper figures.ipynb notebook. (5) License This repository and the associated data summaries are licensed under Creative Commons Attribution 4.0 International (CC BY 4.0).

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.