Skip to contents

Process EM-DAT Flood and Tropical-Cyclone Events Into Country-Week Panels

Source: R/process_EMDAT_data.R
process_EMDAT_data.Rd

Reads the most recent EM-DAT public-extract xlsx file in PATHS$DATA_EMDAT_RAW, filters to (a) flood events and (b) tropical-cyclone / storm-surge events in MOSAIC AFRO countries, expands each event across the ISO-8601 weeks it spans, and writes two complete country-week panels of hazard-occurrence covariates: PATHS$DATA_EMDAT/floods_country_weekly.csv and PATHS$DATA_EMDAT/cyclones_country_weekly.csv.

Usage

process_EMDAT_data(PATHS)

Arguments

PATHS

A list of paths generated by get_paths(). Must include:

  • DATA_EMDAT_RAW: Path to MOSAIC-data/raw/EMDAT.

  • DATA_EMDAT: Path to the processed output directory (MOSAIC-data/processed/EMDAT/weekly).

Value

Invisibly returns a named character vector with the paths to the two written CSVs (flood and cyclone). Each CSV contains one row per (iso_code, year, week) for every ISO-week between date_start (default 2000-01-03) and the latest week covered by any event of that hazard in the raw extract, restricted to iso_codes_mosaic. Flood columns:

iso_code

ISO3 country code.

year

ISO-8601 year of the week.

week

ISO-8601 week number (1-53).

date_start

Monday of the ISO-week.

date_stop

Sunday of the ISO-week.

emdat_flood_active

1 if any flood event overlaps the week, else 0.

emdat_flood_new

Number of flood events that start in this week.

emdat_flood_affected

log1p of summed Total Affected across overlapping events (NAs treated as 0).

emdat_flood_deaths

log1p of summed Total Deaths across overlapping events (NAs treated as 0).

The cyclone CSV mirrors this schema with emdat_cyclone_active, emdat_cyclone_new, emdat_cyclone_affected, and emdat_cyclone_deaths.

Details

The flood label is unchanged (Disaster Type == "Flood"). The cyclone label is a SEPARATE series built from Disaster Type == "Storm" restricted to Disaster Subtype %in% c("Tropical cyclone", "Storm surge") – the crisply-dated, water-producing storm events (49 events / ~74 active weeks in the 2026-05 extract). Storm (General), Severe weather, Lightning/Thunderstorms, Tornado, Hail, and Blizzard/Winter storm are deliberately EXCLUDED: they are not reliably water-producing and would dilute the cyclone signal. Tropical cyclones were previously silently dropped by the flood-only filter, neutering the downstream flood GAM's wind smooth; the separate cyclone series feeds a dedicated wind/coastal GAM (impute_cyclone_probability).

EM-DAT records carry frequent missing day-of-month and occasional missing end-month values. Imputation rules:

  • Missing Start Day or End Day -> 15.

  • Missing End Month -> Start Month.

  • Missing End Year -> Start Year.

  • If the resulting end_date is earlier than start_date (data-entry artifact), it is reset to start_date.

Each event is expanded to the unique set of ISO-8601 (year, week) pairs touched by the daily date range start_date:end_date. The panel grid is the full Cartesian product of iso_codes_mosaic and all ISO-weeks from date_start through the latest event end-week; cells with no flood activity carry zeros, so a downstream left-join against the suitability table does not need to impute missingness.

EM-DAT only records events meeting CRED's inclusion thresholds (>=10 deaths, >=100 affected, or formal emergency declaration), so this covariate captures major flood occurrences rather than continuous flood intensity. The richer continuous signal lives in the OpenMeteo precipitation/anomaly covariates.