Skip to contents

Impute Country-Week Tropical-Cyclone Probability from EM-DAT and Climate

Source: R/impute_cyclone_probability.R
impute_cyclone_probability.Rd

Fits a binomial GAM on observed EM-DAT tropical-cyclone / storm-surge events (emdat_cyclone_active, 0/1) as a function of a wind/coastal mechanistic predictor set that is deliberately distinct from the precipitation-driven flood GAM. The key proxy is the weekly maximum 10m wind speed (wind_speed_10m_max); precipitation channels capture the co-arriving surge/rain, and ENSO/IOD teleconnection lags capture the seasonal steering of cyclone tracks into the Southwest Indian Ocean / Mozambique Channel. The fitted model predicts a continuous cyclone probability for every (iso_code, year, week) row – historical AND forecast – so the downstream psi LSTM consumes the same feature definition in training and inference.

Usage

impute_cyclone_probability(
  d,
  output_col = "emdat_cyclone_prob",
  gam_train_stop = NULL,
  diagnostics = TRUE,
  diag_dir = NULL,
  verbose = TRUE
)

Arguments

d

A data.frame with one row per (iso_code, year, week). Must contain the columns returned by .impute_cyclone_probability_required(): iso_code, year, week, date; the binary target emdat_cyclone_active (0/1; NA in forecast rows is fine); the WHO subregion region; the storm proxy wind_speed_10m_max; precipitation channels precipitation_sum, precip_sum_2w, precip_sum_4w; and the teleconnection indices ENSO34, IOD (lags computed inline by the function).

output_col

Character. Name of the new probability column. Default "emdat_cyclone_prob".

gam_train_stop

Date or character ("YYYY-MM-DD") or NULL. When non-NULL, the binomial GAM is FIT only on rows with date <= gam_train_stop; the fitted model then PREDICTS every row (rows <= gam_train_stop in-sample, rows > gam_train_stop leak-free extrapolation). The leakage-hygiene hook for rolling-origin forecast CV. Default NULL = current full-data fit (back-compatible). Only the fit-row subset changes; select=TRUE/fREML/the formula are identical.

diagnostics

Logical. If TRUE, fits a rolling-year cross-validation and writes diagnostic artefacts to diag_dir.

diag_dir

Character. Directory for diagnostic outputs (created if missing). Ignored when diagnostics = FALSE.

verbose

Logical. Echo progress messages.

Value

The input data.frame with one added column named output_col carrying a non-NA numeric in [0, 1] for every row.

Details

Why a separate GAM (not folded into the flood label). Tropical cyclones carry the opposite climate signature from ordinary precipitation floods: a sharp wind spike with a short, crisply-dated footprint. The flood-only EM-DAT filter historically dropped all 49 tropical-cyclone / surge events, which neutered the flood GAM's wind smooth. Modeling cyclones with their own wind-led GAM restores that signal as a clean input channel; the psi LSTM recombines the flood and cyclone probabilities downstream (the shared cholera-contamination mechanism lives in the LSTM, not in the label).

The GAM is fit with mgcv::bam(), family = stats::binomial(link = "logit") (predictions naturally in [0, 1]), method = "fREML", select = TRUE (smoothness null-space shrinkage – drives uninformative smooths toward zero, the same principled mechanism used by impute_flood_probability). Predictors computed inline (the caller need not supply them):

  • ENSO34 and IOD at lags 8, 16, 24 weeks (seasonal cyclone-season steering; SWIO cyclone activity co-varies with ENSO/IOD phase).

  • Nothing else – the wind and short-window precip channels are the acute proxy; the country random effect absorbs baseline coastal-exposure differences.

The active fraction is very low (~74 weeks out of ~35k country-weeks, ~0.2%). The country random effect s(iso_code_f, bs = "re") concentrates predicted probability on the historically cyclone-exposed coastal countries (MOZ, MWI, ZWE, MDG-adjacent belt) and keeps landlocked / non-cyclone countries near zero. A region-conditional wind smooth (s(wind_speed_10m_max, by = region_f)) lets Southern Africa (the cyclone belt) map wind to risk more steeply than the other subregions.

Sentinel handling mirrors the flood imputer: any NA prediction (lag warm-up edge, residual missing predictor) is filled with that country's mean predicted probability; global mean is the last resort. The function asserts the final column is numeric, in [0, 1], and NA-free.

If diagnostics = TRUE, a rolling-year CV (3 most-recent fully-observed years) writes per-fold AUC / Brier / log-loss to cyclone_gam_cv_metrics.csv. A mean CV AUC below 0.65 emits a warning but the column is still returned.

See also

impute_flood_probability, compile_suitability_data, process_EMDAT_data