From TIROS-1 to impact: global daily landslide susceptibility mapping in data-sparse regions

From TIROS-1 to today’s satellites, rainfall products enable near-real-time monitoring of extreme precipitation, yet landslide early warning is constrained by sparse inventories and weakly transferable surface context. We test whether adding a single global terrain proxy (slope) to satellite-rainfall features improves daily landslide susceptibility assessment in a new region without local training or threshold recalibration. Using NASA COOLR reports (2019–2024), CHIRPS v2.0 precipitation, and Copernicus DEM (GLO-30) slope, we frame prediction as daily 0.1° binary classification for the day t using rainfall summaries over the prior 1–14 days. Because COOLR is event-only, we sample pseudo-absences with spatiotemporal constraints (≥30 km from events; ±7-day exclusion) at ~5 negatives per positive. An L2-regularized logistic regression is trained globally and evaluated on a geographically withheld Southern Africa holdout (N = 535). Rainfall-only predictors yield PR-AUC = 0.912 (precision = 0.959; recall = 0.676). Slope fusion delivers the largest gain (PR-AUC = 0.939; F1 = 0.874), increasing recall to 0.790 while maintaining precision at 0.976. Metrics are point estimates and may reflect reporting/accessibility bias in event-only inventories and pseudo-absence sampling. Overall, a single global susceptibility proxy materially improves transferable daily nowcasts for first-pass alert triage using open inputs.