LND · Machine learning

Landslide

XGBoost · spatially blocked CV · hold-out AUC 0.97 · v1.1 · 2026-04

Karakoram — sharp ridge & valley structure pops at 90 m

Machine-learning susceptibility trained on 80,000+ documented landslides worldwide - terrain, geology, soil, and rainfall.

What ships

Surfaces — calibrated 0–1 probability per pixel.

Every release: 0–1 score · 5-class band · per-pixel uncertainty · input-coverage mask · STAC item · methodology doc.

Inputs

DEM slope/aspect, TWI, landcover, rainfall, NASA-GLC inventory (80K events)

Validation

Spatially blocked hold-out AUC 0.97 — cross-validation blocks prevent leakage across nearby pixels.

Version history

What changed between releases — the same notes that ship with the data. Newest first.

v1.1

2026-04

current

Compact uint16 encoding. The surfaces were re-encoded from float32 to compact integer COGs with a documented scale (physical = raw × scale, nodata 65535) — smaller and self-describing, with no loss of precision at the delivered resolution. Method and values are unchanged from v1.0.

v1.0

2026-04

Initial release. Machine-learning landslide susceptibility (gradient-boosted trees, XGBoost) trained on 80,000+ documented landslides worldwide (NASA Global Landslide inventory) with spatially-blocked cross-validation (hold-out AUC 0.97). Inputs include DEM slope/aspect, topographic wetness, land cover, and rainfall. Ships a five-class susceptibility map plus per-pixel uncertainty and coverage.