Post-event information in calendar-year satellite products inflates landslide susceptibility assessments
Abstract
Annual satellite products, such as the AlphaEarth Foundations and TESSERA embeddings, summarize a calendar year, so the product of a landslide's event year also contains post-event observations. We audited this post-event information in 18 dated events (7 earthquakes, 11 rainstorms; 61,402 landslides), changing only the input year within each event and keeping the points, spatial folds, models and bootstrap clusters fixed. Replacing the prior-year AlphaEarth product with the event-year product raised the out-of-fold AUC by 0.087 (95% CI 0.065–0.110; all 18 events) for a point model and by 0.071 (0.048–0.093) for a patch CNN. The point-model gain was 0.085 after subtracting each event's drift between two pre-event years. On complete maps, the event-year input added on average 0.21–0.22 of the landslide area to the map's top 10% (9 of 9 events; up to 0.46). A training-year × prediction-year decomposition traced the inflation to models that learn the event's surface changes, chiefly its landslide scars; event-year inputs used only at prediction did not inflate within-event scores on average. The inflation grew with the post-event fraction of the year, reaching half its full size when 23–39% of the year followed the event; this relation survived corrections for recovery. TESSERA and plain Sentinel-2 composites behaved alike. In forward tests in three regions, models trained on earlier landslides scored 0.02–0.11 higher AUC on the new event with event-year inputs, whereas legitimate pre-event embeddings helped mainly near earlier landslides. Susceptibility zoning should use only products whose year ends before the event.
Related
- Landslide Susceptibility Modelling and Infrastructure Exposure in the Zillertal, Austrian Alps
- Machine-learning modelling of regional landslide susceptibility in eThekwini, South Africa
- Landslide Susceptibility Modeling and Mapping: A Comparison of Frequency Ratio and Decision Tree Models
- Role of Topo-Climatic Factors in Landslide Susceptibility Mapping: A Machine Learning Perspective
- Adaptive Inversion Optimization of the Static Factor of Safety for Newmark-Based Earthquake-Induced Landslide Susceptibility Assessment
BibTeX
@misc{du2026postevent,
title = {{Post-event information in calendar-year satellite products inflates landslide susceptibility assessments}},
author = {Yifan Du and Tianhe Zheng and Peiyan Yang and Yibo Li and Junhang Zhou and Kun Yang},
howpublished = {EarthArXiv},
year = {2026},
month = oct,
doi = {10.31223/x5rb93},
url = {https://eartharxiv.org/repository/view/15601/},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Post-event information in calendar-year satellite products inflates landslide susceptibility assessments
AU - Yifan Du
AU - Tianhe Zheng
AU - Peiyan Yang
AU - Yibo Li
AU - Junhang Zhou
AU - Kun Yang
PY - 2026
DA - 2026/10/20
PB - EarthArXiv
DO - 10.31223/x5rb93
UR - https://eartharxiv.org/repository/view/15601/
AB - Annual satellite products, such as the AlphaEarth Foundations and TESSERA embeddings, summarize a calendar year, so the product of a landslide's event year also contains post-event observations. We audited this post-event information in 18 dated events (7 earthquakes, 11 rainstorms; 61,402 landslides), changing only the input year within each event and keeping the points, spatial folds, models and bootstrap clusters fixed. Replacing the prior-year AlphaEarth product with the event-year product raised the out-of-fold AUC by 0.087 (95% CI 0.065–0.110; all 18 events) for a point model and by 0.071 (0.048–0.093) for a patch CNN. The point-model gain was 0.085 after subtracting each event's drift between two pre-event years. On complete maps, the event-year input added on average 0.21–0.22 of the landslide area to the map's top 10% (9 of 9 events; up to 0.46). A training-year × prediction-year decomposition traced the inflation to models that learn the event's surface changes, chiefly its landslide scars; event-year inputs used only at prediction did not inflate within-event scores on average. The inflation grew with the post-event fraction of the year, reaching half its full size when 23–39% of the year followed the event; this relation survived corrections for recovery. TESSERA and plain Sentinel-2 composites behaved alike. In forward tests in three regions, models trained on earlier landslides scored 0.02–0.11 higher AUC on the new event with event-year inputs, whereas legitimate pre-event embeddings helped mainly near earlier landslides. Susceptibility zoning should use only products whose year ends before the event.
ER -