Year-Wise Refitted Spatiotemporal Landslide Susceptibility Forecasting Driven by IMD Rainfall and SMAP Soil Moisture, Wayanad, India
Abstract
Landslides in hilly regions threaten life, infrastructure, and ecosystems. Landslide susceptibility is inherently transient, yet conventional analysis produces time-invariant zoning and evaluates it by random partitioning of a temporally structured inventory, which permits later events to inform predictions for earlier ones. This study develops a year-wise refitted framework for the Wayanad district of the Western Ghats, India, in which a Random Forest classifier is refitted independently for each target year from 2020 to 2024 using only events recorded earlier, with all target-year events withheld from training. Dynamic susceptibility is driven by India Meteorological Department (IMD) gridded daily rainfall and NASA Soil Moisture Active Passive (SMAP) soil moisture alongside static conditioning factors. Decomposing the evaluation set shows that the conventional random holdout returns a mean AUC-ROC of 0.913, rising monotonically, whereas withheld target-year events return 0.871 with no trend; the target-year value is lower in every year examined (sign test p = 0.031), and the divergence widens as the inventory grows. A Bidirectional Long Short-Term Memory (Bi-LSTM) network forecasts both triggering factors, extending susceptibility mapping to 2025 and 2026, years for which no meteorological record existed when the maps were produced. Two landslides subsequently reported in Wayanad during that period both fall within zones these maps classify as High. A cloudburst scenario is simulated to assess the response to localised extreme rainfall. The framework shifts susceptibility assessment from retrospective hindcasting to forward prediction, and makes explicit the optimism that conventional validation conceals.
Related
- InSAR observations from multiple tracks reveal rainfall modulation of post-failure reactivation of the Xiongwadashan landslide
- Review article: rainfall-induced landslide early warning systems through the lens of the early warnings for all framework – advances, gaps, and perspectives
- 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
BibTeX
@misc{swati2026yearwise,
title = {{Year-Wise Refitted Spatiotemporal Landslide Susceptibility Forecasting Driven by IMD Rainfall and SMAP Soil Moisture, Wayanad, India}},
author = {Swati Swati and Sireesh Saride},
howpublished = {EarthArXiv},
year = {2026},
month = oct,
doi = {10.31223/x5v52n},
url = {https://eartharxiv.org/repository/view/15621/},
note = {CC BY-NC-ND 4.0},
}
RIS
TY - UNPB
TI - Year-Wise Refitted Spatiotemporal Landslide Susceptibility Forecasting Driven by IMD Rainfall and SMAP Soil Moisture, Wayanad, India
AU - Swati Swati
AU - Sireesh Saride
PY - 2026
DA - 2026/10/21
PB - EarthArXiv
DO - 10.31223/x5v52n
UR - https://eartharxiv.org/repository/view/15621/
AB - Landslides in hilly regions threaten life, infrastructure, and ecosystems. Landslide susceptibility is inherently transient, yet conventional analysis produces time-invariant zoning and evaluates it by random partitioning of a temporally structured inventory, which permits later events to inform predictions for earlier ones. This study develops a year-wise refitted framework for the Wayanad district of the Western Ghats, India, in which a Random Forest classifier is refitted independently for each target year from 2020 to 2024 using only events recorded earlier, with all target-year events withheld from training. Dynamic susceptibility is driven by India Meteorological Department (IMD) gridded daily rainfall and NASA Soil Moisture Active Passive (SMAP) soil moisture alongside static conditioning factors. Decomposing the evaluation set shows that the conventional random holdout returns a mean AUC-ROC of 0.913, rising monotonically, whereas withheld target-year events return 0.871 with no trend; the target-year value is lower in every year examined (sign test p = 0.031), and the divergence widens as the inventory grows. A Bidirectional Long Short-Term Memory (Bi-LSTM) network forecasts both triggering factors, extending susceptibility mapping to 2025 and 2026, years for which no meteorological record existed when the maps were produced. Two landslides subsequently reported in Wayanad during that period both fall within zones these maps classify as High. A cloudburst scenario is simulated to assess the response to localised extreme rainfall. The framework shifts susceptibility assessment from retrospective hindcasting to forward prediction, and makes explicit the optimism that conventional validation conceals.
ER -