Chapter 01 · The Problem
Kullu · Mandi, Himachal PradeshOn a mountain slope,
rain is never just rain.
In the rugged valleys of Kullu and Mandi, intense monsoons trigger devastating slope failures every season. Yet rainfall alone does not determine where the earth collapses. Steepness, fractured Himalayan bedrock, soil density, and road incisions dictate which slopes hold firm and which give way. GeoShield was engineered to separate intrinsic terrain susceptibility from dynamic weather triggers across a continuous 500 m grid.
Why Landslides Happen:
The Physics of Slope Failure
Slope failures in high-relief mountainous topography can involve complex combinations of debris flows, planar slides, and rotational slumps. While rainfall triggers individual events, geological steepness, soil mechanical composition, groundwater pathways, and anthropogenic modifications (such as road benching) govern which slopes are susceptible to collapse.
Under dry or drained conditions, gravitational driving forces on the slope face are countered by the internal friction and cohesion of the soil and rock matrix (Coulomb shear strength).
Assumes continuously monitored pore-water sensors, borehole core lithology, and micro-catchment drainage modeling across individual slope faces. Real-time factor-of-safety calculations require extensive physical field instrumentation that cannot scale across entire mountain ranges.
Because continuous subsurface sensor networks do not exist across Kullu and Mandi, GeoShield decouples the problem: an XGBoost ML model evaluates static topographic, geologic, and spatial susceptibility, while empirical multi-day antecedent rainfall thresholds assess the meteorological trigger state.
Understanding the Landscape:
Spatial Evidence & The 12 Features
GeoShield transforms satellite topographies, global soil databases, open infrastructure vectors, and geological inventories into a unified 500 m regular grid spanning 257 columns by 374 rows across Kullu and Mandi districts. Every collected dataset fulfills a distinct role in the system architecture.
Used by the Trained Susceptibility Model (12 Features)
The trained model uses exactly these 12 static geo-environmental features. Historical landslide density is by far the single largest contributor to the model's susceptibility output. Population is included as a spatial layer summed over each cell's full area, but is not a major driver.
Vertical relief and alpine geomorphology
Gravitational shear gradient
Compaction and subterranean pore volume
Cohesive shear strength and plasticity
Internal friction and percolation rate
Erodibility and piping vulnerability
Chemical weathering index
Topsoil aggregate stability and moisture retention
Toe cut slope steepening and drainage alteration
Spatial anthropogenic presence density
Spatial clustering of past slope failures
Valley toe erosion and basal slope scour
Used Only by the Rainfall-Trigger Layer (Not a Model Feature)
Historical weather data at 21 locations across the region (2015-01-01 to 2026-09-12, sourced from the Open-Meteo Historical Weather API). Each cell is linked to its nearest precipitation location (up to 34.7 km away, mean 15.2 km).
Rainfall feeds the deterministic empirical trigger rules directly. Rainfall is NOT an input to the XGBoost machine learning model.
Collected But Not Used in Training
Vegetation and water indices (NDVI and NDWI) derived from a single Sentinel-2 optical scene (2026-03-09). While this scene captures approximately 70% of historical landslide points, it covers only about 32% of all 96,118 grid cells.
Including partially missing spectral features would have biased spatial predictions or forced artificial imputation over alpine glaciers and valleys. Consequently, NDVI and NDWI are retained in project records but strictly excluded from model training.
The Temporal Blindspot:
A Critical Methodological Pivot
Machine learning projects often begin with an ambitious premise: feed daily rainstorms and past landslide timestamps into a neural network or gradient booster to predict exact days of failure. When GeoShield examined the official Geological Survey of India (GSI) Bhusanket inventory, we uncovered a fundamental data limitation.
We rejected the temptation to invent synthetic event dates or train a temporal black-box model on an unverified 2015-concentrated catalog.
Decouple spatial susceptibility from meteorological stress. Train XGBoost strictly on where terrain is prone, and separately evaluate when rainfall conditions exceed empirical Himalayan thresholds.
The Two-Layer Solution:
Decoupling Terrain Predisposition from Meteorological Triggers
By separating intrinsic geo-environmental predisposition from transient meteorological stress, GeoShield provides operational clarity: each layer answers an entirely distinct physical question with explicit empirical boundaries.
Susceptibility Score
An XGBoost model trained on 12 static topographic, geologic, and infrastructure features. Calculates a static relative terrain index from 0 to 100 per cell that never changes with rainfall.
The training sample was enriched to ~20% positives, whereas real-world prevalence across the grid is only ~0.9% (872 positive cells out of 96,118). Raw model scores reflect this synthetic balance. Furthermore, calibration audits show observed event frequencies exceed predicted scores in low-to-mid bands while over-predicting the highest band. It is strictly a relative index.
Rainfall Trigger State
Dynamic meteorological conditions evaluated against four empirical checks from regional Himalayan literature. Yields a categorical trigger state: LOW, ELEVATED, or HIGH.
Derived from Singh et al. (2025, Shimla, Himachal Pradesh; DOI 10.1016/j.nhres.2024.12.004). This is a same-state empirical reference, NOT calibrated specifically for Kullu or Mandi. Furthermore, the reference paper calibrated duration for discrete storm events; GeoShield applies rolling 1/3/7-day windows because its weather data is daily.
Neither layer alone assigns risk. In the next chapter, see how these independent axes combine to establish operational priority.
How the Layers Become Current Risk:
The 3×3 Decision Matrix
When operational assessments are made, GeoShield does not obscure decisions inside a black box. Instead, the static Susceptibility Bin and dynamic Trigger State intersect deterministically in a 3×3 lookup matrix, producing one of six operational Current Risk levels.
Grid Cells in the Study Box · Snapshot 2026-09-12
Total: 96,118 cells (the box includes neighboring terrain)Interactive What-If Simulation:
Stress-Testing the Slopes Under Extreme Rainfall
Emergency planners cannot wait for catastrophic monsoons to discover which valley sectors will cross trigger thresholds. GeoShield includes a real-time sensitivity simulation engine that stress-tests terrain by scaling antecedent rainfall stress.
Slope, soil, and bedrock morphology remain fixed.
Observed 1d, 3d, 7d, and 30d antecedent precipitation snapshot.
Controlled rainfall increments (+10%, +20%, +30%).
New 3×3 matrix intersection outcome.
Cell Outcome Assessment
“Susceptibility Score does not change; only the rainfall trigger does.”
Launch GeoShield directly with the interactive what-if simulation tour enabled:
What GeoShield Can and Cannot Tell You
In natural hazard engineering, true trustworthiness comes from radical transparency regarding what a model knows and what it cannot know. The assumptions and physical boundaries of GeoShield are documented below.
The score is a 0–100 relative spatial index ranking terrain vulnerability against known historical inventory patterns. It is not an occurrence probability, percentage chance, or statistical odds of failure.
Precipitation inputs represent a static historical weather snapshot as of 2026-09-12. GeoShield is not connected to a live radar feed and does not present real-time streaming rainfall.
The simulation slider tests slope sensitivity under hypothetical precipitation deltas (+10%, +20%, +30%). It is an analytical diagnostic tool, not a meteorological forecast of incoming weather systems.
Trigger criteria originate from Singh et al. (2025, Shimla, HP). While geographically within Himachal Pradesh, the thresholds are not individually calibrated for Kullu or Mandi micro-catchments. Rolling 1, 3, and 7-day windows represent an operationalization of daily records.
The strong influence of road proximity reflects both genuine toe-cutting destabilization and human reporting bias—landslides along highways are documented far more frequently than slides on remote high-altitude peaks. Unmapped terrain may possess lower baseline inventory fidelity.
Rainfall (21 points) and SoilGrids (19 points) are assigned to the nearest grid cell (up to 34.7 km and 48.5 km away). The 128 km × 187 km grid boundary is a rectangular box encompassing both districts and overshoots into neighboring terrain.
Explore GeoShield
Navigate 96,118 spatial cells across Kullu and Mandi. Inspect local susceptibility scores, examine regional rainfall triggers, and test what-if scenarios in the live interactive map.
Open the Operational Map- •NASADEM 30 m via OpenTopography
- •GSI Bhusanket landslide inventory
- •Open-Meteo Historical Weather API
- •SoilGrids (ISRIC)
- •© OpenStreetMap contributors
- •WorldPop (100 m resolution, 2020)
- •Copernicus Sentinel-2 (Scene 2026-03-09)
- •Singh et al. (2025), DOI 10.1016/j.nhres.2024.12.004