30% of existing models fail to identify high-risk zones. 19% of at-risk populations remain unprotected. After Wayanad, here is why the paradigm must shift.
The Human Cost of ‘Known Unknowns’
July 2024. Wayanad, Kerala. A region known for steep terrain, lateritic soils, and intense monsoon rainfall — all well-documented risk factors for landslides. The general vulnerability was mapped. The hazard was understood. Yet when the slopes finally gave way, over 200 lives were lost, more than 100 people were injured or went missing, and over 2,000 homes were destroyed.
The failure was not one of ignorance — it was one of precision. We knew the region was dangerous. We did not know exactly where or exactly when the terrain would fail. Wayanad was, in the starkest possible terms, a predictable tragedy with unpredictable timing.
This gap — between knowing a region is vulnerable and being able to predict specific slope failures — is the central challenge of landslide risk management. And it is a gap that current models are failing to close.
Three Ways Current Models Leave Us Blind
Across India’s 147 landslide-vulnerable districts, disaster management authorities rely on susceptibility maps and prediction models that suffer from three fundamental blind spots:
Blind Spot 1: Inaccurate Prediction
Approximately 30% of current landslide susceptibility models fail to correctly identify high-risk zones. This is not a minor calibration issue — it means that nearly one-third of the time, the areas flagged as dangerous may not be the ones that actually fail, while genuinely critical slopes receive insufficient attention. The root cause lies in how these models are built: most rely on pixel-level spectral analysis that treats each point in isolation, ignoring the spatial relationships between terrain features that actually govern slope stability.
Blind Spot 2: False Alarms
When models over-predict risk, they generate false alarms — warnings that cry wolf. This erodes institutional trust in the warning system itself. When disaster management authorities receive repeated alerts for areas that do not fail, the natural human response is to discount future warnings. The consequence is a dangerous feedback loop: imprecise models generate unreliable alerts, authorities lose confidence in the system, and when a genuine threat materializes, the response is slower and less decisive than it needs to be.
Blind Spot 3: False Sense of Safety
Perhaps the most dangerous failure mode is the false negative — when a genuinely hazardous area is classified as safe. An estimated 19% of at-risk populations remain unprotected because current identification methods miss the zones they inhabit. These are the slopes that appear stable in a static susceptibility map but are quietly losing strength under the cumulative pressure of rainfall infiltration, soil saturation, and seasonal terrain degradation. When these slopes fail, communities have no warning because the model never flagged their location.
The downstream effect of these three blind spots is devastating. Around 15% of disaster mitigation budgets are misallocated because of blurred risk maps — resources directed to the wrong areas, protective infrastructure built in zones of lower actual risk while genuinely vulnerable communities go unshielded.
Why Conventional Approaches Fall Short
To understand why current models underperform, we need to examine their foundational assumptions. Most existing landslide susceptibility models share three technical limitations:
Limitation 1: Pixel-Based Analysis Ignores Spatial Context
The dominant approach in satellite-based landslide mapping treats each pixel independently — analyzing spectral reflectance, elevation, and slope at individual points without considering how those points relate to their neighbours. But landslides are inherently spatial phenomena. A slope’s stability depends not just on its own gradient, but on the drainage patterns above it, the soil conditions around it, and the vegetation network anchoring it. Pixel-based analysis captures none of this context.
Object-Based Image Analysis (OBIA) offers a fundamentally different approach. Instead of analyzing individual pixels, OBIA segments the landscape into meaningful units — terrain objects that correspond to real-world features like ridgelines, valleys, and slope facets. These objects preserve spatial relationships, enabling the model to understand landscape patterns rather than isolated data points. When applied to landslide susceptibility, OBIA delivers markedly superior performance in identifying hazard boundaries and distinguishing genuinely critical zones from background noise.
Limitation 2: Static Models in a Dynamic World
Most susceptibility maps are produced once and treated as static references — a snapshot of risk frozen in time. But slope stability is profoundly dynamic. Rainfall infiltration changes pore water pressure hour by hour during a storm. Soil cohesion degrades with sustained moisture exposure. Vegetation cover shifts seasonally, altering root reinforcement patterns. A model that does not account for these temporal dynamics is answering the wrong question: it tells you where slopes are structurally prone to failure, but not when they are actually approaching the failure threshold.
The distinction matters enormously for emergency response. A static map might identify a thousand slopes as “high risk.” But the operationally relevant question is: which of those thousand slopes is approaching critical instability right now, given the rainfall that fell in the last 72 hours?
Limitation 3: Data Leakage Inflates Confidence
A pervasive technical flaw in many published landslide models is spatial data leakage — the contamination of training and testing datasets with geographically proximate samples. When training and validation data come from the same spatial neighbourhood, models learn location-specific patterns rather than transferable geological principles. The result is artificially inflated accuracy metrics that collapse when the model is deployed in a new region.
This is not an academic concern. It directly explains why models that report impressive accuracy in research papers fail catastrophically in operational deployment. A model trained and validated on data from the same watershed may report 95% accuracy but perform at 60% when applied to the next district. Spatially-aware cross-validation — which enforces geographic separation between training and testing data — is essential for producing models that genuinely generalize.
The Paradigm Shift: Physics-Informed Intelligence
Addressing these limitations requires more than incremental improvement — it demands a fundamental shift in how we model terrain stability. The answer lies in physics-informed AI: models that embed physical laws governing slope mechanics directly into the machine learning framework.
Understanding Invisible Pressure:
When rain falls on a hillslope, it does not simply run off. Water infiltrates the unsaturated soil, raising pore water pressure within the slope mass. This increasing pressure reduces the soil’s effective strength — its ability to resist the gravitational forces pulling it downhill. The process is invisible from the surface, and it is precisely what conventional spectral-only AI models cannot see.
The Transient Rainfall Infiltration and Grid-based Regional Slope-stability (TRIGRS) modeling approach simulates this invisible process. It calculates how rainfall-driven infiltration evolves pore pressures over time, accounting for the specific hydraulic properties of different soil and rock types. The output is not a vague susceptibility score but a time-varying Factor of Safety (FoS) for each slope element — a physically meaningful metric that directly indicates proximity to failure.
Dynamic Material Modeling: The Rain Factor
Different geological materials respond to saturation in fundamentally different ways. Alluvium with a base cohesion of 3,000 Pa and friction angle of 30° loses up to 30% of its strength at full saturation, dropping to an adjusted cohesion of 2,100 Pa and friction angle of 24°. Phyllite, starting at 12,000 Pa cohesion and 24° friction, degrades to 8,400 Pa and 19.2° under the same conditions.
Physics-informed models capture these material-specific responses by integrating real-time precipitation data to dynamically adjust geotechnical parameters. The transformation formulas are straightforward but powerful: cohesion is reduced proportionally to saturation level, friction angles decrease by up to 20% at saturation, and unit weight increases with water content as the slope literally becomes heavier. This is not abstract modeling — it is the direct simulation of the physical processes that precede every rainfall-triggered landslide.
The Factor of Safety: Your Warning Signal
Every slope calculation produces a Factor of Safety value that encapsulates the complete stability story. FoS is the ratio of resisting forces (the soil’s strength) to driving forces (the gravitational and water-pressure stresses acting on the slope). When FoS is above 1.5, the slope is highly stable. Between 1.0 and 1.5, it enters a zone of moderate risk requiring monitoring. Below 1.0, failure is imminent.
This clear, physics-based metric transforms complex geotechnical analysis into actionable intelligence. A disaster management authority does not need to interpret spectral indices or neural network confidence scores — they need to know which slopes are approaching FoS = 1.0 and how fast they are getting there. Physics-informed models provide exactly this.
An End-to-End Intelligence Pipeline
At Jarbits, we have integrated these principles into TerraLux — an end-to-end geospatial intelligence pipeline that flows from raw satellite data to actionable early warnings.
The Glass Box Approach: AI You Can Trust
There is one final dimension that separates physics-informed AI from conventional approaches: explainability. Government decision-makers are understandably cautious about acting on “black box” AI predictions. When an algorithm says “this area is high risk” without explaining why, it creates a trust deficit that slows response times and undermines adoption.
TerraLux addresses this through what we call the Glass Box approach. Using LIME (Local Interpretable Model-agnostic Explanations) and SHAP (SHapley Additive exPlanations) frameworks, every risk alert is accompanied by a transparent breakdown of contributing factors. A typical alert might show: Slope Angle is critical at above 45 degrees with high contribution, Soil Saturation is at 92% with high contribution, and Vegetation Cover is low with moderate contribution. Decision-makers see not just the risk score but the specific physical conditions driving it — enabling them to validate the alert against local knowledge and respond with confidence.
From ‘If’ to ‘When’
The title of this post reflects the fundamental transition that physics-informed AI enables. Current models answer the question “if” a slope might fail — a static, probabilistic assessment of vulnerability. Physics-informed models answer the far more urgent question of “when” — tracking the dynamic evolution of slope stability under real-time weather conditions and identifying the approach to failure thresholds.
After Wayanad, after every monsoon season that claims lives in our hills, the question is no longer whether we can build better models. The technology exists. The data exists. The scientific principles are well understood. The question is whether we will deploy these capabilities at the scale and speed that the crisis demands.
At Jarbits, that is exactly what we are building TerraLux to do. Recognized by MeitY as a Top 30 AI Solution, incubated at SIIC IIT Kanpur, and now part of the LowCarbon.Earth 2026 climate-tech cohort, we are committed to ensuring that the next monsoon season finds India better prepared — with models that see what current approaches cannot, and warnings that arrive before the ground gives way.
About TerraLux: TerraLux is Jarbits’ flagship geospatial AI platform for landslide susceptibility mapping and early warning systems. It combines physics-informed feature engineering, object-based image analysis, spatially-aware cross-validation, and explainable AI to deliver actionable terrain intelligence. Purpose-built for Indian terrain conditions, transferable globally.Want to learn more? Reach out at contact@jarbits.com
