AI & IoT Rockfall Prediction — Geological Hazard Detection

Multi-sensor fusion pipeline for real-time geological hazard detection, running inference at the edge on live monitoring data.

A multi-sensor data fusion pipeline with a real-time hazard detection model achieving 98% AUC-ROC, shortlisted for Smart India Hackathon 2025.

Result: 98% AUC-ROC
Recognition: SIH 2025 shortlist
Inference: At the edge

What it had to solve

Rockfall gives warning, but not in any one signal — the useful evidence is spread across several sensors that individually look like noise and only mean something together. And a prediction that arrives after a round trip to a data centre is not a prediction; slopes under monitoring are exactly the places without dependable connectivity, so the decision has to be made where the sensors are.

The decisions that shaped it

Fusion before classification

Several sensor streams are fused into a single view of slope state before anything classifies it. That ordering is the point of the system: individual channels are too noisy to trigger on, and the correlations between them are what actually carry the signal.

A model tuned for the cost of being wrong

Hazard detection is a heavily imbalanced problem where a miss and a false alarm cost wildly different amounts. AUC-ROC is reported because it measures separability across every threshold rather than accuracy at one convenient cut-off; the model reaches 98%.

Inference at the edge

The model runs on edge hardware next to the sensors rather than in a cloud service, so detection survives the connectivity that monitoring sites actually have and the alert path has no network in its critical section.