SIH25071 — AI-Based Rockfall Prediction & Alert System

An end-to-end geotechnical surveillance and early-warning platform designed for open-pit mines. The system fuses multi-modal environmental, spatial, and sensor telemetry into physics-grounded machine learning models to forecast slope instability and trigger sub-minute evacuation alerts.

Live: Backend · Frontend · API Docs

About The Project

SIH25071 is an AI/ML-powered geotechnical early-warning and slope stability surveillance platform tailored for open-pit / opencast mining operations (aligned with the Ministry of Mines, Disaster Management theme).

In opencast mines (such as SECL Kusmunda, Korba Coalfield, Chhattisgarh), slope failure and bench rockfalls represent critical occupational hazards. While industrial Slope Stability Radar (SSR) systems deliver sub-millimeter displacement tracking, their high capital expenditure (~$250k–$500k/unit) and line-of-sight constraints leave peripheral and smaller-scale pits unmonitored.

This platform bridges that gap by fusing distributed geotechnical sensor telemetry, satellite Earth observation, and meteorological data into a unified, physics-grounded machine learning pipeline:

• Multi-Modal Remote Sensing & Meteorology: Copernicus GLO-30 DEM for slope/aspect/curvature via Google Earth Engine; Sentinel-1 SAR (C-band GRD) backscatter change detection as a surface-disturbance proxy; Open-Meteo ERA5 precipitation data.
• Physics-Informed Geotechnical Telemetry: Displacement, pore pressure, vibration, and strain calibrated under the Fukuzono (1985) Inverse Velocity Method — displacement accelerates and inverse velocity trends to zero before failure.
• Imbalance-Aware ML Engine: Class-weighted loss (not SMOTE — physically-correlated channels risk implausible synthetic interpolation). Models: RandomForest, XGBoost (production champion — 100% evacuation recall), GRU (benchmark). Exportable to ONNX Runtime for offline edge alerting.
• Real-Time 3D Dashboard: Next.js 16 + MapLibre GL pit heatmaps, WebSocket telemetry charts, and evacuation dispatch logs.

---

Problem & Physical Grounding

• Problem Statement: SIH25071 | Ministry of Mines (Disaster Management Theme)
• Goal: Mitigate fatal slope failures in opencast mines where SSR coverage is unavailable.
• Physical Basis:
  • Inverse Velocity Method (Fukuzono, 1985): Displacement rate accelerates before failure; inverse velocity trends to zero — enables lead-time forecasting.
  • Empirical Risk Thresholds (Indonesian open-pit coal SSR case study):
    • Safe: 0–50 mm/day
    • Warning: 50–120 mm/day
    • Evacuation: >120 mm/day

---

System Architecture

[Geotechnical Sensors + Sentinel-1 SAR + GLO-30 DEM + Open-Meteo API]
                               │
                               ▼
                  [FastAPI Stream & Ingestion]
                               │
               ┌───────────────┴───────────────┐
               ▼                               ▼
     [ML Inference Engine]           [Edge Node (ONNX)]
 (RF/XGBoost + GRU TimeSeries)    (Local Siren / Offline Mode)
               │                               │
               └───────────────┬───────────────┘
                               ▼
                   [Real-Time WebSocket Feed]
                               ▼
           [Next.js 16 Dashboard (MapLibre + Recharts)]

---

Model Performance (Test Set — Evacuation Class)

Metric	RandomForest (v2)	XGBoost (v2)	GRU
Precision	0.9949	0.9704	1.0000
Recall	0.9848	1.0000	0.7208
F1-Score	0.9898	0.9850	0.8378
Missed Evacuations	3 / 197	0 / 197	55 / 197

XGBoost is the production champion (zero missed evacuations). RF ships on the live backend for its stronger terrain/SAR SHAP signal (17.03% vs 6.90%). GRU is benchmarked for architectural completeness — all 55 misses land in Warning, not Safe.

---

Tech Stack

Layer	Technologies
Frontend	Next.js 16.3 (App Router, Turbopack), React 19, TypeScript 5.9, Tailwind CSS 4
Mapping & Viz	MapLibre GL + React-Map-GL (open-source 3D terrain), Recharts
Backend API	FastAPI 0.141, Python 3.12, Uvicorn, WebSockets
ML & Inference	Scikit-learn, XGBoost, PyTorch (GRU), ONNX Runtime, SHAP
Deployment	Vercel (Frontend) + Render (Backend)
Containerisation	Docker + Docker Compose · Images on Docker Hub (kaizer777)

---

Repository Structure

SIH2026/
├── frontend/             # Next.js 16 App Router UI
│   ├── app/              # Routes: /dashboard, /alerts, /trends, /pitch
│   ├── components/       # MapLibre 3D heatmap, Recharts trends, TopBar
│   └── lib/              # API client, WebSocket client, TypeScript types
│
├── backend/              # FastAPI microservice
│   ├── main.py           # Entrypoint, lifespan, CORS, router mount
│   ├── app/schemas.py    # Pydantic: SensorReading, RiskPrediction, AlertEvent
│   ├── app/physics_generator.py  # Fukuzono-based live sensor generator
│   └── routers/rockfall.py       # POST /predict, WS /ws/feed, alert logic
│
├── models/               # Trained artifacts (RF, XGBoost, GRU + metadata)
├── data/                 # DEM, SAR, rainfall, synthetic sensors, sequences
├── scripts/              # Phase scripts (terrain → training → integration)
├── tests/                # Pytest: endpoint, WebSocket, alert dedup, physics
├── reports/              # SHAP plots, confusion matrices
├── docs/                 # CONTEXT.md, WORKFLOW.md, session logs, pitch drafts
└── frontend.md           # Frontend design guide (single source of truth)

---

Quick Start

🐳 Docker (recommended — zero setup)

# 1. Copy and fill in env vars (add your GROQ_API_KEY)
cp .env.docker.example .env.docker

# 2. Build and start both services
docker compose up --build

Service	URL
Frontend	http://localhost:3000
Backend API	http://localhost:8001
API Docs	http://localhost:8001/docs

Or pull pre-built images directly:

docker pull kaizer777/sih2026-backend:latest
docker pull kaizer777/sih2026-frontend:latest

Note: models/ and data/ are mounted from the repo root at runtime — clone the full repo before running.

---

Manual Setup

1. Backend (FastAPI)

cd backend

# Create & activate Python 3.12 venv
py -3.12 -m venv venv
.\venv\Scripts\Activate.ps1   # Windows PowerShell
# source venv/bin/activate     # Linux/macOS

# Install dependencies
pip install -r requirements.txt

# Run development server
uvicorn main:app --reload --port 8000

API Docs at http://localhost:8000/docs

2. Frontend (Next.js)

cd frontend

# Install packages
npm install

# Start development server
npm run dev

Dashboard at http://localhost:3000

---

Documentation

Doc	Description
docs/CONTEXT.md	Full engineering spec, scientific references, all 29 phases, API reference, glossary
docs/WORKFLOW.md	Day 0 → Demo execution plan
frontend.md	Frontend design guide (colors, typography, components, responsiveness)
AGENTS.md	AI agent operational directives