METHANEGUARD AI '26
MethaneGuard AI
Every methane leak. Every second. Everywhere.
AI + Computer Vision platform for methane hotspot detection,�risk prioritization, and inspection operations.
Live · methengaurdai-2.onrender.com · GitHub · Dynamicsoumya/MethenGaurdAi
MethaneGuard AI '26 · Every methane leak. Every second. Everywhere.
1/10
The problem
Invisible leaks. Visible climate & financial damage.
Methane is invisible & odorless — ~80× CO₂ over 20 years (GWP20)
Leaks persist for weeks across pipelines, tanks, compressors, refineries
Satellites / drones / sensors already produce oceans of data
What's missing: intelligence to read it — not more inspection schedules
Cost of delay: emissions + regulatory fines + lost product + wasted site visits
MethaneGuard AI '26 · Every methane leak. Every second. Everywhere.
2/10
What the brief asks → what we built
End-to-end methane intelligence OS
BRIEF REQUIREMENT
METHANEGUARD
Detect hotspots (sat / drone / sensor)
Reject background / false positives
Location + severity + confidence
AI risk → inspection priority
Actionable monitoring dashboard
MethaneUNet + /detect modalities
Cls head + TP/FP confusion UI
Geo, kg/h, severity, risk score
GBM + Command queue + /ceo
3D Earth · roles · twin · reports
MethaneGuard AI '26 · Every methane leak. Every second. Everywhere.
3/10
Solution loop
From pixel to repair crew
Detect
Segment
Score
Forecast
Dispatch
Repair
Report
Not a notebook demo — an operations platform with Admin / Analyst / Inspector roles,�Command Center, digital twin, autonomous agents, and executive reporting.
MethaneGuard AI '26 · Every methane leak. Every second. Everywhere.
4/10
Architecture
CV + risk + ops in one stack
CV: MethaneUNet (seg + classification head)
Risk: rules + GradientBoosting priority 0–100
Forecast: 1h / 6h / 24h plume dispersion
XAI: why flagged + key factors + confusion
API: FastAPI · UI: Next.js + Three.js Earth
Deploy: Render · Docker-ready
DATA PATH
AVIRIS-like synthetic → primary metrics
STARCOP / AVIRIS-NG → real fine-tune
Sentinel-5P TROPOMI .nc → hotspots UI
Honest status matrix on /datasets
MethaneGuard AI '26 · Every methane leak. Every second. Everywhere.
5/10
Results
Validated metrics — honest dataset story
DICE
0.933
Synthetic primary
IoU
0.879
Synthetic primary
mAP50
0.833
Proxy · primary
RISK AUC
~0.96
GBM priority model
REAL AVIRIS-NG FINE-TUNE (STARCOP mini · secondary)
Real chips: 14 train / 4 val → Dice 0.826 · IoU 0.704 · mAP50 0.698
False-positive control: classification head + val confusion (TP/FP/TN/FN) on /detect & /datasets
Primary published scores remain synthetic; real fine-tune is secondary & disclosed
MethaneGuard AI '26 · Every methane leak. Every second. Everywhere.
6/10
Product walkthrough
What judges should click
1. Landing — cinematic 3D Earth, plumes, satellites, wind
2. /login → Demo users — Admin / Analyst / Inspector (one click)
3. /command — leak queue by risk, workflow, XAI, repair plan
4. /detect — run AI detection + Dice/IoU + FP confusion
5. /datasets — AVIRIS synthetic + real + TROPOMI honesty matrix
6. /ceo · /twin · /inspector — executive, twin, field flow
MethaneGuard AI '26 · Every methane leak. Every second. Everywhere.
7/10
Stack & links
Ship-ready hackathon build
FRONTEND
Next.js 16
React 19
Three.js / R3F
Tailwind · Zustand
BACKEND / AI
FastAPI
PyTorch MethaneUNet
scikit-learn GBM
NetCDF TROPOMI
LIVE
App: methengaurdai-2.onrender.com
API: methengaurdai.onrender.com
Repo: MethenGaurdAi
Docs PDF in /docs
MethaneGuard AI '26 · Every methane leak. Every second. Everywhere.
8/10
Impact
Faster detection. Fewer wasted inspections.
Minutes to triage super-emitters instead of weeks of blind schedules
Risk-ranked inspection queue cuts unnecessary site visits
CH₄ → CO₂e, trees, cars, ₹/$ impact for compliance storytelling
Role-based ops: HQ admin, analyst verification, field inspector close-out
Audit trail: reports, workflow states, explainable AI factors
MethaneGuard AI '26 · Every methane leak. Every second. Everywhere.
9/10
NEXT & ASK
MethaneGuard AI
Smell trouble before the planet does.
Next: larger real AVIRIS splits · live TROPOMI feeds · field OGI companion
Try it: methengaurdai-2.onrender.com · Demo users on /login
MethaneGuard AI '26 · Every methane leak. Every second. Everywhere.
10/10