Blockage severity82/100
Visible obstruction may retain and mobilize street waste during rainfall.
40% weight · 32.8 points Environmental decision-support prototype
Street-to-waterway monitoringDecision support
Which drain should your crew inspect before the storm?
DrainGuard turns visible street evidence and changing conditions into explainable inspection priorities—then verifies whether cleanup worked.
DETECT→PRIORITIZE→ALLOCATE→ACT→VERIFY
1crew
2inspections available
3reports competing
A decision has to be made.
Why intervene before rainfall?Potential pathway
- 01Street litter
- 02Blocked storm drain
- 03Rainfall mobilizes waste
- 04Potential transport toward waterways
- 05Environmental concern
DrainGuard intervenes through Detect → Prioritize → Priority Shock → Allocate → Verify → Close.
01 · Inspect
Photo in. Priority out.
The model runs on the image, then a transparent risk formula combines obstruction, litter, and rainfall. No black-box flood claims.
mixed litter · 91%
organic debris · 86%
5th Cross · KoramangalaDemo scenarioChoose a photo — analysis starts automatically.
Evidence gateStrong visual evidence
Detected objects2 boxes
Model pathDrain-domain evidence gate
Why this priority?
Every point has a visible reason.
Cleanup priority63/100High
Rainfall exposure28/100
Uses 18.0 mm as the current or selected scenario input.
30% weight · 8.4 pointsVisible litter62/100
Visible waste increases potential environmental concern; it does not measure pollution volume.
20% weight · 12.4 pointsEnvironmental context100/100
High proximity concern: mapped water feature about 180 m away.
10% weight · 10.0 pointsEnvironmental impact risk64/100 · High
Evidence coverage100% · high confidence
Recommended action: Schedule inspection and cleanup within 24 hours.
Environmental decision-support estimate. This helps prioritize inspection using visible blockage, litter evidence, rainfall conditions, and mapped environmental context. It is not a hydrological model or a prediction of pollution volume.
⚡ Priority Shock
🌧 Moderate rainSame drains. Different conditions. Different priorities.
Change the rainfall scenario and watch the existing scoring logic update priority, environmental concern, and the ranked queue. This is scenario exploration for decision support—not a weather forecast or flood prediction.
Rainfall input18 mm / 24h
0 mm · Dry64+ mm · Heavy
Current condition🌧 Moderate rain18 mm selected
Urgent inspection0reports at 80+ priority
Priority movementReadymove rainfall to compare
#1DG-1045th Cross · Koramangala63/100—
Severe visible obstruction remains the strongest priority contributor.
#2DG-098Market Road · Shantinagar55/100—
Visible evidence remains lower than the reports above it.
#3DG-0911st Main · Indiranagar46/100—
Visible evidence remains lower than the reports above it.
#4DG-0878th Block · Jayanagar27/100—
Cleanup already verified; this report remains visible for routine monitoring.
Rainfall scenario explorer
One drain. Three operating conditions.
Scenario exploration for decision support. These are controlled inputs, not a weather forecast or a prediction of flooding or pollution volume.
Live impact story
What changed because we prioritized these drains?
Calculated only from reports added on this device. Demo scenarios are excluded.
No live reports yet.
Add a location after inspecting a photo, or select Demo preview to see how the dashboard responds.
Controlled scenarios
Sample data for demonstrationDemo mode
Select a controlled scenario to update the inspection, scoring, explanation, and map workflow. These examples are not real municipal observations.
02 · Prioritize
What should crews inspect next?
Crews see cleanup priority, environmental concern, evidence status, and the recommended action—not simply the newest report.
Loading saved reports
Reports, scores, and compressed before/after evidence survive refreshes on this inspection device. Shared multi-user storage requires a connected municipal database.
5th Cross · Koramangala4 mapped reports · symbols + labels
▲ High concern ● Moderate ✓ Verified ! Review
Cleanup queueSorted by cleanup priority
SelectedDG-104
Dispatch now. Cleanup priority 84/100 · environmental concern 87/100.
Human review
1 waitingEvidence that needs a person.
DG-0911st Main · IndiranagarOpen
Low-confidence or non-drain evidence requires a field officer to check the inlet.
🚚 Action Planner
2 of 2 reports plannedIf you only have one crew, where should it go?
Transparent capacity allocation—not route optimization. DrainGuard selects the highest-priority dispatchable reports using the current scenario and distributes them across available crews. Human-review cases stay with the review queue.
Today's capacity2 / 2top-ranked open reports
Recommended nowWithin capacity
Crew 1
1DG-1045th Cross · Koramangala63 priority
2DG-098Market Road · Shantinagar55 priority
Monitor / next wave0 outside capacity
All open reports fit within today's capacity.
Controlled demo timeline
09:00 Report received
→09:01 Evidence detected
→09:02 Rainfall scenario updated
→09:03 Added to crew plan
→✓ Verified clear after cleanup evidence
Why DrainGuard?
DrainGuard closes the decision loop.Most systems stop at reporting.
Basic / static reporting workflowDrainGuard workflow
Someone reports a problemAI evaluates visible evidence
Reports remain in a queueConditions influence priority
Static task orderingPriorities adapt
Teams choose manuallyLimited capacity gets an action plan
Task marked completeCleanup evidence is verified
Detection is not the finish line. Verified action is.
03 · VerifyClose the loop,
Close the loop,
not just the ticket.
After the crew cleans this drain, upload a second photo. The AI compares obstruction and litter with the original evidence before closing the report.
Detection alone does not prove the problem was solved. DrainGuard verifies the cleanup evidence.
1 Upload after photo2 AI compares evidence3 Map status updates
Same scene confidenceWaiting for an after photo.
Drain evidenceWaiting for an after photo.
Obstruction improvedRequires at least a 15-point reduction.
Litter thresholdResidual litter must be 48 or lower.
Before · 82% blocked+
After · waitingAfter photo appears here
Before82% blocked
After— awaiting photo
Change— points
Same drain— awaiting photo
StatusOpen
How the intelligence worksAI proposes. Evidence explains.
AI proposes. Evidence explains.
People decide.
See
A research-backed ResNet-50 classifies blockage locally in the browser. A drain-domain gate rejects uncertain photos; COCO-SSD is used only for visible litter.
Score
Central configuration drives cleanup priority and a separate environmental decision-support estimate. Missing waterway context lowers coverage instead of becoming a fake value.
Allocate
The system ranks inspections and turns limited capacity into a concise, traceable field plan for cleanup teams.
Verify
A normalized scene fingerprint must match the original drain before blockage reduction can close the task. Uncertain pairs go to human review.
Held-out AI evaluation
92.5% accuracy on unseen cameras.
Balanced blocked-versus-clear audit using the four cameras held out by the source research.
40labelled test images
Blocked recall100%20 / 20 blocked caught
Clear specificity85%17 / 20 clear correct
Precision87%3 false alarms
F1 score0.93balanced classification quality
Confusion matrix · held-out audit
| Actual | Predicted blocked | Predicted clear |
|---|---|---|
| Blocked | 20 TP | 0 FN |
| Clear | 3 FP | 17 TN |
Audit protocol
Threshold fixed on 28 images from seven different cameras, then tested once on 40 balanced images from four source-paper test cameras.
95% accuracy interval: 80.1–97.4%Open dataset + research weights ↗Scope: this is a reproducible proxy audit on UK trash-screen imagery, not claimed Bengaluru street-drain accuracy and not flood prediction. Uncertain field images still go to human review.
Safety regression
12/12 expected decisionsWorkflow logic also stays tested.
| Controlled test group | Cases | Expected behavior | Decision |
|---|---|---|---|
| Blocked drain controls | 3 | Remain open and ranked | 3/3 pass |
| Clear drain controls | 3 | Low obstruction signal | 3/3 pass |
| Same-drain cleaned pairs | 2 | Match scene, then verify reduction | 2/2 pass |
| Unchanged after evidence | 1 | Reject zero-point reduction | 1/1 pass |
| Different-scene after photos | 2 | Reject scene mismatch | 2/2 pass |
| Non-drain input | 1 | Route to human review | 1/1 pass |
Twelve deterministic checks cover ranking, clear evidence, unchanged cleanup, wrong-scene evidence, and non-drain review routing. These verify workflow policy separately from the model audit above.
Trust principles
Why should you trust this recommendation?
Every priority exposes the blockage, litter, rainfall, and context evidence contributing to it.
Unavailable map or weather context lowers coverage instead of becoming a fabricated value.
Low-confidence images, mismatched scenes, and uncertain cleanup evidence route to review.
Validation & limitations
Automated decision suiteWhat the prototype proves—and what it does not.
Executable checks
- ✓ Cleanup-priority weighting and thresholds
- ✓ Environmental-risk weighting and missing-context behavior
- ✓ Rainfall sensitivity across controlled scenarios
- ✓ Before/after scene mismatch detection
- ✓ Human-review routing and cleanup gates
Still requires field validation
- Drain detection and blockage-estimation accuracy
- Environmental-risk calibration
- Real-world precision, recall, and false-positive rates
- Bias across lighting, devices, seasons, and weather
- Coverage and freshness of mapped water features
Passing workflow tests demonstrates deterministic product behavior. It does not establish scientific validity or field-model accuracy.
Responsible use
DrainGuard supports inspection prioritization. It does not predict floods, measure pollution volume, or replace hydrological and engineering assessment. Scores depend on image quality, rainfall inputs, and available map context.
Prototype v0.15 · Field pilot