BahaWatch: Every CCTV and phone camera becomes a flood gauge

Problem Statement

Flooding in Metro Manila and other urban areas is a yearly crisis. Typhoon Ondoy (2009) and Typhoon Carina with the enhanced habagat (2024) left streets submerged and commuters stranded. Official water-level sensors measure rivers, not streets, and there are far too few to cover thousands of flood-prone roads. MMDA’s flood gauges use categories like “gutter-deep” and “tire-deep,” but reports are manual and sparse. Meanwhile, thousands of barangay, LGU, and private CCTV cameras already watch these streets, and residents post flood photos on Facebook. None of this is turned into a structured, real-time flood-depth map that people can use to decide whether to leave, reroute, or evacuate.

Proposed Solution

BahaWatch pulls in frames from opted-in CCTV feeds (barangay, LGU, MMDA, malls) and photos submitted through a Messenger bot or web app. A vision pipeline detects reference objects such as car tires, people’s legs and waists, jeepney bumpers, curbs, and utility posts, then estimates water depth relative to their known heights. It maps the result to MMDA-style passability levels. Each estimate is geolocated and time-stamped, and multiple sources are combined to boost confidence. A live map shows flood hotspots, trends over time, and “passable for” labels (pedestrians, light vehicles, trucks). Commuters can subscribe to routes and get SMS or Messenger warnings (“EDSA–Quezon Ave underpass: waist-deep, avoid”). LGUs get a dashboard with alert thresholds that trigger barangay evacuation advisories, and the data also improves flood models over time.

Target Market

• Primary users: Commuters, motorists, delivery and ride-hailing riders, and residents of flood-prone barangays in Metro Manila, Cebu, Davao, and Pampanga.
• Buyers (B2G): MMDA, city DRRMOs, DPWH, and barangays with CCTV networks.
• B2B: Logistics and delivery fleets (Grab, Lalamove, couriers), insurers (motor flood claims), malls, and property managers.
• Size: Millions of daily commuters in Metro Manila alone, plus all highly urbanized cities.

How AI Is Used

• Object detection and segmentation: Finds water surfaces and reference objects in noisy, rainy, night-time frames. Without it, raw CCTV is just video nobody watches.
• Depth estimation via reference-object geometry: Converts the waterline on a tire or leg into centimeters. This is the key innovation and impossible to do by hand at city scale.
• Multimodal LLM verification: Sanity-checks crowdsourced photos (is this actually flooding, is it recent, does the location make sense) and filters out old or fake viral images.
• Short-term forecasting: Combines depth time series with PAGASA rainfall data to project “rising” or “receding” for the next hour.

PLDT / Smart / ePLDT / Rokid Angle

• PLDT Enterprise managed CCTV and fiber: Many LGU and barangay cameras run on PLDT connectivity. BahaWatch becomes an AI value-add on top of existing contracts.
• ePLDT cloud/edge GPU for video analytics, with data stored locally in PH data centers.
• Smart SMS and location-based alerts for route subscribers, and Smart commuters as a distribution base.
• Rokid (stretch): Rescue boat crews or MMDA traffic enforcers see live depth readings for the road ahead on AR glasses.

Demo Scenario (3 minutes for judges)

1. Show actual public YouTube/news footage of a flooded Metro Manila street during Carina (2024), with credit.
2. BahaWatch draws boxes around cars and pedestrians and labels them “Waterline at 60% tire height, about 35 cm, not passable to light vehicles.”
3. A judge snaps a photo of a staged scene (toy car in a basin, or a printed flood photo) and sends it to the Messenger bot. Within seconds a new pin appears on the map.
4. The map lights up three hotspots. A route from Taft Ave to España turns red, and a teammate’s phone gets an SMS reroute warning.
5. Close: “Thousands of cameras are already watching our streets. BahaWatch makes them warn us.”

30-Hour Build Plan

Suggested stack: Python, YOLOv8-seg / Segment Anything, a multimodal LLM (Gemini/GPT-class), FastAPI, PostGIS, Mapbox/Leaflet, OSRM for routing, Messenger Platform API, SMS gateway. Public data / APIs: Project NOAH flood hazard maps, PAGASA rainfall advisories, MMDA flood gauge advisories, OpenStreetMap roads, public flood datasets on Hugging Face/Kaggle (e.g., flood segmentation and “European Flood 2013”-style depth sets), Sentinel-1 SAR for validation.

Risks & Privacy Notes

• Depth estimation error: Show depth bands rather than exact cm, show confidence scores, require agreement from multiple sources before public alerts, and calibrate each fixed CCTV with a one-time reference marker.
• Fake or old images: EXIF/time checks, LLM reverse-context checks, and reporter reputation.
• CCTV access: Partner with LGUs through data-sharing agreements, and process frames at the edge so only depth values leave the site.
• Data Privacy Act 2012: Blur faces and plates automatically before storage, don’t store raw video, keep only aggregated depth data, and request minimal personal info from photo submitters.
• Clause 24 compliance: Use only public, licensed footage and staged photos in development. No identifiable personal images are sent to third-party AI tools without anonymization.

AI Use Disclosure (draft)

BahaWatch uses YOLO/SAM-class vision models for detection and segmentation, a multimodal LLM to verify crowdsourced photos, and a simple forecasting model for flood trends. AI coding assistants helped with code scaffolding and drafting this document. Public alerts are shown with confidence levels and reviewed by LGU operators, and the team reviewed all AI-assisted outputs.

Rubric Scoring