CellUlan: Turning Smart cellular microwave towers into a zero-hardware national rain and flash-flood radar

Problem Statement

Flash floods in the Philippines often start from intense, localized mountain cloudbursts that fall between physical rain gauges. Maintaining thousands of physical rain gauges and Doppler radars across 7,641 islands costs millions of pesos each year, and gauges frequently fail during typhoons due to clogged funnels, stolen solar panels, or cut power lines. Satellite rainfall estimates such as GPM IMERG suffer from 30-minute to 3-hour latency and 10-kilometer pixel resolution, which is too coarse to tell a municipal disaster officer whether heavy rain hit the exact sub-watershed feeding their town’s river. Meanwhile, Smart Communications already operates thousands of point-to-point commercial microwave backhaul links across Philippine mountains and valleys that measure radio signal strength every minute.

Proposed Solution

CellUlan repurposes Smart’s existing cellular microwave backhaul network into a high-density virtual rain gauge array with zero new field hardware. Point-to-point microwave links operating between 15 GHz and 38 GHz experience measurable signal attenuation whenever raindrops pass through the beam between two cell towers. Network management systems already log Transmitted Signal Level and Received Signal Level every 1 to 15 minutes to monitor link quality.

CellUlan ingests this anonymized link telemetry on ePLDT, separates rain-induced attenuation from baseline atmospheric noise, applies the ITU-R P.838 power-law physics relationship, and runs a spatio-temporal Graph Neural Network to reconstruct a 500-meter rainfall intensity grid across every barangay every 60 seconds. When accumulated upland rainfall along a watershed exceeds the runoff threshold on Project NOAH maps, CellUlan sends an automated Smart SMS warning to downstream barangay captains and streams live rain-corridor vectors to responders wearing Rokid AR glasses.

Target Market

• Primary users: Municipal and provincial DRRM offices across 1,600+ LGUs, PAGASA hydro-meteorologists, and NDRRMC operations centers.
• Internal sponsor user: Smart Communications Network Operations Center and PLDT Enterprise B2G climate resilience teams.
• B2G and B2B buyers: Local government units using their 5 percent Local DRRM Fund under RA 10121, water utilities and dam operators such as MWSS and NIA, renewable energy operators, and agricultural insurers such as PCIC.

How AI Is Used

• Wet-dry period classification via 1D-CNN and LSTM: Distinguishes genuine rain attenuation from multipath fading, antenna wetting, and temperature drift in noisy Received Signal Level time series.
• Physics-informed Graph Neural Network: Treats cell towers as graph nodes and microwave line-of-sight paths as directed edges, solving the tomographic inverse problem to turn path-averaged attenuation into a 2D gridded rainfall field.
• Hydrological runoff nowcasting: Combines the reconstructed 60-second rainfall grid with digital elevation models and soil saturation indices to predict downstream flash-flood arrival times 30 to 90 minutes in advance.
• LLM dispatch summarizer: Generates plain Tagalog, Bisaya, and Ilocano flash-flood advisories naming the exact upstream microwave corridor where heavy rainfall was measured.

PLDT / Smart / ePLDT / Rokid Angle

• Smart Communications: Unlocks a new public-good and commercial product from network telemetry that Smart already collects in its Network Management System at zero hardware capex.
• ePLDT: Hosts the real-time telemetry ingestion pipeline and spatio-temporal GNN inference engine on sovereign VITRO data centers so critical telecom topology never leaves PLDT infrastructure.
• PLDT Enterprise: Packages hyperlocal weather and flood intelligence as a managed B2G subscription for LGUs, dam operators, and logistics fleets.
• De La Salle University: Connects directly with DLSU Gokongwei College of Engineering and Center for Complexity and Emerging Technologies research on opportunistic wireless sensor networks and urban flood modeling.
• Rokid Philippines: Field rescue leaders wearing Rokid AR glasses see live upstream rain intensity cells and projected river rise timers overlaid on their field of view during night operations.

Demo Scenario (3 minutes for judges)

1. Display the map of the Marikina and Rizal upper watershed during a simulated monsoon squall where only two physical rain gauges exist, but 48 Smart microwave backhaul links crisscross the valley.
2. Stream a replay of microwave Received Signal Level drops in dBm across the 23 GHz and 38 GHz links. The wet-dry classifier flags active precipitation along the Montalban ridge within 15 seconds.
3. The Graph Neural Network reconstructs the moving rain cell at 500-meter resolution, revealing a 65 mm/hr cloudburst over an un-gauged sub-basin that standard satellite feeds miss.
4. CellUlan computes a 45-minute flood crest warning for three downstream barangays and dispatches a Tagalog SMS to a judge’s phone while projecting the flood arrival countdown onto the Rokid AR display view.
5. Close with the core metric: “Every Smart cell tower becomes a weather station, giving the Philippines 10 times more rain sensors overnight with zero hardware cost.”

30-Hour Build Plan

Suggested stack: Python FastAPI, PyTorch Geometric, pycomlink and rainbow open-source CML toolkits, PostgreSQL/PostGIS, Uber H3, Next.js + deck.gl, Gemini API for localized advisories. Public data / APIs: OpenMRG dataset from SMHI for real commercial microwave link attenuation and reference radar validation, OpenCelliD and NTC public tower coordinates, PAGASA historical gauge records, Project NOAH watershed shapefiles, Copernicus DEM.

Research & Literature References

1. Hagit Messer, Artem Zinevich, and Pinhas Alpert, “Environmental Monitoring by Wireless Communication Networks,” Science, Vol. 312, Issue 5774, p. 713, 2006. DOI: 10.1126/science.1120034
   • How CellUlan uses this study: Establishes the physical law that radio signal attenuation $A$ across commercial cellular microwave links relates to path-averaged rainfall rate $R$ via the power law $A = aR^b$, proving that cellular backhaul networks function as accurate atmospheric rain sensors.
2. Aart Overeem, Hidde Leijnse, and Remko Uijlenhoet, “Country-wide rainfall maps from cellular communication networks,” Proceedings of the National Academy of Sciences, Vol. 110, Issue 8, pp. 2741-2745, 2013. DOI: 10.1073/pnas.1217961110
   • How CellUlan uses this study: Provides the national-scale tomographic interpolation algorithm and wet-dry baseline separation method used to transform thousands of individual tower-to-tower links into continuous country-wide precipitation maps.
3. Maximilian Graf, Christian Chwala, Julius Polz, and Harald Kunstmann, “Rainfall estimation from a German-wide commercial microwave link network: optimized processing and validation for 1 year of data,” Hydrology and Earth System Sciences, Vol. 24, Issue 6, pp. 2931-2950, 2020. DOI: 10.5194/hess-24-2931-2020
   • How CellUlan uses this study: Supplies the open-source pycomlink processing pipeline and wet-antenna attenuation correction formulas that CellUlan adapts for tropical convective rainfall in the Philippines.

Risks & Privacy Notes

• Zero subscriber personal data: CellUlan processes only tower-to-tower radio signal strength in dBm and link coordinates. It never touches call records, subscriber identities, or mobile traffic payloads, complying with the Data Privacy Act of 2012.
• Telco topology confidentiality: Exact backhaul link coordinates are sensitive infrastructure data. In the public dashboard, individual tower links are aggregated into 500-meter H3 hexagonal rainfall cells before leaving ePLDT.
• Wet-antenna overestimation: Water film on microwave radomes adds 1 to 3 dB of extra attenuation during heavy downpours. CellUlan applies the Graf et al. wet-antenna correction curve and cross-calibrates against nearby PAGASA ground gauges.
• Clause 24 compliance: All hackathon development uses the public OpenMRG microwave dataset and OpenCelliD coordinates rather than proprietary Smart internal logs.
• Strict 30-hour feasibility constraint: Student teams do not have direct access to Smart internal Network Management System (NMS) SNMP feeds. During the hackathon, the model can only be validated on European OpenMRG microwave logs mapped onto Philippine terrain.

AI Use Disclosure (draft)

CellUlan uses a 1D-CNN/LSTM classifier for wet-dry link detection, a PyTorch Geometric Graph Neural Network for spatial rainfall reconstruction, and an LLM to draft localized Tagalog and Bisaya flood warnings. The team used AI coding assistants to scaffold the PostGIS aggregation pipeline and deck.gl visualization layers. All physical attenuation formulas and hydrological thresholds were verified against peer-reviewed literature.

Rubric Scoring