Anticipating climate hazards in the Senegal River Valley.
The Senegal River Valley is a strategic zone for Mauritania. The Gorgol and Trarza regions support agricultural production, livestock, and rural communities whose livelihoods are tied to the rhythms of the river and the rains. They are also exposed to a documented intensification of hydrometeorological hazards. Seasonal floods following intense rainfall, prolonged droughts during the dry season, Sahelian heatwaves, and sandstorms are all becoming more frequent and more severe, threatening food security, economic stability, and the social cohesion of the riverside populations.
UNDP Mauritania, working with national authorities including the National Meteorological Office, the Directorate General for Civil Security and Crisis Management, the National Society for Rural Development, and the local authorities of Kaédi and Rosso, mandated PANEOTECH to design and deliver a national early warning system that anticipates these hazards using AI-driven prediction. PANEOTECH led the engagement in joint venture with Effica SYS Co SARL, the Mauritanian technology firm that contributed in-country institutional coordination and continuity of presence on the ground. The result is HydroMet AI, an operational early warning system that combines deep learning, deterministic hydrological modelling, and a multi-channel diffusion architecture to give Mauritanian institutions and rural communities the lead time they need to act.
The mandate. Build a national early warning system for the Senegal River Valley that anticipates floods, droughts, heatwaves, and storms with the scientific credibility institutions require and the operational reach rural communities depend on. Combine AI prediction with deterministic hydrological modelling, validate every critical alert through institutional review, and diffuse across every channel that reaches the populations at risk.
An eight-microservice architecture, centred on the AI engine.
The platform is structured around eight autonomous microservices, each responsible for a discrete domain and communicating through standard interfaces. The modular architecture supports the operational reality that an early warning system has to evolve continuously. New stations join the network. Models recalibrate. Diffusion channels expand. Beneficiaries multiply. Each component evolves independently without disrupting the others.
AI Prediction Engine
The analytical heart of the system. A stacked LSTM recurrent neural network for time-series forecasting, an Isolation Forest for anomaly detection, and the SWAT hydrological model for watershed simulation work together as a hybrid prediction stack. The combination delivers a measured 12 percent precision gain over SWAT alone on rapid flood detection, and the LSTM achieves RMSE under 0.15, MAE under 0.1, and AUC above 0.9 on validation against ten years of historical observations.
Multi-Source Data Ingestion
Seven complementary global sources feed the prediction engine: Open-Meteo for weather forecasts, Open-Meteo Flood and GloFAS via Copernicus for fluvial discharge ensembles, NASA POWER for validated weather archives, the Copernicus ERA5 reanalysis for eight decades of climatological context, CHIRPS via Google Earth Engine for satellite precipitation, SMAP via Google Earth Engine for soil moisture, and the Copernicus Climate Data Store for historical flood reconstruction. Sentinel-2 and SRTM provide land cover and topographic data for the watershed model.
Alert Engine and Validation Workflow
The prediction engine produces graded risk scores classified as WARNING or CRITICAL across four hazard types: floods, droughts, heatwaves, and storms. Every critical alert passes through a mandatory institutional validation step before diffusion. Technical validation by the National Meteorological Office, operational validation by the Directorate General for Civil Security, and configurable rule-based pre-filtering protect the chain of accountability that public-safety alerting requires.
Multi-Channel Diffusion
Validated alerts diffuse across SMS, WhatsApp, email, web and mobile push notifications, radio scripts, and television scripts. Each channel reaches a distinct audience: SMS for institutional contacts, WhatsApp for community relays, email for formal institutional reporting, web and mobile for the operational dashboards, and broadcast scripts for the rural populations whose connectivity to digital channels remains limited. The combination is not redundant. It is layered for resilience.
Institutional Dashboards and Simulator
Five functional modules in the institutional dashboard. Pilotage gives supervisors a real-time overview. The Alert Centre orchestrates the validation workflow. The AI Simulator lets experts run hypothetical scenarios and inspect the algorithmic reasoning behind every prediction. Data Management governs ingestion, quality control, and station configuration. System and Security exposes user management, audit trails, and configuration. Every module is designed for institutional users with no command-line intervention required.
Mobile Application for the Last Mile
A Trusted Web Activity application published on Google Play under org.hydrometmr.twa, designed for community focal points, village relays, agricultural technicians, and women's cooperatives. Push notifications for validated alerts, geolocated field reporting with photos, an emergency contact directory, and degraded-mode operation with deferred synchronisation when connectivity is poor. Authentication is simplified through phone-based one-time passwords by SMS or WhatsApp.
The discipline behind a credible AI early warning system.
The architectural choices follow directly from the operational reality of an early warning system that informs decisions affecting lives. False alerts erode trust. Missed events erode trust faster. The model has to be defensible to the scientific community, traceable to the operators using it, and resilient to the data scarcity that defines its operating context. Each constraint shapes a specific design decision.
The hybrid model: SWAT plus LSTM, not one or the other
The prediction engine combines two scientific traditions that are usually treated as alternatives. SWAT, the Soil and Water Assessment Tool, encodes the physical principles of watershed hydrology in deterministic equations validated for decades. LSTM, the long short-term memory recurrent neural network, learns non-linear patterns from data. Used in isolation, each has known weaknesses. SWAT struggles when the regime departs from its calibration. LSTM struggles when the data is sparse. Used together in the architecture deployed on this engagement, they reinforce one another. SWAT provides the physical baseline that grounds the LSTM in scientifically defensible territory. The LSTM captures the non-linear dynamics that pure physical modelling misses. The hybrid stack delivers a measured 12 percent precision gain on rapid flood detection over SWAT alone, with the additional advantage of scientific traceability that pure deep learning cannot offer.
Engineering around national data scarcity
National Mauritanian hydrometeorological telemetry was not available in real-time machine-readable form at the start of the engagement. The architectural answer was a strategy built on global, institutional, scientifically validated sources rather than a wait for national infrastructure to mature. Open-Meteo, NASA POWER, the Copernicus Emergency Management Service, the Copernicus Climate Data Store, Google Earth Engine, the CHIRPS satellite precipitation product, the SMAP soil moisture mission, the ERA5 reanalysis, Sentinel-2, and SRTM together provide a continuous, scientifically defensible operating picture for the Senegal River Valley. The architecture remains open: when national sources mature, they slot in as additional inputs without redevelopment.
Upstream-downstream sentinel architecture
The Senegal River flows from southeast to northwest. When discharge rises at Bakel in eastern Senegal, it takes between two and five days to reach Kaédi in Mauritania. The system exploits this propagation explicitly. Four operational stations on Mauritanian territory, Kaédi, Rosso, Boghé, and Gouraye, define the public perimeter and receive alerts directly. Five transboundary sentinel stations, Bakel, Matam, Saldé, Manantali, and Diama, feed the prediction model as upstream inputs but never appear in public-facing outputs. The architecture mirrors standard practice in operational hydrology and gives Mauritanian populations the maximum possible lead time.
Human-in-the-loop validation for every critical alert
Every critical alert passes through a mandatory human validation step before diffusion. The architecture enforces it: no path exists in the system through which a CRITICAL alert reaches the public diffusion channels without an authorised operator validating it first. The choice is deliberate and structurally consequential. It preserves the institutional accountability that public safety alerting requires, filters anomalies the model may produce in genuinely unprecedented climatic regimes, keeps operators in the decision loop and maintains their expertise, and reassures public authorities that the AI dispositif remains under controlled supervision rather than operating autonomously.
Eight microservices on AWS, with operational discipline
The platform runs as eight autonomous microservices on Amazon Web Services with measured availability above 99.9 percent. Cloudflare provides web application firewall, DDoS protection, content delivery, and edge caching. Grafana exposes monitoring dashboards and alerts. RabbitMQ decouples ingestion from prediction. Amazon S3 holds generated reports, audio bulletins, log files, and backups. The operational discipline includes ISO/IEC 27001 alignment, OWASP development practices, TLS 1.3 for transport, AES-256 for storage, immutable audit logs, two-factor authentication, role-based access control, and a documented disaster recovery procedure with a recovery time objective under four hours.
API-first design and institutional sovereignty
A versioned REST API is exposed at app.hydrometai.tech/api/v1 with comprehensive documentation, including code examples in Python for IoT sensor integration on Raspberry Pi devices and JavaScript and React Native for partner mobile applications. The API turns the platform into an open infrastructure on which other actors, researchers, NGOs, startups, and partner institutions can build derivative services. The full source code is delivered to UNDP Mauritania, with a documented protocol for transfer to a Mauritanian sovereign infrastructure or to ONM-controlled hosting whenever the institution chooses to activate it.
Built for institutions, communities, and continuity.
The system serves the institutional and community actors mandated by the project. The National Meteorological Office uses the platform for technical supervision and alert validation. The Directorate General for Civil Security and Crisis Management uses it to trigger response plans. The National Society for Rural Development supports producers through the platform. Regional Agricultural Delegations operate it across the pilot zones. Local authorities in Kaédi and Rosso, community focal points, agricultural cooperatives including women's groups, and the rural producers and pastoralists at the receiving end of every alert constitute the population the system was built to protect. A field training campaign was conducted in Kaédi from the twelfth to the fourteenth of January 2026 and in Rosso from the sixteenth to the eighteenth of January 2026, followed by an institutional supervision mission in Nouakchott from the twenty-fifth to the thirtieth of January 2026. The launch ceremony was held on the thirteenth of February 2026.
Joint venture and institutional ownership.
HydroMet AI was developed by PANEOTECH in joint venture with Effica SYS Co SARL, the Mauritanian technology firm that contributed in-country coordination, technical support, and continuity of presence on the ground in Nouakchott and across the pilot zones. PANEOTECH led the technical architecture, the AI prediction engine, the data integration pipeline, the institutional dashboards, the multi-channel diffusion infrastructure, the mobile application, and the API. Effica SYS led the institutional coordination with the National Meteorological Office, the Directorate General for Civil Security, and the National Society for Rural Development, and ensured continuity of presence during the field training and the supervision mission.
The platform is in production at app.hydrometai.tech for institutional users and at hydrometmr.org for partner access, with the mobile application published on Google Play. UNDP Mauritania holds full operational ownership, and the engagement included structured capacity transfer with administrator training, user guides per profile, hosting and maintenance documentation, an API specification, the consolidated test acceptance file, and a centralised resources platform accessible to authorised administrators of UNDP, the National Meteorological Office, the Directorate General for Civil Security, and the National Society for Rural Development. The full source code is delivered, with a documented protocol for transfer to a Mauritanian sovereign infrastructure whenever the institution chooses to activate it.
