Inspiration and Track 6 alignment
Urban river monitoring teams need to turn a weather warning into a useful field question: where should we investigate, what should we measure, and how can we preserve the evidence? StreamSignal addresses Track 6: Resilience Informatics with weather alerts and a reviewable sampling workflow for researchers, citizen scientists, and monitoring teams.
What it does
The prototype covers five European regions, 106 source sites, and 21 distinct weather groups. It shows calibrated probabilities for heavy rain, heat, low rainfall, and hot-and-dry conditions over three and seven days. Users can inspect alert definitions, compare cities, and generate up to five visit suggestions covering different weather groups. Historical urban and One Health context explains investigation priorities without asserting that contamination has occurred.
Mission control produces a printable investigation brief and tentative calendar export. The field notebook records measurements and links eligible entries to earlier forecasts. Citizen mode supports plain-language observations. FHIR R4 Location/Observation exports provide an interoperability starting point. Field records remain local; the offline workspace requires an initial successful online visit, and live forecast refresh needs internet. Existing translations are drafts.
How it was built
The interface uses HTML, CSS, JavaScript, and a service worker, with live weather from Open-Meteo. Python and scikit-learn support reproducible experiments and probability calibration. Forecast inputs, parameters, dates, and model versions are archived so later observations can be interpreted in context. The public repository includes documentation, tests, and complete results.
Evidence, challenges, and lessons
All eight aggregate task comparisons show improved retrospective Brier skill on matched windows. For example, seven-day heavy rain improves from 20.1% to 52.1%, and three-day heat from 28.6% to 79.2%. These measure probability-error reduction against a seasonal baseline, not classification accuracy. Some city-task results remain uncertain. Modeled reference weather, repeated inputs, overlapping windows, and differences between scored and deployed configurations limit interpretation. The rain threshold revision after earlier 2026 outcomes were viewed is disclosed in the methods.
The central lesson is to keep weather forecasts, investigation hypotheses, and actual water measurements distinct. The prototype does not predict contamination, water quality, or health outcomes, and live performance has not been prospectively validated.
Expected impact and next steps
The intended benefit is more useful ecosystem observations within a fixed visit budget. A prospective pilot with monitoring teams would compare verified water-quality changes, missed changes, response time, and workload against routine monitoring. Ecological or human-health improvements have not yet been demonstrated.
Demo, source, and attribution
The captioned video is about 4 minutes 42 seconds and uses the 4 October 2026, 18:07 UTC forecast snapshot; live forecasts change. Original StreamSignal code is MIT licensed; third-party datasets, weather, and assets retain their separate terms. This is an independent prototype with no implied IEEE or OneAquaHealth endorsement.
Assistance disclosure
Codex assisted with audio editing, captions, video production, and preparation of this submission.
Log in or sign up for Devpost to join the conversation.