A forecast you can check, not just read

Attribution

Ashif Ahmed Shuvo (Master's thesis author, ORCID 0009-0003-5734-1519) — HazardNet: multi-hazard early warning for Bangladesh agriculture. Master's thesis, Department of Agrometeorology, Bangladesh Agricultural University (BAU), supervised by Dr. Ahmed Khairul Hasan (Supervisor).

Shuvo, A. A. HazardNet: multi-hazard early warning for Bangladesh agriculture. Master's thesis, Department of Agrometeorology, Bangladesh Agricultural University. https://github.com/myself-aas/HazardNet

Start here: what to do with an outlook

Use an outlook as lead time, in this order: find your district, read the hazard, the horizon, the forecast date, the severity and the confidence together, then check the official bulletin for that hazard before you act. The sections below explain each of those pieces; the links here are the working surfaces.

• For a farm decision (planting, spraying, irrigating or harvesting), a district outlook tells you when to pay attention, not what to do in a particular field. Confirm the step locally, with extension advice where you can.
• For an extension or union parishad decision, start from the published alerts and the sector advisories: each alert carries its evidence and policy version, and each advisory names the issuing authority at each step.
• For a research or review question, start from the validation scorecard and the model card rather than from a colour on the map: they state what was checked, and what the counts cannot support.
• In an emergency, do not act on this page. Call 999, and follow your local administration and the official BMD or FFWC bulletin over anything published here.

What HazardNet is for, and who it serves

The operational question in Bangladesh agriculture is rarely whether a hazard is coming; it is which districts, at which crop stage, with how many days of lead time, and how much weight the number deserves. HazardNet is built to answer that second set of questions. For each district and horizon it publishes a hazard class, a severity index, a confidence bin and an independent physics estimate: and the run they came from.

It is written for three readers at once: farmers and rural residents checking the outlook for their own district before planting, spraying, irrigating or harvesting; agricultural extension workers and union parishad officials deciding whom to warn and what to prepare; and researchers and reviewers judging whether a published number deserves the weight it asks for.

• 64 districts on the live map; the published advisory archive covers 507 ADM3 units (495 upazilas plus 12 city corporations).
• Eight hazard classes: cold wave, drought, fire, flash flood, flood, heat wave, severe local storm and tropical cyclone.
• The aim is a better district-level decision, on a farm, in a parishad office or in a review, made knowing exactly how much to trust the number. It is not a replacement for an official warning.

The current outlook, and where to see it

This page is the record; the live map is the outlook. The split is deliberate and documented, because the two questions need different pages: "where is the hazard now" is answered by a map, and "who produced this, from what, and how would I know if it stopped working" is answered by a document with dates on it.

The strip under the headline reads the committed alert artifact on every load: how many alerts are published at each level, how many districts the last run covered, and how many assessed rows the publication gate withheld. When nothing is published it says so, and says why: silence on a hazard platform is read as safety, so this page does not leave that to be inferred.

• The live map addresses all 64 districts. A district the current run does not cover keeps its static baseline colour and says so on its own page, rather than borrowing a neighbour's number.
• Every published alert carries its own evidence: the drivers behind the level, the policy version applied, and whether a duty officer reviewed it or the engine published it on its own at or below its configured ceiling.
• The status page carries the age of every artifact this deployment ships, measured against the SLO it is held to, together with the notes each run reports against itself.

What a forecast is, and how it differs from an alert

Every forecast on HazardNet is a published record: one district, one hazard class, a severity value, a confidence score and 7- and 15-day horizons, stamped with the date of the run that produced it. No forecast, alert or figure on this site is produced by a language model: every number is read from a committed artifact, and no number is interpolated to fill a gap in a run.

An alert is narrower than a forecast. It is a forecast row the publication gate has released as a card, with its level, evidence, policy version and reviewer state attached. A district can therefore have a forecast on the map while no alert is published for it; and, as the strip under the headline states, no published alert is not the same as no hazard.

• Results only: how a record is generated is research-private and is not published.
• Every figure on this site traces to a published forecast record with a forecast date.
• Two severity tracks are reported per record, a skill track and a physics track, each shown with its own honest label.
• The 7-day horizon covers the week ahead. The 15-day horizon covers planning further ahead, and should be read as the less certain of the two.

Severity, alert levels and confidence are three different things

A forecast record carries three values that are easy to merge into one impression of danger. Severity is the value recorded for one district, hazard class and horizon. The alert level is the published tier a released alert carries. Confidence is neither of those: it is an uncalibrated model score (a softmax output), not a probability that the event will happen, and it should always be read beside the severity value and the forecast date.

• No alert: nothing unusual for the season in the published alert record. On forecast records the lowest advisory tier is named NORMAL; the two labels belong to two related lists, and neither is a promise that a district is safe.
• Watch (monitor): conditions are favourable for a hazard.
• Warning (prepare): a damaging event is plausible.
• Severe (act): an event is imminent or likely. What action is safe is set by the official bulletin and your local administration, not by this platform.
• A level is carried as a word as well as a colour, so it stays readable in glare, and for readers who cannot rely on colour alone.

These are HazardNet's decision-support tiers. They are not BMD cyclone signals, FFWC flood bulletins or evacuation instructions, and an official bulletin outranks them wherever the two differ.

How it works, and when it updates

Each run publishes dated data files. This site calls them artifacts. The live map, the alert cards, the status page and this front door all read the same files, so any figure can be followed back to the file and forecast date behind it instead of being taken on trust from a page design.

Forecasts are updated in daily runs, so freshness is something you can check rather than assume: the strip under the headline gives the alerts file's age at the moment of this read, every forecast record carries its forecast date, and the status page states each artifact's age against the limit (its SLO) it is held to.

• What is published here is the result, and the evidence needed to check it. How a record is generated (model code, dataset collection and training) is research-private and is not described on this site.
• No forecast, alert or figure on this site is produced by a language model, and no number is interpolated to fill a gap in a run.
• A district a run did not cover says so on its own page. It does not borrow a neighbour's number, and it is not shown as zero.

Validation, and how to check any number on this site

A claim that cannot be checked is marketing. The platform therefore publishes the evidence next to the claim, including the places where the evidence is thin.

• The status page reads the committed artifacts and states the age of each against its SLO: including when a source is stale, failing or unprovenanced, which is the case the page exists to show.
• The validation scorecard reports per-episode detection counts for five historical episodes and states plainly what those numbers cannot support.
• The model card lists the known failure modes and the out-of-scope uses, and carries the 2,931-event training claim as reported rather than verified: that archive is not redistributed by this repository, and the pages that would need it say it is absent instead of quoting the figure as if it had been read.
• The code, the artifacts and the build scripts are in one public repository, and this site is rebuilt from them: the same pipeline that publishes a number also publishes the check on it.

Access, languages, and when something is missing

HazardNet is free and non-commercial on the web, and the map and the published record need no account. The app installs as a PWA, a web app you can add to a phone's home screen. It caches its tiles and keeps working when the connection does not. The pages are built for low-end Android phones, slow or metered connections and reading in sunlight, and the download centre carries the forecast snapshot and the offline builds for use away from a connection. Printed district briefs are a first-class output, not an afterthought.

Bangla and English sit side by side. The switch at the top of this page renders the front door's editorial copy and its controls in Bengali; the alert cards, the map and the district tables were translated first. The long-form knowledge products (the model card, the methodology and the data-source ledger) remain English-only, and where a sentence has no Bengali version the English one is shown rather than a gap.

• An em dash (—), a value left blank, or the words “not reported” mean the run or the file did not supply that value. They never mean zero.
• An unread file is not a count of zero, and no published alert is not the same as no hazard. The live strip states those cases separately, so a loading failure cannot be mistaken for safety.
• If a panel cannot load, use its Try again button where one is offered, then check the status page to see whether the source file itself is stale or failing. The rest of the page stays readable while one panel is missing.
• To report a broken page, a stale run or a prediction that looks wrong, use the contact page and include the district and the date: that is enough to act on.

Knowledge products

The dated products this deployment publishes. The reviewed date is the review date the page itself carries, so a stale product can be seen to be stale.

The official record, and where it lives

HazardNet does not issue warnings, cyclone signals or evacuation instructions, and it must never be read as doing so. The bodies that do are named below; their bulletins outrank anything on this site.

• Bangladesh Meteorological Department (BMD): weather forecasts, cyclone signals, and heat, cold and storm warnings.
• Flood Forecasting and Warning Centre (FFWC): river-level forecasts and flood bulletins.
• Department of Disaster Management (DDM), Ministry of Disaster Management and Relief: the Standing Orders on Disaster, relief and response.
• In an emergency, call the national hotline 999, and follow your local administration’s instructions over anything published here.

Take part

The platform is maintained as a Master’s thesis project at Bangladesh Agricultural University, and it improves fastest where practitioners push back on it.

• Use it in the field and report where it was wrong. A district, a date and what you observed is enough to act on.
• Reuse the data: the download centre carries the forecast snapshot and the offline builds, and the code is MIT-licensed with the citation given on this page.
• Report a broken page, a stale run or a prediction that looks wrong from the contact page: every report is read.

Questions and answers