The Agromet Gap

Executive Summary. Bangladesh has the forecasts. In April 2026, they failed again. Flash floods inundated roughly 16,000 hectares of boro paddy in Sunamganj and 7,000 hectares in Kishoreganj at the peak of harvest season, affecting over 112,000 farmers across the haor belt. ECMWF model output indicated the rainfall anomaly days in advance. The Bangladesh Meteorological Department subscribed to it. No upazila-specific crop-stage advisory reached the farmers in time. The Tarique Rahman government, sworn in February 2026, has committed to digital public services as a pillar of its economic programme; closing the agromet bridge is the cheapest, most verifiable test of that commitment in the agricultural domain.

A boro paddy farmer in Sunamganj plants a crop in December. By late March, the spikelets are forming inside the booting panicles, the most temperature-sensitive moment of the rice life cycle. If the daytime maximum crosses thirty-five degrees Celsius for more than a few days at this stage, sterility sets in and the grains never fill. If a sixty-millimetre flash rainfall slams the haor in early April, the standing crop drowns before it can be cut. Both events have happened in the past decade. The 2017 haor flood washed out roughly nine hundred thousand metric tonnes of boro before harvest, the worst single-season loss in Bangladesh agricultural history. Heat-stress sterility events have hit boro fields in 2014, 2017, and 2024.

The forecast that would have warned about each of these events existed. It exists now. The European Centre for Medium-Range Weather Forecasts (ECMWF) integrated forecast system runs a 0.25-degree global model four times a day; on the days leading up to the 2017 haor flood, it correctly indicated a heavy precipitation anomaly over the upper Meghna basin three to four days in advance. The Bangladesh Meteorological Department subscribes to and re-distributes this output. The Department of Agricultural Extension has a block-level supervisor structure across all four hundred ninety-two upazilas of the country.

What did not exist, in 2017 or in 2024, is the bridge between the forecast on the ECMWF server and the action on a farmer's field. Bangladesh has the science. It has the institutions. It even has the forecasts. What it does not have is the agromet bridge that turns numbers in a server in Reading into a Bangla SMS that reaches a Sunamganj boro farmer at five in the morning of the right day. This is the agromet gap. It is the most fixable, most under-discussed, and most consequential failure in Bangladesh's climate-adaptation stack.

What an agromet system actually is

A working agromet system has four layers, in order. The first is a forecast that resolves to a useful spatial scale. For Bangladesh that is the upazila, of which there are four hundred ninety-two. The second is a translation layer that converts the forecast into crop-stage-aware risk. A thirty-six-degree maximum temperature is a benign weather forecast. A thirty-six-degree maximum during the booting stage of the boro that a specific union is currently growing is a sterility warning. Without the crop-stage layer, the same forecast either over-warns farmers (every March is hot) or under-warns them (an April heat day is fine for harvest, fatal at flowering). The third layer is operational language. A farmer does not need a probability distribution; the farmer needs to know whether to drain the field, delay topdressing, or cut the crop today. The fourth layer is a delivery channel that reaches the farmer in time. SMS, push notification through the bKash or Nagad farmer-account base, IVR for non-literate farmers, and union-parishad-level mass loudspeakers each cover a different segment of the farmer base.

India runs all four layers through the Gramin Krishi Mausam Sewa, an India Meteorological Department programme that delivers crop-and-stage-specific advisories from district-level Agromet Field Units. Vietnam's Vietnam Agromet Service does similar work through the Hydro-Meteorological Forecasting Centre. Kenya runs the Kenya Meteorological Department Agromet bulletin in partnership with the Ministry of Agriculture and IFAD. None of these countries has better forecasts than Bangladesh has access to. Each has more layers of the bridge.

What Bangladesh has, and what it does not

Bangladesh has the first layer. The BMD subscribes to ECMWF, runs the regional WRF with assimilation from a network of automatic weather stations, and operates an agromet division. The Bangladesh Water Development Board's Flood Forecasting and Warning Centre runs a hydraulic model on the major river basins and issues daily flood forecasts during the monsoon. The BARC publishes a weekly agromet bulletin in English, distributed primarily to research institutions and government offices. None of this is missing.

What is missing is the second, third, and fourth layers. The translation from forecast to crop-stage risk, the operational Bangla language, and the delivery to a farmer's hand at the right time, are simply not happening at scale. The DAE has the union-level supervisor network that could deliver, but it does not receive a structured upstream feed from BMD that says "boro at flowering, Sunamganj union 17, heat warning Tuesday." It receives the same general-public weather bulletin everyone else does, and is asked to interpret it on the fly with no decision support.

The agromet bulletin that BARC publishes weekly is a useful research document. It is not an actionable advisory: it arrives by Friday email, covers the past week and a generic outlook, and is in English. By the time a Sunamganj agromet officer translates the relevant pieces into operational guidance for the union, the forecast window has often closed. Cyclone Sitrang in October 2022 showed the gap clearly. The ECMWF and JTWC tracks were available three days in advance. BMD signal warnings were timely. But farmers in Patuakhali and Bhola had no upazila-and-crop-specific advisory telling them to harvest the eighty-percent-ripe aman paddy in the field before the cyclone landed. Aman losses in the affected coastal districts were substantial; many were avoidable.

This is not a science problem. The forecast was right. It is a translation, language, and delivery problem.

What we built, and why that proves the point

To make this argument concrete rather than abstract, BDPolicyLab built a reference implementation of the missing layers in a single developer week, using public APIs alone. It runs at /early-warning. The system pulls hourly forecasts from Open-Meteo, which itself blends ECMWF with regional models, for thirty-two grid points covering all eight divisions of Bangladesh. It computes per-district seven-day extremes for six variables: cumulative rainfall, maximum and minimum temperature, maximum gust, peak convective available potential energy, and peak relative humidity. It joins these to a static crop calendar covering boro, aman, aus, wheat, potato, jute, mango, and hilsa, with each crop's monthly stage encoded as sowing, vegetative, reproductive, or harvest. And it applies a rules engine that fires district-and-crop-specific Bangla advisories when forecast variables cross thresholds drawn from BRRI, BARI, DAE, BMD, and BFRI extension materials.

The thresholds are sourced. Boro spikelet sterility at maximum temperatures above thirty-five degrees Celsius during reproductive stage comes from Bangladesh Rice Research Institute extension literature. Cold-induced sterility at minimum night temperatures below seventeen degrees comes from the same source. The Smith period proxy for potato late blight, minimum temperature ten to eighteen degrees with maximum relative humidity above eighty-five percent, is the standard DAE potato-block advisory trigger. The kalbaishakhi signature, convective available potential energy above fifteen hundred joules per kilogram combined with a fifty-kilometre-per-hour gust forecast, is the conventional pre-monsoon thunderstorm climatology used in BMD bulletin context. Cyclone gust thresholds match the BMD signal scale: sixty kilometres per hour for local warning, ninety for storm signal, one hundred twenty for great danger.

The rules engine emits Bangla actions, not weather metaphors. For mango farmers in the Rajshahi-Chapainawabganj belt during a forecast kalbaishakhi, the advisory reads: harvest mature fruit within twenty-four to forty-eight hours; reinforce supports for younger fruit clusters. For boro at harvest under the same forecast: cut ripe paddy today, cover cut fields, avoid open fields during lightning. For potato at vegetative stage under a Smith-period forecast: apply preventive mancozeb seventy-five percent WP at two grams per litre before sunrise, switch to metalaxyl-based fungicide if symptoms appear. Each card on the page shows the threshold that fired, the institution that owns the threshold, and the source.

This is not a sophisticated system. It is one developer-week of work using free public APIs. The point is exactly that. If a single developer can build the missing translation, language, and delivery layer in a week with no proprietary infrastructure, the question of why a national programme has not done so for two decades is institutional, not technical.

Why has it not happened

The institutional answer has three pieces. First, none of the three relevant agencies sees this work as its primary mandate. BMD's mandate is general public weather and aviation; agromet is a sub-division within it but does not get the cross-agency authority to push structured forecasts into DAE delivery channels. DAE's mandate is extension and input distribution; weather forecasting is not a core competence. BARC publishes the weekly bulletin but does not run an operational delivery channel. Each agency does its own piece; nobody owns the bridge.

Second, the funding architecture mirrors the mandate gap. BMD's budget is set by the Ministry of Defence; DAE's by the Ministry of Agriculture; BARC's by the Ministry of Agriculture but as a research line. None of the three line items has a head dedicated to operational agromet. Donor-funded pilots have happened, including World Bank and IFAD-supported initiatives, but each has wound down at project close because the recurrent costs were not absorbed into a host agency's budget.

Third, the user is invisible in the bureaucratic frame. The boro farmer in Sunamganj, the mango farmer in Chapainawabganj, the potato farmer in Munshiganj, the hilsa fisher in Bhola, are all clients of one agency or another but not of a coordinated agromet service. A user's success is no agency's KPI. When the haor floods in 2017 or Sitrang lands in 2022, no agency is held to account for the absence of an upazila-crop-stage advisory, because no agency's mandate document says it should have produced one.

What would close the gap

The reform proposal flowing from this analysis is structural, not technical. The science is settled. The institutions exist. What is needed is a coordinating cell with a mandate to bridge them.

A National Agromet Cell, jointly staffed by BMD, DAE, and BARC, with a director who reports to a committee chaired by a senior official with cross-ministry authority, can in principle take the ECMWF and BMD outputs, run the crop-stage join against the BBS and DAE crop-block registry, generate operational Bangla advisories, and route them through DAE's existing union-level supervisor network and through SMS gateways that already serve government channels. The technical components are off-the-shelf. The personnel exist within the three agencies. What is missing is the cell that holds the workflow together and the budget head that pays its recurring costs.

A pilot can be defended on cost-benefit grounds without overstating the numbers. Boro alone produces approximately nineteen to twenty million metric tonnes of paddy a year and accounts for over half of Bangladesh's domestic rice supply. A single avoided heat-sterility event of even one percent of national production, well within the documented range, is six figures of metric tonnes of rice; the recurring cost of an Agromet Cell of fifty staff with operational budget would be a fraction of that loss in a single typical year. The same arithmetic holds for the mango belt during pre-monsoon kalbaishakhi season and for hilsa fishing-vessel safety during cyclone season.

Five concrete moves would put the architecture in place.

First, a memorandum of understanding among BMD, DAE, and BARC, with a Cabinet directive making the National Agromet Cell a coordinating body with line authority across the three agencies for agromet outputs.

Second, an interoperable data layer that allows BMD's forecast outputs and BWDB's flood model to flow directly into a DAE-managed crop-block database. The Smart Bangladesh data architecture has the components; what is needed is the agromet schema and the pipeline.

Third, an SMS-and-IVR delivery channel built on top of the existing mobile financial services penetration. Bangladesh has more than seventy-five million mobile financial accounts; the addressing infrastructure already exists. Pushing a Bangla advisory to a farmer's registered handset is a question of pipeline plumbing, not a question of network coverage.

Fourth, a transparent threshold registry, owned jointly by BRRI, BARI, BFRI, and DAE, that lists every crop-stage-hazard combination, the threshold value, and the source. The /early-warning demonstration shows what this looks like in operational form. The registry should be public, updatable on a quarterly cycle, and citable in extension material.

Fifth, a verification and skill-score system, owned by BMD's research wing and audited annually, that compares each year's advisories to the actual weather and crop outcomes. Without this, no agency can answer the question of whether the system is improving over time.

What this is and what it is not

The /early-warning page is a research demonstration. It is not, and was not built to be, a substitute for BMD or DAE. The disclaimer on the page itself is explicit: for operational warnings, follow BMD bulletins and the local DAE office. The demonstration's purpose is to make the gap legible. When a Bangla policy-conversation participant says we need better warning systems for farmers, the demonstration is the answer to the implicit question of what better looks like, in concrete operational form, on real forecast data, citing real institutional thresholds.

This is the BDPolicyLab argument: the gap between the forecast on a server in Reading and the field in Sunamganj is fixable, the components exist, and the obstacle is not science. It is mandate, coordination, and a budget head that nobody has yet written into the next five-year plan. The next time a haor floods or a kalbaishakhi flattens a mango belt, the question worth asking is not whether the forecast existed. It did. The question is why, nine years after 2017, the bridge has still not been built.

Sources

• Bangladesh Agro-Meteorological Information Service (BAMIS), the national agromet platform; agromet forecasts and advisories
• BMD Agromet Forecast, Bangladesh Meteorological Department live operational portal
• WMO: Agrometeorological Information for Climate-Resilient Agriculture in Bangladesh, institutional framework and programme assessment
• The Daily Star: Haor flood impact on Bangladesh rice production 2026, 2026 season damage estimates
• The Business Standard: Floodwaters swallow haor harvest in Sunamganj, field reporting on 2026 losses
• Dhaka Tribune: Haor harvest in peril, 7,000 hectares submerged in Kishoreganj, district-level loss data
• Weather Impact: Bangladesh Water for Food Programme, S2S forecasting project with DAE and BMD
• ECMWF: Forecasts, primary global forecast data source underlying BMD and regional models