Climate-smart agriculture · Pre-pilot concept
Bangladesh already holds the data. The gap is the last mile: turning registries, forecasts and satellite passes into advice a smallholder can act on this week, in Bangla, on the phone they own.
Film
A short introduction to the platform: where the data comes from, what the model does with it, and how the advice reaches a farmer holding a feature phone.
Hosted on YouTube. Playback starts only when you press play.
The gap
BAMIS, DAE-linked farmer registration and agro-meteorological services from BMD and related agencies are all in place. What smallholders still lack is hyper-local, crop-stage-specific guidance at the moment a decision is made.
Three things get in the way: forecasts resolved at district scale rather than field scale, landholdings fragmented across plots that behave differently, and data streams that never meet in a single system.
National strategies — the NAP, the updated NDC — already call for climate-smart agriculture. The implementation gap sits at the last mile.
The field GIS step closes the first gap: a DAE officer standing in the plot drops a pin, picks the crop from a standard Bangla list, and the entry syncs to a national geotagged crop layer.
Ground truth
Before proposing a new layer, we tested whether existing ones already do the job — screenshots below, not just a description of what we found.
Crop zoning · BARC dashboard
Real and running today, not a future promise — searchable by division, district and upazila. It answers "where can this crop grow." It does not answer "what is growing on this farmer's plot right now."
Open full size →
Land records · Search
The land ministry's plot search is real and searchable. We ran it for a plot in Gopalganj to see what it actually returns.
Open full size →
Land records · Result
The record traces to a decades-old survey, classifies the plot as both farmland and urban residential, and returns no khatian (ownership) data at all. Nothing here links it to a living cultivator.
Open full size →Government registries record who a farmer is; they do not record what is growing in a given plot this season. DAE field officers close that gap directly, on foot.
01 · Open
A mobile-first web GIS tool, built for a DAE officer standing in the field.
02 · Pin
GPS-located on the spot — no relying on a 1985 survey line.
03 · Tag
Chosen from a standard Bangla list of fruits and vegetables.
04 · Sync
Links farmer, plot and current crop in one geotagged layer — for the first time.
Why the fields first
Farmers are not farming badly. They are farming without the hyper-local signal they need at the exact moments — sowing, irrigation, pest pressure — when a district-level forecast is not enough. Sources: Bangladesh Rice Research Institute; peer-reviewed rice yield-gap modelling
Architecture
Three layers, acting together as a single national decision brain: ingesting government data and returning one plot's advice, not a district-wide guess. Only the middle layer is quantum-inspired, and it never sits on the path between a farmer and their advice.
Classical statistical tools handle single-variable prediction well. They struggle with combinatorial search across spaces like this one — compute needed grows exponentially as variables are added, not linearly. That is the specific, narrow case the quantum-inspired layer is tested against.
Layer 1 · Classical
Ingests Sentinel-2 optical, Sentinel-1 radar for monsoon cloud cover, IMERG rainfall, BMD forecasts and BAMIS ground truth, joined to the DAE registry and the field crop layer.
Layer 2 · Quantum-inspired
Prediction: quantum kernel methods and QSVMs run on simulators to separate flood and drought risk zones. Optimisation: irrigation scheduling expressed as a QUBO and solved by annealing, tested against classical solvers.
Layer 3 · Classical
A distilled student model emits plain JSON to DAE dashboards, SMS gateways, the automated voice (IVR/OBD) platform and the Bangla app. Ordinary servers or cloud; no quantum dependency at runtime.
Not a general claim about quantum advantage — three specific, testable ones, each checked against a classical baseline before it's trusted.
Feature selection
Quantum kernels map soil, weather and satellite variables into a higher-dimensional space, looking for separability a classical SVM misses when identifying which of hundreds of variables actually predict crop failure in a specific upazila.
Constrained optimisation
Irrigation scheduling under shared pump capacity is a resource-allocation problem of the same family as the classic knapsack problem. Annealing explores the space directly rather than approximating it.
Small-data generalisation
Crop-disease niches are often data-scarce. Quantum-inspired teacher models sometimes generalise better there than data-hungry classical deep learning — sometimes; the pre-pilot tests it rather than assumes it.
The two formulations
The two pieces of maths the argument above actually rests on — named in full, not just referenced.
QUBO
Quadratic Unconstrained
Binary Optimization
This is the exact shape quantum annealers are built to minimise — the reason irrigation scheduling gets framed this way, rather than the other way round.
QSVM
Quantum Support
Vector Machine
The point: separating structure without classically building every one of those dimensions by hand.
Method
No quantum hardware is procured, and nothing a farmer depends on runs on a quantum machine. Quantum methods are used offline, during training, and only where they beat a strong classical baseline on the same data.
Which mix of satellite bands, soil variables and weather features actually predicts irrigation need in a given upazila? We use QAOA-style feature selection and tensor-network simulation on GPU servers to explore a space that grows exponentially with each variable added.
Quantum kernel methods act as offline teacher models where field data is scarce. Their behaviour is distilled into a small classical student model — the one that actually ships. Every teacher is scored against a strong classical baseline, and we publish the comparison either way.
The student model runs on ordinary servers and pushes plain-language advice through the channels farmers already use: SMS, automated voice calls for feature phones, and an app for smartphone users. Delivery stays with trusted DAE field officers.
SMS
কাল বৃষ্টি হবে, সার দিও না
Rain tomorrow — hold the fertiliserVoice call
১৫ দিন পর ফসল কাটবেন
Harvest in about 15 daysApp
ক্ষেতে মাটি ভেজা থাকতে সার দিন
Apply fertiliser while the soil is still wetOn honesty about the method: quantum advantage for these workloads is not proven. The pre-pilot is designed to find out, with a classical fallback that is already an improvement on current practice. If the quantum-inspired path adds nothing, that result gets written down too.
After 365 days
A working pre-pilot package, ready for independent evaluation — not a prototype demo.
Inputs
Nothing here starts with a new national survey. The pre-pilot reads existing registries and open Earth observation, subject to data-sharing agreements with each custodian.
See how it's phased in →The 365 days
The phases run in order because each one blocks the next: no data agreements, no dataset; no dataset, no model; no model, no field test.
Governance. MoUs with DAE, BMD and BWDB for data access; steering committee formed.
Infrastructure. Cloud sandbox (AWS Braket or IBM Quantum access) plus local GPU servers for tensor-network simulation.
Data audit. Cleaning 22,065 farmer records, verifying geotags, gathering historical yields.
Gate: signed data-sharing agreements and a clean golden dataset, v1.0.
Prototyping. Encoding soil moisture and NDVI into quantum feature maps; kernel methods tested against classical baselines.
Formulation. Irrigation scheduling written as a QUBO and solved for a 500-node network.
Backend. Middleware connecting model outputs to the dashboard, target latency under two seconds.
Gate: a documented benchmark result, positive or negative.
Deployment. Advisories to 2,000 farmers across two upazilas, delivered by DAE officers.
Controlled comparison. One group receives the new advisories; one continues with standard BAMIS advice.
Measurement. Water use, input cost and yield variance tracked across both groups.
Gate: a clean, analysable dataset from the field.
Analysis. Statistical comparison of the two groups, with effect sizes and confidence intervals.
Reporting. A readiness report for the Ministry of Agriculture, published in full.
Handover. A blueprint for national scaling, including integration with existing a2i infrastructure.
Gate: a scale, adjust or stop recommendation supported by the data.
Risk
| Risk | Likelihood | Impact | Response |
|---|---|---|---|
| No quantum speed-up found | Medium | High | Fall back to quantum-inspired classical methods — tensor networks — which already improve on the legacy pipeline. The negative result is published. |
| Data access denied or delayed | Medium | Critical | MoU signing is a Phase 1 blocker, not a parallel task. Open satellite and rainfall products serve as the fallback layer. |
| Farmers do not trust the advice | Low | Medium | Delivery through DAE officers farmers already know, voice messages in local dialect, and advice that names its reason. |
| Cloud and compute overrun | High | Medium | Simulation-first policy with hard budget caps; paid quantum backends reserved for final validation runs only. |
Cost posture: the pre-pilot is scoped well under one million USD because it procures no quantum hardware and builds on registries that already exist. Quantum-inspired workloads run on GPU servers; cloud quantum backends are used sparingly and by the hour.
National alignment
Governance
Sovereignty is a design constraint, not a policy paragraph added at the end.
Project roles
| Role | Responsibility |
|---|---|
| Project oversight | Strategic direction, and coordination with DAE, BMD and ministry counterparts. |
| Lead scientist, quantum | Designs the QML circuits and the tensor-network simulations; owns benchmark integrity. |
| Agri-domain expert | Validates every advisory rule, and translates model output into language a farmer trusts. |
| Data engineering lead | API integration with BAMIS and BMD; owns the data lake and pipeline uptime. |
| Field coordinator | Runs the pilot upazilas and the DAE officer feedback loop. |
The sceptic's questions
The fastest way to start is one upazila, one crop cycle, and a clear question about what a farmer should have done differently.
Immediate next steps
The ask is narrow: a small task force inside the Ministry of Agriculture to fast-track those MoUs. Everything else in this plan, the team can start on its own.
