The enabling data layer for AI in smallholder farming

This matters more than the successes, because the current thesis rests on it.

The platform held field location, field size and a field-mapping output in its data architecture, and the aircraft generated GPS and operational data on every job. The spatial raw material was there. What it never became was an object: field size was a number typed into a booking form rather than a measured geometry, nothing carried a persistent identifier, and no boundary was ever reconciled against another. The word “parcel” does not appear in the venture material.

That is the gap the current work closes, and it is a smaller gap than it looks. The service already flies over the field, already knows where it flew, and already records what it applied. Everything the parcel needs is being generated and discarded.

Two things changed, and together they invert the problem.

Field boundary data can now be moved. Through the XAG partnership we access boundary geometry from their platform, and we have built an extraction and internal mapping pipeline against it. Partnerships with local drone manufacturers mean this is not a single-vendor dependency.

This is the bridge between the two ventures. The earlier work proved the service, the operator network and the farmer relationship. The mapping pipeline turns the data that service was already generating into something that can persist. Device and service-generated boundary information becomes the foundation for a spatial object, which makes the real product question askable for the first time — what should the Digital Farm Parcel contain, and what should it enable?

Farmers respond to yield, not to digital tools. When shown what yield they could achieve, they engage. Nobody asks for a mapping platform, a farm record or a digital service. They ask what their land could produce.

Answering that takes more than interest. To tell a specific farmer what their specific field could yield, you need what that field did before, what was applied to it, and what happened as a result. That is a history, and a history needs something stable to attach to.

That is the opening this venture takes. Not a market gap in cadastral coverage, but a service we cannot deliver to farmers who already want it.

The common framing — “APAC needs cadastral digitisation” — is wrong, and the evidence is unambiguous about why. Three distinct layers are evolving independently.

Vietnam has been building a national Multi-Purpose Land Information System (MPLIS) as the platform for a National Land Database, financed in part through the a multilateral development bank's VILG project across 30 provinces. Vietnam holds more than 100 million land parcels.

In parallel, FAO — under a donor-funded project, implemented with ITC, University of Twente — developed an AI workflow to delineate crop field boundaries from freely available Sentinel-2 imagery, and tested the prototype in Cambodia and Vietnam. The associated open dataset, AI4SmallFarms, contains 439,001 manually digitised field polygons across 62 tiles in Vietnam and Cambodia.

This is the single strongest piece of evidence in the document. A land database is not a field layer. The gap between them is where this venture lives.

Indonesia's land registration programme has registered over 60 million parcels in eight years, created more than 38,000 jobs and increased the number of private surveying firms twelve-fold. World Bank financing alone accounted for over 8 million parcels, 42 percent registered to women. A follow-on project targets the land rights of more than 11 million people.

The World Bank has separately identified fragmented and incomplete land information as a constraint on land and natural resource governance in Indonesia, which is the problem the One Map programme was created to address.

Proposing cadastral mapping to Indonesia would be close to insulting. The opportunity is making existing spatial infrastructure answer agricultural questions it was never built to answer.

The World Bank-financed SPLIT project subdivides collective land ownership awards into individual titles. As of July 2025, 163,176 individual electronic titles had been issued, 84 percent bearing female names. On completion around 750,000 agrarian reform beneficiaries are expected to gain individual rights over more than 1.3 million hectares.

Project implementation reporting notes cadastral surveys were available for only 20 percent of collective CLOAs, with roughly a further one million hectares still requiring survey.

This is structurally different from Indonesia. Here an incomplete cadastral problem sits directly inside the agricultural system, which changes both the partner set and the sequencing.

Of an estimated 3.5 million land parcels in Lao PDR, around 1.5 million had been registered and titled as of 2021, mostly in towns and cities, against a government target of a further 1.2 million titles. a multilateral development bank reporting shows titles or use rights in the Lao LandReg digital system rising from 482,233 at baseline to 1,203,331 by September 2023, with the system operational in 32 districts against a target of 100.

As of October 2025, Lao LandReg rollout was active in 8 provinces with 10 more planned by June 2026, over 430,000 paper land files had been scanned in Vientiane with 95,000 entries digitised, and a decision had been taken to centralise the system at the government data centre.

Laos is the cleanest illustration of the two-speed problem: land digitisation is actively progressing while agricultural parcel measurement is still being piloted. Small market, high development-partner interest, useful as a proof site rather than a commercial target.

The Digital India Land Records Modernization Programme (DILRMP), a central sector scheme launched in 2016 and extended to 2025-26, assigns each land parcel a 14-digit Unique Land Parcel Identification Number (ULPIN, or Bhu-Aadhaar) based on the geo-coordinates of the parcel's vertices. As of November 2025, more than 360 million land parcels across 29 states and union territories had received ULPINs, and ULPINs had been generated for roughly 66 percent of agricultural land parcels. Computerisation of Records of Rights was complete in 625,137 of 657,397 villages as of December 2023.

Land remains a state subject under the Indian constitution, and implementation varies substantially by state.

India does not have a missing-identifier problem. It has a spatial accuracy, interoperability, currency and state-variance problem. That makes it a platform and integration market, not a mapping market, and it is large enough that it should not be treated as one country for go-to-market purposes.

The ratings below are my synthesis. There is no published maturity index for this and I am not presenting one. Each cell is a judgement about what a venture would encounter, not a measurement.

Vietnam is the launch market. It has the demonstrated technical gap, the existing venture history, the XAG and local manufacturer relationships, and — decisively — a national policy programme that creates institutional demand for exactly this object. That last point is section 05.

Vietnamese family farms average around 0.4 hectares. In the Mekong Delta the average area per parcel has been measured at 0.14 hectares, with over 80 percent of farmers holding under one hectare. Nationally there are roughly 75 million land parcels, averaging seven to eight plots per farm household. Vietnam's agricultural land endowment of around 0.3 hectares per person is among the lowest in the world.

Research on field boundary delineation describes the Vietnam and Cambodia landscapes as highly fragmented, with sub-hectare paddies separated by narrow earth dykes, and notes that high cloud frequency forces reliance on sparse clear-sky observations.

The AI4SmallFarms work notes that Sentinel-2's 10-metre spatial resolution poses challenges given the prevalence of small fields, and its published results show higher-resolution imagery capturing small fields that Sentinel-2 misses.

This is the crux and it should be stated plainly. The question is not whether satellites can see agriculture. They obviously can. The question is whether pixels can be turned into reliable individual field polygons at the scale of fragmented smallholder farming — and then whether those polygons can be attached to a farmer, a crop, a history and an economic activity.

If remote sensing alone could produce trustworthy sub-hectare field boundaries at scale, this venture would be a data product and someone would already have built it. It cannot, reliably, at the resolution that matters. A drone flying a treatment mission is physically present at ten to thirty centimetre resolution with RTK positioning. It resolves what a 10-metre pixel cannot.

So the argument for service-generated boundaries is not primarily commercial. It is technical. The service is on the field, and being on the field is what produces geometry good enough to be worth keeping.

Earth observation still earns its place, but in three specific roles rather than as the primary boundary source.

DJI's SmartFarm platform automatically identifies farmland borders using AI to obtain the area of each parcel, and records the operation information of each parcel to create farmland management archives. It performs field scouting analysis, NDVI-based growth detection, prescription map generation and 3D route planning for orchard terrain.

DJI's own SmartFarm Web user guide states that field management is currently independent of field planning, that at present it is for area measurement and yield analysis, and that in future field management will be integrated with Agras operation data, yield data, seedling data, agricultural notes and pesticide records.

That second citation is worth reading twice. The largest agricultural drone manufacturer in the world documents that its boundary object and its operation history are not yet joined. The join is the product. It is stated as roadmap by the incumbent whose interest is in a single-vendor stack.

We have a partnership with XAG and access to boundary data from their platform, against which we have built an extraction pipeline and internal data mapping. We also hold partnerships with local drone manufacturers in the region.

The multi-vendor position is what makes this defensible rather than parasitic. A venture whose only asset is a pipeline into one manufacturer's system is a feature that manufacturer can remove. A venture that normalises boundaries and history across XAG, local manufacturers and — where farmers permit — DJI, is doing something no manufacturer will do for a competitor's customer.

The clearest way to explain equipment-agnosticism to a non-technical audience is the cooperative that uses two vendors. Half its members' fields are described in one platform, half in another, and neither platform will ever hold the whole picture. The cooperative therefore cannot answer a question about its own land, despite every individual field being mapped somewhere.

Three things must be resolved before this data can be described as an asset in any external document.

Four consequences, and they compound.

The earlier venture had to manufacture demand through demo events and loyalty caps. This programme generates it. The venture's job is to be the infrastructure the programme is currently missing.

Whether cooperatives in the programme currently experience traceability as a painful unmet need, or whether commune-level verification is working adequately with paper and spreadsheets, is not established. Policy requiring something is not the same as an operator feeling the cost of it. This is the first thing to test and it is in section 11.

A Digital Farm Parcel is a persistent, uniquely identified representation of a piece of cultivated land, with four defining properties.

Three alternatives were considered and rejected.

The parcel is the smallest object that is stable over time, meaningful to a farmer, addressable by a machine, and the natural key for every downstream product.

Identity rules, chosen so the object survives the messiness of real land.

That last rule is the one people argue about, so the reason should be stated. A cultivated field and a legal parcel routinely disagree. Farmers cultivate across boundaries, rent adjacent land, leave margins uncultivated and subdivide informally. Forcing them into equivalence corrupts both.

Six states, with the transitions defined.

Every geometry enters the system with a source tier. The tier travels with the geometry permanently and governs what the geometry may be used for. Accuracy figures below are typical ranges from published instrument specifications, not measured results from our system — they must be replaced with measured values before external use.

The critical distinction is P3 against P4. A treatment footprint records where the drone sprayed, which is not where the field is. It omits headlands, margins, obstacles and any area the operator chose to skip. Treating them as interchangeable is the most likely source of quiet data corruption in this system, and it is the failure the XAG question in section 04 is trying to prevent.

When two geometries claim the same land, the system does not pick a winner and discard the loser. It records the conflict.

The reason for the adjudication queue rather than automatic resolution is the standard irony of automation. Resolving the easy conflicts automatically makes the residual ones rarer and harder, and they must surface somewhere a person is actually looking.

Every parcel carries a confidence score derived from source tier, age of last verification, number of independent confirmations, and disagreement between sources. Confidence is displayed anywhere the geometry is used, and it gates the products: a finance or MRV claim requires a floor that a service-planning query does not.

These two get collapsed into one another constantly, and the backcast separates them. One is a spatial problem that the parcel solves. The other is a labelling problem that the parcel does nothing for.

The four buildable applications all depend on the same two things, which are the parcel and the operation record attached to it. Neither of them is an intelligence product. The build order therefore puts the substrate first and the models second, which is the opposite of how these programmes are usually funded.

Whether an application that needs seasons of history can survive the wait is an open question, and it is carried into the validation plan rather than assumed away.

Farmers own their boundary data. We discuss boundaries with them directly and hold an anonymised derivative in the system for aggregate questions such as crop and yield distribution.

That instinct is right and needs to be made structural rather than left as practice.

A vendor partnership grants us access to a vendor's system. It cannot grant us a farmer's consent, because the farmer's consent ran to the vendor. Moving a farmer's geometry into a new platform with new purposes requires its own basis obtained from the farmer.

Anonymisation does not solve this. A field boundary plus a commune resolves to a household in most of the Mekong Delta, where holdings average 0.4 hectares and neighbours know each other's land. Boundary geometry should be treated as identifying by default. The anonymised aggregate layer is legitimate for crop and yield distribution questions; it is not a route around consent for the geometry itself.

The consent conversation is placed at first field contact, conducted by the operator, in person, and it is the same conversation as the yield conversation. This is a design decision with a commercial rationale: the farmer is being told what their land could produce, and the reason we can tell them is that we will keep a record of this field. The permission and the value proposition are the same sentence.

Consent is recorded per parcel, per purpose, and is revocable. Revocation removes the farmer's identity link and the geometry from active use while retaining the operation record needed for the service provider's own accounting.

The architecture separates three things that are commonly and dangerously merged.

The commercial products sit almost entirely in the third band. Making that explicit early is what allows the first band to be genuinely farmer-owned without hollowing out the business model.

Eleven entities. Fields listed are the ones that carry design weight, not an exhaustive schema.

Five rules that govern the whole system.

The system is an event log with materialised views over it, rather than a mutable record store. The reasons are specific rather than architectural fashion.

Aggregate queries return results only above a minimum cell size, defined by parcel count and distinct-farmer count within an administrative unit. Below that, the query returns a suppressed result rather than a rounded one. Rounding leaks; suppression does not.

The data system above describes what the platform stores. What follows describes what a farmer or cooperative actually gets, because a data model is not a product.

This is not a dashboard and it is not a farm management system. It is a record of one field that a person can read on a phone in under a minute. Anything longer will not be read.

This is the product surface that opens every relationship, so it should be designed first.

The comparison set must be real and local. A national average is not persuasive to someone who knows their own delta. The published cooperative results in section 05 are the initial comparison basis, and they are strong precisely because they are field-level and nearby.

Whether a farmer acts on a yield comparison, or merely finds it interesting, is untested. Interest is observed. Behaviour change is not.

What a cooperative director gets, which is a different product from what a farmer gets.

The second and third items are the ones that make this a purchase rather than a nice-to-have. A cooperative enrolled in a programme with traceability requirements has an obligation it currently discharges on paper.

Derived, not separately maintained. Production area, variety, season, water management, fertiliser application, residue handling, all keyed to a parcel and drawn from operations already recorded. The farming diary problem from the earlier venture — where growers neglected a lengthy manual task that offtakers needed — is solved by generating the diary from operations rather than asking anyone to keep one.

Offtakers interviewed during the earlier venture said they would use a platform that automated farming diary entry for residue level and crop protection usage. That finding was never acted on. It is directly relevant now.

Reconstructed from the pilot and current operating knowledge, described as method rather than as a measured baseline.

The same job, with three additions that cost the operator almost nothing and change what the system knows.

The design constraint is that the operator's job must not get longer. Every additional step must be either automatic or under thirty seconds, because operators are paid by hectares covered and will abandon anything that costs them a job.

Drone operators refuse small isolated jobs because transport and setup dominate the economics of a single sub-hectare field.

This is why the cooperative is the entry unit rather than the individual farmer. A cooperative aggregates adjacent parcels into a single visit. Where the earlier venture proposed a smart scheduling tool to group nearby bookings after the fact, the cooperative provides the grouping structurally.

Illustrative arithmetic, on the pilot's own figures, to show the shape rather than to claim a result. At the pilot's rice spraying price of USD 7.9 per hectare, a single 1.3-hectare field yields around USD 10 of revenue against a trip. Twenty adjacent parcels averaging 1.3 hectares yield around USD 205 against one trip. The pilot material noted a provider with five drones could cover up to 50 hectares per day. Density is the entire economic argument.

Seven components, each with an owner and a failure mode.

Cooperatives enrolled in the One Million Hectare programme in the Mekong Delta, starting with the provinces where the pilot already scouted: Dong Thap, An Giang, Kien Giang, Long An, Tay Ninh.

That last point is a targeting signal, not stated demand. No cooperative has asked for this. What is observed is that farmers engage when shown yield potential.

Three parties, each of which must gain something they currently lack.

The earlier venture had already developed a scheme where large crop protection orders could earn a drone, and planned to support those customers in becoming service providers.

Four phases, each gated on the previous one's kill test passing.

Nothing about expansion, other countries or additional products belongs in the plan until Phase 3 passes. The earlier venture proposed an Indonesia pilot and a Thailand and Philippines launch structure before the Vietnam pilot had completed a growing season, and that sequencing is a mistake worth not repeating.

Four streams, in the order they can credibly be turned on.

Streams 3 and 4 are the ones that justify the infrastructure investment. Stream 1 pays for field presence. Anyone pitching this as a drone booking business has misread it.

Willingness to pay is untested at every level. The earlier venture tested whether farmers would pay a platform fee on top of spraying and never reached an answer.

Gate three kills the most candidates in this domain, and it should. The entire premise of this venture is that the data does not currently exist. Building AI capabilities that assume it does is the fastest way to fail.

These were considered seriously and killed. Keeping the reasons visible is the point.

Three bands, held apart, and this is the most consequential single decision in the system.

The test: the platform must remain correct and usable with the intelligence band switched off. Boundaries would come only from instruments, records would still assemble, compliance would still generate. Anything that fails this test is a dependency rather than a capability.

The reason this matters commercially as much as technically: a traceability claim that cannot be explained is worth nothing to a verifier, and a yield figure a farmer cannot interrogate is worth nothing to the farmer.

Seven surfaces. Each is named by the question it answers, because a surface that cannot be described as a question does not need to exist.

S6 is the surface that makes the rest defensible, and it is the one most likely to be cut for looking unglamorous. It should be built in the first phase, not the third.

Three horizons, sequenced against the programme calendar and the growing season rather than engineering preference. The Winter-Spring season is the natural pilot window, as it was in the pilot.

The shadow-mode discipline in Horizon 1 is deliberate. Every model in this system can be run against ground truth before anyone relies on it, because the drone is measuring the same thing the model is predicting. Very few domains offer that and it would be wasteful not to use it.

The open items above are not all the same kind of question, and they will not be settled by the same method. Three are measurement problems that a season of operations answers on its own. Three need people in a room who do not normally share one.

The next-stage discovery programme uses cross-disciplinary workshops to resolve the remaining spatial, data, device and adoption questions. Each session is convened around a single decision that cannot be made by one discipline alone.

The programme is part of the product strategy rather than a research overhead. A spatial object that has to satisfy an agronomist, a drone operator, a compliance verifier and a lender is a negotiation, and the workshop is where that negotiation happens before the schema hardens.

Six, ordered by how likely they are to kill the venture.

Six load-bearing assumptions. Each has a test, a pass condition and a stated consequence of failure. A validation plan that only describes success is a plan to rationalise whatever happens.

Order matters. Test one before building anything. Tests one and two together decide whether this venture is possible; the rest decide what it is worth.

The domain expertise in Vietnamese smallholder agriculture belongs to the growers, the cooperatives and the agronomists. The land administration programmes belong to governments and their development partners. The field boundary machine learning belongs to FAO, ITC Twente and the named researchers. The drone technology belongs to XAG, DJI and the local manufacturers.