The eleven preceding articles in this cluster flagged errors as their subjects arose. This one gathers them and classifies them by when they occur, a method permitting a difficulty to be detected where it appears rather than at delivery.One observation runs through this inventory: the seasonal window makes correction far more expensive here than in other vision fields. It extends the article on sensors.
Why these errors cost more here
Three properties aggravate correction in weed detection.The acquisition window is seasonal, which means a design error is paid for in a year rather than a few weeks.Ground truth belongs to botanical judgement, which makes rework dependent on an agronomist’s availability.And the corpus does not transpose between crops or regions, which precludes compensating a gap with images from elsewhere.One important practical consequence follows. Those three properties mean the asymmetry between prevention cost and correction cost, already strong in other fields, reaches its maximum here.Scoping errors
This family occurs before any image.It is the most expensive in a weed detection project.Adopting species identification where the binary distinction sufficed, which multiplies the corpus cost without improving the decision.Aiming at both earliness and species identification, two partly incompatible objectives at the cotyledon stage.Underestimating the number of contexts, the product of crops, regions and stages almost always being larger than anticipated.And committing to a project with no agronomic competence available, a disqualifying obstacle rather than a difficulty to work around.Nomenclature errors
This family determines the ceiling of everything that follows in weed detection.Building the nomenclature from a botanical reference rather than from the decisions the system will trigger.Omitting the plant to protect class, covers, margins and intercrops.Omitting the undetermined object class, which forces annotators to settle at random.Omitting the unusable zone class, which produces a silent omission.And fixing the nomenclature before counting the numbers per class on a first batch.Convention errors
This family concentrates most of the real variability between annotators.Not writing the convention for handling weed mats, the principal source of divergence.Not fixing the window of stages covered and excluded.Neglecting the boundary between crop and volunteer from the previous crop.Illustrating a reference without showing the confusions associated with each species.And not stating what to do when no criterion decides.Corpus errors
This family concerns what the images cover.It is poorly corrected after the season.Collecting before fixing the task formulation.Building the corpus as outings allow rather than to a sampling plan.Acquiring every image under the same lighting and soil conditions.Discarding degraded images that will be present in production.Awaiting rare species instead of deliberately seeking them.And not recording the field identifier, which precludes a correct separation between training and evaluation.Production errors
This family belongs to execution.It affects consistency as much as throughput.Using annotators with no botanical training on a species nomenclature.Presenting the complete nomenclature at once rather than progressively.Tolerating rather than encouraging recourse to the undetermined class.Annotating a plant together with its cast shadow.And letting annotators discard difficult images for want of an explicit instruction.Control errors
This family produces results that look good without being so.Relying on a global rate of correctly annotated objects.Not measuring disagreement between annotators before committing to a requirement.Conflating agreement with correctness, two annotators being able to be wrong together.Reading a fall in the undetermined rate as progress.Neglecting the measurement of omissions, which requires exhaustive re-annotation.And omitting the constant control sample that alone reveals drift over time.Use errors
This family concerns how the corpus meets what exploits it in weed detection.Reusing a parameter setting designed for spraying on a mechanical weeding implement.Annotating to a precision the section treatment regime did not require.Annotating the weeds where the implement required recognising the crop.Neglecting atypical crop plants, weak or lodged, in a use with irreversible action.And enriching the sensor without correspondingly enriching the corpus.Durability errors
This family occurs after delivery.It is the least anticipated of all.Not planning a campaign for the following season, the only source of inter-annual variability.Omitting the operating line from the budget.Leaving knowledge of the conventions in one person’s head.Not retaining the raw images, which precludes any re-exploitation with a different pre-treatment.And not settling at scoping the corpus ownership and the right of reuse.Errors specific to this field
Five errors have no equivalent in other vision fields and they surprise teams coming from industrial work.Treating botanical variability as the principal difficulty, when the variability of conditions often exceeds it in amplitude.Assuming an annotator can identify a species without training, a transfer that fails where recognising an industrial defect succeeded.Planning an acquisition without accounting for the agronomic window, a calendar constraint absent from every other field.Neglecting that the target and the object to preserve belong to the same kingdom, a proximity with no equivalent in detecting manufactured objects.And expecting a single season to produce a representative corpus, when inter-annual variability is structurally absent from it.One observation follows for a weed detection provider. Those five errors are committed almost systematically by a team approaching agriculture for the first time, which makes them the priority content of a project preparation.The comparative cost of the families
Not all of them cost the same in weed detection.This economic reading clarifies the priorities.A scoping error costs the entire project, nothing produced serving.A nomenclature error costs the corpus, whose rework is the dominant line.A convention error costs a partial re-annotation, more bearable and nonetheless heavy.A corpus error costs a season, a calendar constraint no means lifts.A production error costs a batch rework, generally contained.And a durability error costs reconstitution, more than the initial build since the context has been lost.That gradation leads to a practical conclusion. Prevention effort is most efficiently placed on the first two families, where it costs days and avoids months.The common structure of these errors
One cross-cutting observation explains their persistence despite being avoidable.Three properties characterise them all.They are committed by omission rather than by action, a decision not taken being less visible than a wrong one.They produce no immediate effect, their consequences appearing weeks or a season later.And correcting them produces no visible deliverable, which makes it hard to justify to a sponsor awaiting annotated images.That asymmetry between prevention cost and correction cost is the dominant characteristic of this inventory, a written mat convention costing an hour and its reconstitution after production costing the corpus.Warning signals
Six signs indicate a project is starting badly and they are visible early.Nobody can name the decision the system will trigger.The species list comes from a regional botanical reference rather than from the client’s records.No agronomist features in the arrangement.The weed mat convention appears nowhere.The schedule contains no pilot batch.And no later campaign features in the budget.One observation follows. Those six signs are verified by reading a specification, they require no agronomic competence, and their presence predicts a difficulty better than any analysis of the technical arrangement adopted.The errors that are not errors
Four situations are regularly treated as failures when they are structural.The undecidability of a seedling at the cotyledon stage, which belongs to biology and not to competence.Disagreement between annotators on ambiguous cases, expected and constituting information rather than a defect.Lower performance on rare classes, structural and not correctable by better work.And the absence of generalisation to an uncovered region, a property of the field rather than a corpus weakness.One practical consequence follows for a weed detection project. Distinguishing an error from a limit avoids spending effort correcting what cannot be corrected, and that distinction belongs at diagnosis rather than after several unsuccessful attempts.The prevention arrangement
Six measures address the majority of the weed detection errors listed.Their total cost stays modest.Scoping by the decision, by questioning the end user on what the system will trigger.The pilot batch on real images, which establishes feasibility, throughput and disagreement.The nomenclature built with an agronomist on that batch, with an illustrated reference and written conventions.The sampling plan covering stages, conditions and rare species.Measuring disagreement before any contractual commitment.And documentation written during production, reference, record of decisions and quality report.Reworking a project in difficulty
One frequent situation deserves treating, a provider being regularly approached after unsatisfactory first work.Four questions determine what can be salvaged.Which family does the defect belong to, a production error being corrected by batch rework and a scoping error requiring starting again.Are the conventions used documented, their presence permitting identification of what must be reworked and their absence requiring reconstitution by sampling.Have the raw images and the metadata been retained, their absence making any re-exploitation and sometimes any diagnosis impossible.And is the following season reachable in the client’s calendar, a question determining whether a corpus gap is correctable or definitive.One important observation follows for a weed detection provider. The fourth question is specific to this field and often decisive, a gap identified in October not being filled before the following spring whatever resources are committed.Auditing a project under way
Four questions suffice.They establish in an hour where a project stands, from a team met for the first time.Ask to see the weed mat convention. Its absence indicates the project has not addressed its principal source of divergence.Ask how disagreement between annotators was measured. A figure indicates a grounded commitment, an impression indicates an empty promise.Ask which agronomist is involved and how often. A vague answer caps the achievable quality.And ask what the sampling plan covers in conditions. An answer limited to stages signals the commonest gap.One observation follows. Those four answers characterise a project more reliably than any performance indicator, and they require no access to the data.Why these errors persist
One cross-cutting observation deserves stating because known errors keep occurring.Three mechanisms sustain them beyond ignorance.The pressure of the agronomic window favours shortcuts, a project launched late preferring to acquire fast rather than design well, an understandable and almost always losing trade-off.The information asymmetry prevents a client from distinguishing a rigorous corpus from a fast one before the consequences appear, a situation aggravated here by a season’s delay.And the separation of actors disperses the decisions, the agronomist, the engineer, the implement manufacturer and the annotation provider each intervening on a part with nobody holding the whole.That third mechanism is the most specific to this field. It explains why errors obvious to each actor taken separately nonetheless run through a project, for want of a moment when the complete chain is examined by someone.One consequence follows for a weed detection provider. Occupying that moment, by asking the end-to-end questions at scoping, constitutes a value distinct from annotation itself and one few actors offer.Setting these errors out to a client
Naming possible failures looks counterproductive and is not in weed detection.Three reasons justify that transparency.An informed client asks better questions, which produces a better scoped project and fewer disputes at delivery.The field is young and most clients have no experience of its failure modes, which makes the information genuinely useful rather than redundant.And a difficulty announced before it occurs is handled jointly, where the same difficulty discovered at delivery becomes a reproach.One caveat accompanies that recommendation. Setting out the errors without setting out the arrangement that addresses them installs an anxiety without answering it, which argues for presenting both together systematically.A checklist before production
Ten points condense this inventory into a check applicable before launching a weed detection production.The decision the system will trigger is named.The formulation adopted is the simplest compatible with that decision.The species list comes from the client’s records crossed with the images.The plant to protect class features in the nomenclature.The undetermined and unusable zone classes feature as well.The weed mat convention is written.The window of stages covered and excluded is documented.The sampling plan covers conditions as much as stages.The field identifier will be recorded.And disagreement between annotators has been measured before commitment.A project ticking those ten has addressed most of the causes listed. A project ticking fewer than half will probably meet several of the families set out, and it is better to know that before committing a season.The three things to fix first
A closing simplification reduces this inventory to what a team should address before anything else.Write the weed mat convention. It costs an hour, it addresses the first source of divergence between annotators, and it cannot be reconstituted after production.Measure disagreement between two annotators before any commitment. It costs a day and it determines what can be promised.And record the field identifier at acquisition. It costs a field and its absence makes any announced performance unverifiable.Those three share a property. Each is cheap, each is available immediately, and each protects against a failure whose correction cost bears no relation to the prevention cost.Everything else in this inventory improves a project. These three determine whether it can be defended, corrected or extended, which places them in a different category.What this inventory offers a provider
Three commercial consequences emerge.The first is that a scoping engagement has demonstrable value. An interlocutor able to pose the feasibility questions before commitment spares their client a season whose cost far exceeds that of the scoping.The second is that the dominant families, scoping and nomenclature, belong precisely to an annotation provider’s competence domain, which makes their intervention more decisive than that of the team choosing the hardware.And the third is that these errors are little documented publicly, agricultural literature dealing mainly with models, which makes setting them out differentiating.That third consequence deserves emphasis. A provider able to name what makes a project fail demonstrates an experience no capability presentation replaces, and that demonstration costs one conversation.What this inventory teaches
One cross-cutting observation deserves closing this examination because it matches the preceding series’ finding.The errors listed are almost all non-technical.Three findings compose it.None follows from a shortfall in skill, the operations involved being simple to describe and to execute.Almost all follow from a decision not taken, a formulation not settled, a convention not written, a measurement not conducted.And their prevention is demonstrated through documents rather than assertions, which makes it verifiable by a third party.That finding matches exactly the one the series on satellite image segmentation and aerial LiDAR established in domains with very different constraints, which suggests a general property of corpus-building projects rather than an agricultural peculiarity.What the first project always costs
One observation deserves stating plainly, since it reframes several of the errors above as a predictable phase rather than as failures.A team’s first agricultural project bears costs the following ones do not.Four of those costs are unavoidable. Establishing the botanical conventions with an agronomist. Training annotators to recognise plants rather than objects. Learning the agronomic calendar and its consequences for planning. And discovering the failure modes this inventory lists, some of them by experiencing them.Two consequences follow for how that first project should be run.Scope it to one crop and one region, since the errors will be made and their cost scales with the breadth they affect.And capture what is learned deliberately, the conventions and the case library being the outputs that make the second project cheaper, and being lost by default if nobody is responsible for writing them.That second consequence is the one teams miss in weed detection. A first project completed without documented conventions has paid the entry cost and kept none of the benefit, which is the most expensive outcome available in a field where the next attempt is a season away.The most expensive errors
These failures recur often enough that naming them is usually enough to avoid them.- Adopting species identification where binary would have sufficed.
- Aiming simultaneously at earliness and species identification.
- Committing to a project with no agronomic competence available.
- Omitting the convention for handling weed mats.
- Omitting the plant to protect class.
- Collecting images before fixing the formulation.
- Not recording the field identifier.
- Not measuring disagreement between annotators before committing.
- Acquiring under a single lighting and soil condition.
- Omitting the following season’s campaign from the budget.
