A weed detection corpus is not built by collecting images and then having them annotated. The reverse sequence produces most of this field’s failures, and the seasonal constraint does not forgive scoping errors.This article sets out the method, in order, with what each stage determines. It extends the article on detection at early stages.
Why order matters more here than elsewhere
Three constraints make rework expensive in weed detection.The acquisition window is seasonal, which means a design error is paid for in a year rather than a few weeks.Conditions do not recur, one field at the same stage presenting a different flora, weather and soil state the following year.And the corpus does not transpose, which the preceding articles established, precluding compensating a gap with images from elsewhere.One important practical consequence follows. Those three constraints make the design phase, which produces no images, the most decisive stage of a weed detection project.Stage one: establishing the scope
Five decisions bound a weed detection corpus.They are taken with the end user.The crops concerned, each of which constitutes a distinct context requiring its own coverage.The regions concerned, whose flora and soil conditions differ.The stage window, determined by the use and by the destruction method.The task formulation, binary, by functional group or by species.And the acquisition modality, mounted, drone or hand-held, which do not produce interchangeable corpora.One observation follows. Those five decisions multiply rather than add, two crops across two regions and three stages producing twelve contexts to cover, which makes reducing the scope the most powerful cost lever available.Stage two: the pilot batch
This restricted first batch determines everything that follows.It is not skipped.A few hundred images acquired under real conditions suffice to establish what no meeting can.Four decisive pieces of information follow.The feasibility of the formulation adopted, measured by agreement between two annotators.The list of real edge cases, which feeds the reference and which no anticipation produces exhaustively.The annotation throughput observed, which replaces any estimate with a fact and serves as the basis of every subsequent quotation.And the effective quality of the images, blur or unsuitable exposure being correctable at this stage and not afterwards.One practical consequence follows for a weed detection project. This batch represents a modest fraction of the final corpus and it determines most of the conventions, which justifies devoting to it a care out of proportion to its size.Stage three: the sampling plan
Explicit design replaces opportunistic collection.It changes the value of a weed detection corpus.A corpus built as outings allow reflects the conditions of those outings rather than those of the use.Six dimensions must be covered deliberately.Stages, spread across the whole window rather than concentrated on the most convenient one.Infestation levels, from clean fields to dense mat zones.Lighting conditions, from full sun to overcast, including low-angle hours.Soil states, dry, wet, crusted, with and without residue.Target species, including rare ones, whose presence must be sought rather than awaited.And atypical configurations, headlands, gaps and weak plants.One important observation follows. The fifth dimension requires targeted selection rather than a random draw, which the article on problem species set out, and that decision belongs to prior design rather than to later filtering.Sizing
Four factors determine the volume needed.They do not reduce to a number of images.The number of contexts to cover, the product of the crops, regions and stages adopted.The number of classes, each having to reach a sufficient count to be learned.The internal variability of each class, a species changing appearance strongly requiring more examples.And the difficulty of the task, a fine distinction demanding more examples than an obvious one.One practical consequence follows. The useful volume is reasoned in objects per class and per context rather than in a number of images, an image of dense infestation potentially containing fifty objects and an image of a clean field none.One observation completes that. Counting objects per class on the pilot batch supplies the only serious sizing basis, and its absence explains the considerable gaps between announced and necessary volume in weed detection projects.Three sizing errors
Three lines of reasoning produce wrong volumes and they recur regularly.Transposing an order of magnitude from another vision domain, ordinary object detection presenting neither the visual proximity nor the variability of this field.Counting in images without considering object density, two corpora of the same size potentially differing by a factor of ten in useful content.And sizing on the majority class, when it is the rare classes that determine the volume to be reviewed in order to reach them.One observation follows. The third line of reasoning is the most expensive: reaching a sufficient count on an infrequent species requires handling a volume of images unrelated to what the majority class would have demanded, and that factor is not guessed without prior counting.What to record with the images
Six pieces of metadata condition the corpus’s exploitation and they cost a few fields at capture.Date and time, which permit reconstructing lighting conditions.The field, indispensable to a correct geographic separation between training and evaluation.The crop and the stage, expressed in observable markers.Acquisition conditions, equipment, height and speed.Soil state and the presence of residue.And the cropping history where available, notably prior treatments that explain the flora present.One important practical consequence follows. The second is the one whose absence is most damaging: with no field identifier, it becomes impossible to guarantee that an evaluation set is genuinely separated, and an announced performance becomes unverifiable.Exploiting an existing corpus
One frequent situation deserves treating, a client often holding images accumulated before any structured project.Those images have real value and often less value than is attributed to them.Four verifications establish what they permit.Context coverage, by crossing the available images with the target scope, an exercise that almost always reveals gaps.The presence of metadata, whose absence strongly limits exploitation even where the images are good.Acquisition homogeneity, a collection built with several devices over several years mixing incompatible resolutions and viewpoints.And representativeness of conditions, a collection built during demonstrations or visits systematically presenting good conditions.One practical consequence follows for a weed detection project. That fourth point is the commonest: images accumulated during commercial outings describe showcase fields rather than ordinary ones, and a corpus built on them will produce a system that disappoints on the real estate.Separating training from evaluation
One rule governs that separation.Breaching it produces flattering and false figures.Two images of the same field taken the same day share too much context to sit on either side.Three levels of separation exist, and their demands increase.Separation by image, insufficient, two consecutive images of one pass often showing the same plants.Separation by field, correct for most uses, which guarantees the system is evaluated on an unseen context.And separation by region and by season, the most demanding, which measures what the system will actually do at a new client.One observation follows. The third level produces the lowest and most useful figure, since it corresponds to the real commercial situation, and it is the one a provider has an interest in measuring before a client measures it for them.The evaluation corpus, apart
One distinction deserves stating because these two corpora do not follow the same rules.The training corpus must cover what the system will meet; the evaluation corpus must measure what it is worth.Four differences follow.Evaluation requires a far smaller volume, a few hundred objects per class sufficing to measure where training demands far more.It requires higher annotation quality, an error in the reference distorting the measurement where an error in training dilutes.It must over-represent difficult configurations, the objective being to measure where the system is tested rather than to reflect the real distribution.And it must stay stable over time, an evolving evaluation set precluding comparison between two versions of the system.One important observation follows. The second difference justifies having the evaluation corpus annotated twice and disagreements arbitrated by an agronomist, an effort disproportionate on training and fully justified here.Progressive construction
An organisation in successive batches is better than a mass acquisition.It deserves defending to the client.A corpus built in one block fixes conventions before they are tested.Four advantages accompany the progressive approach.Conventions are corrected between two batches rather than after the whole.Throughput and cost are measured early and projected.A first model trained on the initial batches permits pre-annotating the following ones.And the gaps that first model reveals direct the content of the next batch.One practical consequence follows for a weed detection project. That fourth advantage is the most important: it turns corpus building into a loop rather than a straight line, and it concentrates effort on what is actually missing rather than on what is easy to acquire.Organising an acquisition campaign
Six logistical decisions condition the success of an outing and they are prepared before the window.Scouting the fields, ensuring they cover the target contexts and that access is authorised.The visit calendar, set on stages rather than on availability, with several passes over the same fields.The capture protocol, height, angle, speed and settings, recorded so as to be reproducible from one outing to the next.Metadata capture, organised to be filled in during the outing and not reconstituted afterwards.The presence of an agronomist on at least one outing, which permits identifying in the field what the image will not show.And checking the images the same day, an unsuitable exposure or a focus defect being correctable while still on site.One important practical consequence follows for a weed detection project. The last point is the most profitable on the list: an outing lost for want of an immediate check is not recovered before the following stage, and sometimes not before the following year.What not to do with the corpus
Four practices degrade a weed detection corpus.They nonetheless appear reasonable.Discarding poor quality images, which will be present in production and whose absence produces a fragile system.Artificially balancing classes by removing examples of the majority class, which destroys information instead of adding any.Cropping images around objects, which removes the context the model needs and distorts the distribution.And artificially inflating volume through transformations before covering the real contexts, which multiplies examples without adding information.One important observation follows. Those four practices share a common logic, they make the corpus cleaner and less representative, which matches the finding every preceding series established on the difference between a tidy corpus and a useful one.Public data and its use
Several weed image sets are freely distributed.Their place in a weed detection project deserves situating.They do not replace a local corpus and they render three real services.Training annotators, an annotated public set permitting a team to be trained before client images are available.Pre-training a model, whose general layers benefit from exposure to a volume the local corpus will not reach.And prior feasibility estimation, a trial on public data giving an order of magnitude before any commitment.Three limits bound their use.Acquisition conditions differ, resolution, height and equipment not being those of the project.The flora differs, the available sets covering certain regions and crops and not others.And annotation conventions differ, which makes merging with a local corpus delicate or even counterproductive.One practical consequence follows for a weed detection project. Pre-training is the most profitable use of these sets, and direct merging with a local corpus is the one that most often produces a degradation rather than a gain.Documenting the corpus
Six elements accompany a delivery and their absence strongly limits reuse.The nomenclature document, with its classes, thresholds and illustrated case library.The sampling plan actually carried out, distinct from the one that was planned.The counts per class and per context, a simple table that immediately shows what the corpus permits.The quality report, with the control sampling method and the agreement rates measured.The list of contexts not covered, which explicitly bounds the domain of use.And the definition of the training and evaluation sets, with the separation criterion adopted.One important observation follows. The fifth element is the rarest and the most useful: declaring what the corpus does not cover turns an invisible limit into a known one, and that protects the provider as much as the client, a principle every preceding series established.The corpus as the client’s asset
One contractual question deserves asking at scoping since it settles badly afterwards.An annotated corpus is an asset whose ownership, reuse and retention must be established.Four points must be settled.Ownership of the images and of the annotations, which may differ, the client often supplying the first and the provider producing the second.The provider’s right of reuse, notably to improve models or train teams, which must be explicitly granted or refused.Retention after the project ends, a corpus destroyed at closure no longer serving a later extension.And handover of the case library, a document whose value for a future re-annotation often exceeds that of the corpus itself.One observation follows for a weed detection project. The second point produces the most misunderstandings, a client frequently assuming an exclusivity the contract does not mention, and clarifying it upfront is better than discovering it in a later discussion.Approaching corpus construction
Five questions scope this phase and they are asked before any field outing.How many distinct contexts does the scope contain. That product determines the volume and it is almost always underestimated.Is a pilot batch planned before production. Its absence makes any costing arbitrary.Does the sampling plan cover the six dimensions set out. An omitted dimension surfaces at evaluation.Will the field identifier be recorded. Its absence precludes correct separation.And will construction proceed in successive batches. A mass acquisition fixes untested conventions.Those five answers determine the volume, the calendar and the validity of the result. Asking them before the first outing avoids this field’s most expensive outcome, a complete and unusable corpus a whole season away from its correction.What this chapter teaches
One cross-cutting observation deserves closing this examination because it matches a finding of the preceding series.Building a corpus is an operation of design and not of collection.Three findings compose it.What determines a corpus’s value is what it deliberately covers, not the number of images it contains.That coverage is designed before acquisition, since an omitted dimension is not added retrospectively in a field with a seasonal window.And measurement on a pilot batch replaces any estimate, which makes that stage non-negotiable for a fixed-price engagement.That finding matches exactly the one the series on satellite image segmentation and aerial LiDAR established on the role of prior assessment: what costs days upfront avoids a factor of discrepancy in production, and the asymmetry is aggravated here by the agricultural calendar.What a season of data is actually worth
One valuation question comes up when a client weighs building a corpus against waiting, and it deserves a direct answer.A season’s acquisition costs field time, annotation and organisation, and it delivers something whose value is easy to understate.Three properties determine what that season is worth.It is not recoverable later. The stages, the flora and the conditions of that year existed once, and a project that skips a season does not catch up, it starts a year behind.It compounds with the next one. A second season adds inter-annual variability, which is the single dimension a first corpus cannot contain and the one that most determines whether a system generalises.And it decays slowly. Unlike much data, field images of a crop at a stage stay valid for years, since the plants have not changed and the acquisition conditions can be matched.One practical consequence follows for a weed detection project. The cheapest season to acquire is the current one, the argument for waiting is almost always weaker than it appears, and a client hesitating in spring is usually deciding between one season and two rather than between acting and not acting.Common mistakes
These failures recur often enough that naming them is usually enough to avoid them.- Collecting images before fixing the task formulation.
- Underestimating the number of distinct contexts the scope contains.
- Skipping the pilot batch and costing on estimate.
- Reasoning in numbers of images rather than in objects per class and per context.
- Building the corpus as outings allow rather than to a plan.
- Not recording the field identifier.
- Separating training from evaluation by image rather than by field.
- Discarding degraded images that will be present in production.
- Balancing classes by removing examples.
- Acquiring in bulk before testing the conventions.
- Awaiting rare species instead of deliberately seeking them.
