The agronomic value of a detection falls with time. A weed treated at the cotyledon stage competes little with the crop; the same plant treated three weeks later has already consumed water, nutrients and light.Yet it is at the earliest stage that recognition is hardest. This article sets out that tension, its practical consequences and what a corpus must contain to handle it. It extends the article on robotic mechanical weeding.
Why the early stage matters
Four agronomic reasons justify intervening early.They compound.Competition begins at emergence, a weed established before the crop taking an advantage the rest of the season does not recover.The critical period of harmfulness, during which the presence of weeds determines yield loss, sits in the first weeks of the cycle.Destruction methods are more effective on young plants, a seedling being uprooted or destroyed with an energy unrelated to that a developed plant demands.And late intervention lets a share of the weeds set seed, which feeds the seed bank and carries the problem into following seasons.One important practical consequence follows for any weed detection project. Those four reasons converge the agronomic optimum onto a stage where detection reaches its maximum difficulty, a tension that structures everything after.Why recognition fails at the early stage
Five properties make a seedling hard to identify in weed detection.The distinctive characters are not yet formed, a species’ leaf morphology appearing after the cotyledon stage.Cotyledons resemble one another across neighbouring species, which makes the distinction impossible for part of the flora.Size is small against the resolution available, a seedling occupying only a few pixels in a drone image.Contrast with the soil is reduced where the seedling is small and the soil damp.And individual variability is at its maximum, two seedlings of the same species and the same age differing with sowing depth and soil structure.One important observation follows. Those five properties compound and none is correctable by a more performant model, which shifts the answer towards acquisition resolution and task formulation rather than towards architecture.The value curve and the difficulty curve
A simple argument clarifies the choice of stage.It deserves stating explicitly in any weed detection project.The value of an intervention falls with stage and detection reliability rises with it.Three readings emerge.The optimum is neither at the earliest nor at the latest, it sits where the product of the two is greatest.That point depends on the crop and flora pairing, a crop with a long critical period tolerating a later intervention.And it depends on the destruction method, an implement effective on developed plants relaxing the earliness constraint.One practical consequence follows for a weed detection project. That point is not determined in a meeting room: it is measured by evaluating detection performance per stage on a corpus covering the whole window, which presupposes annotating beyond the target stage in order to be able to arbitrate.What resolution imposes
This physical constraint bounds early detection.It is regularly underestimated.A seedling must occupy enough pixels to be detected, and more still to be identified.Three consequences follow.The drone at standard altitude does not suit the early stage, its ground resolution placing a seedling below the detection threshold.The camera mounted at low height becomes the reference modality, its proximity supplying the resolution needed.And forward speed finds itself doubly constrained, resolution demanding a short exposure and the small object size demanding careful processing.One observation follows. This constraint explains why early detection systems are predominantly implement-mounted rather than airborne, an orientation that determines the type of corpus to build and must be settled before any acquisition.Reformulating the task for the early stage
Three reformulations make weed detection tractable where recognition fails.Moving to binary, the crop against non-crop distinction remaining possible when the distinction between weed species no longer is.Recourse to geometry, position relative to the row substituting for morphology as the preceding articles set out.And reasoning by density, departure from the expected count signalling an infestation with no individual identification.One important observation runs across those three reformulations. All abandon species identification and retain the information the treatment needs, which makes them compatible with most weeding uses and incompatible with selective treatment by species.What the corpus must contain
Five requirements characterise a corpus intended for the early stage.Stage coverage must be fine, an interval of a few days between two stages producing distinct appearances a coarse corpus does not represent.Soil conditions must vary, moisture and texture strongly modifying contrast with the seedling.Crop residue must feature, it occupies the soil at the early stage and is the principal source of false detections.The undetermined object class must be widely used, a significant share of seedlings being genuinely undecidable.And the stage must be recorded as metadata rather than inferred from appearance, which permits evaluating performance per stage.One practical consequence follows for a weed detection project. The last requirement costs a field to fill in at acquisition and it conditions the trade-off set out above, which makes it disproportionately profitable.Naming the stage correctly
A question of vocabulary conditions a corpus’s consistency and it is often settled by default.A plant’s stage can be designated in several ways and they are not equivalent.Four conventions coexist in practice.Days after sowing or after emergence, convenient to record and unreliable, development speed depending strongly on temperature and moisture.The standardised phenological scale, precise and requiring competence to be applied correctly.The number of true leaves, a simple marker observable directly on the image and largely sufficient for most projects.And size in centimetres, measurable and uninformative about real development since two species of the same size can be at very different stages.One practical consequence follows for a weed detection project. The third convention offers the best compromise, it applies without advanced training, it is verified on the image itself, and it suffices to decompose an evaluation by stage, which is this metadata’s principal use.The residue problem
This configuration dominates false detections at the early stage.It deserves separate treatment.Residue from the previous crop covers the soil at the moment seedlings emerge.Four difficulties follow.Green residue fragments are confused with seedlings, notably after rain rehydrates them.Residue partially masks seedlings, which produces truncated objects hard to qualify.Its distribution is irregular, which makes the difficulty variable from one zone of the same field to another.And its quantity depends on the cultural practice, a direct-drilled field presenting far heavier cover than a ploughed one.One important observation follows. Those four difficulties make soil conservation systems the most demanding context for early detection, which is doubly problematic since these practices are advancing and since they also limit recourse to mechanical weeding.Staggered emergence
A characteristic of the weed flora complicates reasoning by stage and it deserves setting out.Weeds do not all emerge at the same time, one field presenting seedlings and developed plants simultaneously.Three consequences follow.One image contains several stages at once, which precludes attributing a stage to the image and requires carrying it at object level where that information is needed.A single pass treats only part of the population, those emerging later staying intact.And detection difficulty varies within one image, recent seedlings being far more delicate than the established plants around them.One important observation follows here. That heterogeneity argues for a strategy of repeated passes rather than a single optimally placed one, which matches the logic of mechanical weeding and sits poorly with that of a single chemical treatment.What annotation requires at the early stage
Five requirements distinguish this work from annotation at a developed stage.Zoom becomes the principal tool, a seedling occupying too few pixels to be qualified at the usual display scale.Throughput falls markedly, each object requiring examination where a developed plant is recognised at a glance.Visual fatigue sets in faster, which recommends shorter sessions and reinforced checking towards the end of one.Recourse to the undetermined class must be encouraged rather than tolerated, failing which annotators settle at random on a significant share of objects.And double annotation becomes more useful than elsewhere, disagreement being both more frequent and more informative at this stage.One practical consequence follows for a weed detection project. Those five requirements translate into a cost per object markedly higher than for a corpus at a developed stage, a gap that should feature in a costing rather than be discovered in production.The role of non-destructive ground truth
One method lifts part of the undecidability and it is little used.A seedling undetermined on the image can be identified by letting it develop or by collecting it.Three distinct protocols follow.Photographic tracking of one seedling at several dates, which permits attributing the species retrospectively to the early images once the plant is identifiable.Collection with precise positioning, which permits laboratory determination and supplies an incontestable ground truth.And physical marking of tracked seedlings, which guarantees finding the same individual between two passes.One important observation follows. The first protocol is the most economical and it produces exactly what this field lacks, a corpus of seedlings whose species is known by a means external to the image, which permits both training and measuring what a human observer could legitimately have said about it.One practical consequence follows for a weed detection project. That protocol requires field organisation over a season, it is not improvised, and a corpus so built has markedly higher value than one annotated by judgement on the image alone.Evaluating early detection
Four principles govern that evaluation.They differ from an ordinary weed detection evaluation.Performance must be decomposed by stage, an average over the window masking the collapse at the opening.The rate of undetermined objects must be reported, it measures the genuinely undecidable share and bounds what a model can reach.False detections on residue must be counted separately, they constitute a distinct failure mode calling for a distinct correction.And agreement between annotators must be measured per stage, its value at the earliest stage supplying the achievable performance bound.One important practical consequence follows. The last principle frequently produces an uncomfortable figure, and that is precisely its value: it permits refusing an unrealistic contractual requirement with a measured fact rather than with an opinion.Harmfulness thresholds
An agronomic notion qualifies the earliness requirement and it is rarely brought into technical discussions.Not every weed justifies intervention, there being a density threshold below which yield loss stays lower than the cost of treatment.Three consequences follow for a weed detection project.Exhaustive detection is not always the objective, a density estimate potentially sufficing to trigger or withhold an intervention.The threshold depends on the species, a highly competitive weed justifying intervention at far lower density.And it depends on the cost of treatment, cheap mechanical weeding lowering the threshold above which it becomes worthwhile.One interesting observation follows. This notion sometimes turns a problem of individual detection into one of density estimation, a formulation markedly more accessible at the early stage and one that should be examined before concluding infeasibility.What early detection does not permit
Four limits deserve stating.This objective is often stated without qualification.Species identification stays out of reach for a substantial share of the flora at the cotyledon stage, including for an expert.The acquisition window is very narrow, which complicates building a corpus and sometimes imposes several seasons.Weeds with staggered emergence partly escape, an early pass not meeting those that will emerge later.And detection under the canopy stays impossible, which is not limiting at the early stage and rapidly becomes so.That first limit is the most consequential. It means a project aiming at both earliness and species identification pursues two partly incompatible objectives, which should be established at scoping rather than observed after annotation.Two passes rather than one
One strategy answers this chapter’s central tension and it deserves proposing explicitly.Rather than seeking the single optimal stage, two interventions at different stages handle separately what each can do.Three arguments support it solidly.The first pass, very early, treats in binary with no identification, a formulation feasible at that stage and sufficient for most of the emerged population.The second pass, later, treats subsequent emergence and the first pass’s failures, with identification now possible.And each pass works within its domain of feasibility, which avoids sizing a single system on the hardest case.One practical consequence follows for a weed detection project. This strategy requires two corpora rather than one and it produces two systems simpler than the single one it replaces, a trade-off whose balance is often favourable and that is rarely examined.Approaching an early detection project
Five questions scope this type of project and they are asked in one meeting.What stage is targeted, expressed in observable markers rather than in days after sowing.Is species identification required at that stage. A positive answer requires verifying its feasibility before any commitment.What ground resolution does the acquisition supply. That answer determines whether the target stage is detectable.What is the residue cover on the fields concerned. That answer sizes the principal difficulty.And will the stage be recorded at acquisition. That answer conditions any per-stage evaluation.Those five answers determine feasibility, modality and corpus. Asking them before acquiring avoids this field’s commonest outcome, a corpus built at a stage where the stated objective was not achievable.The question to ask first
A single question settles the feasibility of an early detection project, and it is asked before any technical discussion.What will the system do with the information, at that precise stage.If the answer is treat uniformly, the binary distinction suffices and the project is feasible.If the answer is choose a product by species, feasibility must be verified on a sample before any commitment.And if the answer is decide whether to intervene across a zone, a density estimate suffices and the project becomes markedly simpler.That question precedes all the others. It determines whether the difficulty of the early stage is an obstacle or merely a resolution constraint.What this chapter teaches
One cross-cutting observation deserves closing this examination because it holds beyond this field.Early detection illustrates a configuration where value and feasibility vary in opposite directions.Three findings compose it.The optimum sits at neither extreme and it is not guessed, it is measured.That measurement presupposes a corpus wider than the objective, since the stages one will discard must be covered in order to justify discarding them.And measuring disagreement between annotators at the hardest stage supplies the only bound that can be set against an unrealistic requirement.That finding matches exactly the one the series on satellite image segmentation and aerial LiDAR established: the performance bound is not a property of the model but of an observer’s capacity to decide on the data available, and knowing it before committing protects both parties.What an expert can and cannot do
One calibration exercise deserves running early, because it settles an argument that otherwise runs for months.Show an agronomist a set of cotyledon-stage images and ask them to name the species.Three outcomes are informative and all three are useful.Where they name it confidently and correctly, the distinction exists in the image and a trained annotator can be brought to it.Where they hesitate but get there with effort, the distinction exists and will not survive production throughput, which argues for merging the classes.And where they decline, the distinction is not in the image, and no model, corpus or architecture will recover it.The third outcome is the one worth having on record. A client asking for species-level weed detection at the cotyledon stage is asking for something an expert cannot do on their own data, and demonstrating that with their own images and their own agronomist ends the discussion in an afternoon.One observation completes that. Running this exercise costs almost nothing and it reframes the provider from someone who might underdeliver into someone who established what was possible before promising it.Common mistakes
These failures recur often enough that naming them is usually enough to avoid them.- Aiming at species identification at the cotyledon stage.
- Using a drone at standard altitude for early detection.
- Expressing the target stage in days after sowing rather than in observable markers.
- Building a corpus without recording the stage at acquisition.
- Excluding residue-covered fields from the corpus.
- Not decomposing performance by stage at evaluation.
- Neglecting the undetermined object class at the earliest stage.
- Restricting the corpus to the target stage without covering neighbouring ones.
- Assuming the acquisition window permits completing a corpus in one season.
- Accepting a performance requirement without having measured annotator agreement at the stage concerned.
