Field conditions explain a substantial share of the gap between a system’s announced performance and its behaviour in the field. They belong neither to the model nor to the nomenclature, and they are handled at acquisition or not at all.This article sets out the three families of unfavourable conditions, what they produce and how a corpus takes them in. It extends the article on building a dataset.
Why these conditions matter so much
Three properties of field work explain their weight in weed detection.They are endured rather than chosen, a weeding operation taking place when the agronomic window dictates and not when the light is favourable.They vary within the same day, a field worked in the morning and in the afternoon presenting two distinct appearances.And they compound, wet soil under an overcast sky with residue producing an image in which each factor aggravates the others.One important practical consequence follows. Those three properties mean the variability of conditions often exceeds, in a weed detection project, the botanical variability the corpus nonetheless sets out to cover first.Light
This family produces the strongest variations in appearance.It is also the least controllable.Six distinct configurations occur in the field.Full sun, which produces hard cast shadows and overexposed areas on glossy leaves.Overcast sky, which gives homogeneous light and reduces contrast between plant and soil.Low-angle hours, which lengthen shadows to the point that a plant and its shadow form a composite object.Passing cloud, which changes exposure within seconds inside a single image series.Shadow cast by the machine itself, whose geometry depends on the hour and the direction of travel.And artificial lighting, used under a shroud to escape the previous five configurations.One important observation follows. The last configuration is the only genuinely effective answer, and it transforms the problem by requiring a corpus acquired under controlled lighting, incompatible with one built in natural light.Soil
This family modifies the background against which the plant stands out.It is often neglected in weed detection corpora.Five surface states produce distinct appearances.Dry pale soil, which offers the best contrast with a green seedling.Wet dark soil, which reduces that contrast and alters reflectance.Crusted soil, whose texture produces patterns segmentation methods confuse with plant structures.Cloddy soil, whose internal shadows create dark areas resembling gaps in foliage.And soils of atypical colour, red clays or very pale limestone, which shift the whole colour distribution.One practical consequence follows for a weed detection project. The last category explains part of the transfer failures between regions, a system trained on brown loam degrading markedly on red earth with no botanical difference involved.Residue
This family is the principal source of false detections.Its importance grows with the evolution of cultural practice.Four difficulties follow.Green fragments are confused with seedlings, notably after rain rehydrates them.Stems and straw produce elongated shapes models associate with grass leaves.Cover partly masks plants, which produces truncated objects hard to qualify.And the cover is irregular, which makes difficulty vary from one zone of the same field to another.One important observation follows. These difficulties increase with direct drilling and permanent covers, practices whose share is advancing, which makes this factor more important in future corpora than it was in historical ones.Rain and moisture
This secondary family affects the optics as much as the scene.It deserves flagging in any weed detection project.Four effects are clearly distinct.Droplets on leaves produce specular reflections that locally saturate the image.Condensation on the optics veils the whole image, a frequent effect in early morning.Mud splash soils the protective glass, a progressive degradation over the course of an operation.And moisture raises the reflectance of residue, which aggravates the confusion set out above.One practical consequence follows. The third effect is the only one that degrades within a single session, which produces a performance drop at the end of an operation that no evaluation on fixed images reveals.What these conditions produce
Four failure modes follow directly from them.They are diagnosed separately.False detections on non-plant elements, residue, clods and debris.Omissions through insufficient contrast, a seedling on wet soil going unnoticed.Delineation errors caused by shadows, an object annotated with its shadow producing an offset localisation.And instability between consecutive images, one plant detected then missed as exposure varies.One important observation follows for a weed detection project. The fourth mode is the most revealing and the least often measured: stable performance on a set of fixed images can mask strong instability on a sequence, and it is the sequence that corresponds to real use.What the corpus must contain
Five requirements answer these conditions in weed detection.They are decided in the sampling plan.The same fields at several times of day, which isolates the effect of light from the other factors.The same fields before and after rain, which isolates the effect of moisture.Fields with contrasting soil types, including ones atypical for the region.Fields under direct drilling and under conventional tillage, whose residue cover differs radically.And continuous sequences rather than isolated images, which permits measuring stability.One practical consequence follows. Those five requirements do not increase the volume to annotate proportionally, one field revisited under other conditions supplying a variability new fields under the same conditions would not.What these conditions ask of the annotator
Five additional difficulties weigh on annotating images acquired under unfavourable conditions.Delineation becomes uncertain, a plant partly in shadow presenting no clean outline.Confusion with the background increases, which raises the share of objects classed as undetermined.Time per object grows, each case requiring examination rather than immediate recognition.Disagreement between annotators increases, which makes measuring agreement per condition more informative than a global figure.And a temptation appears, that of discarding the image rather than handling it, which must be explicitly forbidden by the convention.One practical consequence follows for a weed detection project. The last difficulty is the most insidious: with no explicit instruction, annotators naturally discard what is difficult, which produces a corpus biased towards easy conditions with no conscious decision having been taken.The technical responses
Four approaches reduce sensitivity to conditions.Their effectiveness differs markedly.Physical control, through shrouding and artificial lighting, the only response that removes the problem rather than compensating it.Image normalisation, exposure and balance correction, which reduces part of the variability and not the cast shadows.Data augmentation, lighting transformations applied to the corpus, which improves robustness and does not replace real conditions.And learning on diversified data, the most reliable and most expensive response since it presupposes acquiring under every condition targeted.One important observation follows. The first approach is the most effective and it constrains the hardware, which makes it available to implement manufacturers and unavailable to a project working on existing images.Acquisition configurations and their exposure
Three arrangements present very different sensitivities and the choice commits everything else.The shrouded camera with artificial lighting removes most of the lighting variability, at the price of heavier equipment and a proximity constraint to the ground.The open-air mounted camera endures every condition and remains the most widespread configuration, notably on existing implements retrofitted afterwards.And the drone endures the light and partly escapes residue, the vertical view at distance attenuating the effect of ground fragments while losing resolution.One important observation follows for a weed detection project. Those three arrangements produce non-interchangeable corpora, and a project contemplating a move from the second to the first must plan a complete reacquisition rather than a simple adjustment.Evaluating robustness to conditions
Four measures account for it.They differ from an ordinary weed detection evaluation.Performance decomposed by condition, lighting, soil and residue level, which locates the weaknesses.The gap between the best and the worst condition, a fragility indicator more informative than an average.Stability across a sequence, the proportion of objects detected consistently over several consecutive images.And drift over a session, a comparison between the start and the end of an operation.One practical consequence follows for a weed detection project. The second indicator deserves featuring in any performance presentation, since a system with a high average and a wide gap is less usable than one with a lower average and consistency, a difference an average alone does not show.Diagnosing a performance drop in the field
One frequent situation deserves a protocol, a deployed system regularly degrading with no obvious cause.Five verifications locate the origin within half a day.Compare the images from the operation concerned with the corpus images, which immediately reveals an uncovered condition.Examine the evolution over the session, a progressive degradation pointing at the optics rather than the scene.Check the hour and the weather of the operation, a time slot absent from the corpus explaining a gap on its own.Compare the soil type with those in the corpus, an atypical soil being a frequent and rarely suspected cause.And measure stability across a sequence, strong instability indicating a borderline condition rather than an outright gap.One practical consequence follows for a weed detection provider. Those five verifications distinguish a corpus gap, which belongs to annotation, from a hardware or integration problem, which does not, a distinction that protects the relationship as much as it accelerates the correction.The season as a condition
One further dimension escapes the three preceding families and it deserves flagging.One field, one crop and one stage present different appearances from year to year.Four causes contribute simultaneously.The season’s weather, which alters plant vigour and the relative development of crop and weeds.The rotation, which changes the residue present and the flora emerging from the seed bank.The cropping programme, whose prior treatments select the species that persist.And the variety sown, whose morphology differs within one cultivated species.One important observation follows. This variability is the only one a corpus built in one season structurally cannot contain, which makes a second acquisition campaign a qualitatively different contribution from a mere increase in volume.One practical consequence follows for a weed detection project. A system validated over a single season presents a performance whose inter-annual stability is unknown, which should be declared rather than passed over at a first deployment.What these conditions do not permit
Three limits deserve stating since they bound what a corpus can compensate.An image whose information is absent is not recovered, a seedling under a hard shadow simply not being visible.A progressive degradation of the optics is not learned, it is prevented by maintenance.And soil variability between regions is not convincingly simulated, which makes local acquisition necessary for deployment in a new region.That third limit is the most consequential commercially. It means a system validated in one region requires an acquisition campaign before being offered in another, which should feature in a development plan rather than be discovered at the first distant deployment.What covering the conditions costs
Four lines compose that effort and their distribution is surprising.Additional outings, the principal line in logistics and a modest one in image volume, since it is a matter of revisiting known fields.Additional annotation, whose volume grows more slowly than the variability covered, one scene under two conditions representing twice the work for far greater value.The surcharge per difficult image, set out above, which applies to the fraction of the corpus acquired under degraded conditions.And decomposed evaluation, which requires a test set structured by condition rather than a single set.One practical consequence follows. The ratio between the cost of this coverage and the risk it removes is highly favourable, a system unusable under ordinary conditions representing a loss out of all proportion to a few extra outings.One observation completes that for a weed detection provider. This argument presents better upstream than downstream, a client readily accepting a reasoned sampling plan and poorly accepting a request for additional budget motivated by an observed failure.A coverage plan on one page
A simple table suffices to steer this dimension and it fits on one sheet.It crosses the conditions to cover with the acquisitions carried out.Four columns compose that table.The condition, expressed in observable categories rather than continuous values.The number of images acquired under that condition.The number of annotated objects arising from them, more informative than the image count.And the performance measured on that condition, filled in after the first evaluation.One observation follows for a weed detection project. This table is kept from the first outing and it reveals the gaps while it is still possible to fill them, where a review at the end of the campaign reveals them once the window has closed.The condition never covered
One category escapes any sampling plan and it is better assumed than denied.Some conditions are too rare to feature in a corpus and too real to be ignored.Three examples illustrate the point.Exceptional weather events, hail or marked drought, which alter the crop’s appearance to the point of making it unrecognisable.Cropping accidents, a failed sowing or uneven emergence, which produce fields with no exploitable structure.And rare lighting configurations, fog or twilight, which an operation’s calendar almost never meets.A practical answer exists and it does not consist of covering them. The system must detect that it is outside its domain and say so, rather than produce a decision nothing indicates is hazardous.Approaching the question of conditions
Five questions scope this subject and they are asked at the sampling plan stage.Will the system work under controlled lighting. A positive answer radically simplifies the problem and requires a dedicated corpus.What time slots do operations cover. That answer bounds the lighting conditions to cover.What soil types does the deployment zone present. That answer reveals the complementary acquisition needs.What cultural practices, with or without tillage. That answer sizes the residue problem.And will stability across a sequence be measured. That answer conditions detecting a frequent failure mode otherwise invisible.Those five answers determine the sampling plan and the evaluation protocol. Asking them before acquisition avoids this field’s commonest situation, a corpus rich in botanical variability and poor in condition variability.What this chapter teaches
One cross-cutting observation deserves closing this examination because it matches a finding of the preceding series.The variability that makes a system fail is not the one attention goes to.Three findings compose it.Design effort naturally concentrates on the project’s subject, the weed flora, because that is where the technical interest sits.The variability of conditions often exceeds botanical variability in amplitude, and it is easier to cover since it suffices to revisit the same fields.And its omission produces a system that works in demonstration and disappoints in production, a gap whose origin is rarely correctly identified.That finding matches exactly the one the series on aerial LiDAR established about surveys acquired in favourable conditions: what is discarded from a corpus because it appears defective is often the part most representative of real use.The demonstration problem, stated plainly
One consequence of this chapter deserves stating on its own because it explains a pattern buyers meet repeatedly.Demonstrations happen in good conditions. Not through dishonesty, but because a demonstration is scheduled, and anything scheduled is scheduled for a clear morning on a clean field.Three consequences follow for anyone evaluating a weed detection system today.The performance seen is the upper bound, not the expected value, since every condition in this chapter was absent from the demonstration.The gap between that bound and routine performance is not knowable from the demonstration itself, and it varies enormously between suppliers.And the supplier may not know it either, if they have never measured performance decomposed by condition.One practical question resolves this. Ask for the performance gap between the best and the worst condition measured, and for the conditions those were. A supplier who has the figure has done the work; a supplier who has never computed it has a system whose field behaviour nobody has characterised, including them.Common mistakes
These failures recur often enough that naming them is usually enough to avoid them.- Acquiring every image at the same hour of the day.
- Discarding images acquired under cloud or after rain.
- Neglecting soil types atypical for the reference region.
- Building a corpus only on conventionally tilled fields.
- Annotating a plant together with its cast shadow.
- Evaluating only on isolated images without measuring stability across a sequence.
- Presenting an average performance without the gap between conditions.
- Expecting data augmentation to replace real conditions.
- Neglecting drift over the course of an operation.
- Offering a system in a new region without a local acquisition campaign.
