Targeted spraying is the most widely deployed use of weed detection and the one with the hardest constraints. The system does not decide in the abstract: it commands a nozzle that must open while the machine advances.This article sets out what that constraint imposes upstream, how it modifies the trade-off between errors and what it requires of a corpus. It extends the article on row crops.
The decision chain
Four stages separate the image from the droplet in weed detection.Each consumes time.Acquisition, whose exposure must stay short to avoid motion blur.Inference, which produces the location of targets in the image.Transformation into machine coordinates, which projects the detection into the boom’s frame.And nozzle command, whose opening and closing have their own mechanical inertia.One important practical consequence follows. The distance travelled during that chain determines the lead needed between the detection point and the treatment point, and that lead requires the camera to look ahead of the boom rather than level with it.What latency imposes
This physical constraint governs the whole design.It bounds the technical choices of a weed detection system.The time available between detection and treatment is fixed by forward speed and by the distance between the field of view and the boom.Three consequences follow.Model complexity is bounded, a more performant but slower network potentially being unusable at the target working speed.Computation happens on board, sending to a remote server introducing a latency and a network dependency incompatible with field use.And forward speed becomes a system parameter rather than an operator preference, acceleration degrading placement precision.One important observation follows for a weed detection project. Those three constraints are settled before the corpus is built, since they determine the resolution and the rate of the images the system will have to work on.The asymmetry of errors
This property distinguishes targeted spraying from most vision applications.Both error types carry a cost, but that cost is very unequal and its ratio depends on context.An omission leaves a weed untreated, which produces a yield loss, seed set and increased infestation the following year.A false detection triggers an unnecessary spray, which costs a dose of product and reduces the saving sought.Three readings follow.The cost ratio leans strongly towards tolerating false detections, a wasted droplet bearing no relation to an untreated plant.That ratio partly inverts where the contractual objective concerns input reduction, performance then being measured on the volume saved.And it inverts completely where a false detection strikes a plant to protect, a cover or an intercrop.One practical consequence follows. The decision threshold is not set once and for all: it is set by context, and that dependency should be exposed to the client rather than hidden behind a single performance figure.The two treatment regimes
Two logics coexist in targeted spraying.They do not call for the same weed detection system.Nozzle-level treatment opens only the nozzles situated over a detection, which produces the maximum saving and demands a placement precision of the order of a nozzle width.Section-level treatment opens a group of nozzles as soon as a detection appears within their span, which reduces the saving and strongly relaxes the precision constraint.Three differences follow for annotation.The localisation precision required differs by an order of magnitude between the two regimes.Section-level treatment tolerates approximate localisation, which makes a coarse enclosing box acceptable where nozzle-level treatment demands a precise centre.And section-level treatment makes the weed mat unproblematic, the whole zone triggering opening with no individual being located.One important observation follows. The regime adopted must be known before annotation, since it changes the mat convention the pillar article set out and it determines whether precision effort is useful or wasted.Green on brown and green on green
One commercial distinction structures this market and it deserves stating because it covers two very different problems.Detecting vegetation on bare soil requires no botanical recognition: any plant present before sowing or between two crops is undesirable.Three consequences follow.The problem reduces to vegetation segmentation, an operation simple spectral indices largely resolve with no learning.The corpus needed is minimal, the variability concerning lighting and soil conditions rather than the flora.And the corresponding systems are mature and have been deployed longer than those operating in a standing crop.Detecting in a standing crop is an entirely different problem, since two vegetations must be distinguished.One important practical consequence follows for a weed detection project. These two cases are regularly conflated in commercial discussions, a convincing demonstration on bare soil saying nothing about performance in a crop, and clarifying which is intended is the first question to ask.What the corpus must contain
Five requirements characterise a corpus intended for targeted spraying.Images must be acquired under real conditions of use, at working speed and with the mounted equipment, a corpus built under favourable conditions producing a disappointing system.Motion blur must feature in the corpus rather than be excluded from it, since it will be present in production.Localisation must be annotated with the precision the treatment regime demands, no more and no less.Plants to protect must be present in sufficient numbers for the system to learn to spare them.And boom-edge configurations must be covered, the extremities of the field of view presenting an optical distortion the centre does not.One practical consequence follows for a weed detection project. The second requirement is counterintuitive and regularly breached: discarding blurred images produces a cleaner corpus and a system that fails precisely in the conditions where it will work.Real working conditions
Six factors degrade images under spraying conditions and they must feature in the corpus.Motion blur, growing with speed and with low light.Boom vibration, which displaces the field of view between successive images.Dust raised by the machine, which veils the image and scatters the light.Shadows cast by the machine itself, whose geometry changes with the hour and the direction of travel.Rapid lighting variation, a passing cloud changing exposure within seconds.And moisture on the optics, early in the day or after a treatment.One practical consequence follows for a weed detection project. Those six factors are met in production and rarely during a dedicated acquisition campaign, which makes acquisition under real working conditions more valuable than careful acquisition, contrary to intuition.The question of geometric calibration
This operation links the image to the world.Its failure produces errors invisible to annotation.A correct detection in the image must translate into a correct position under the boom, which presupposes knowing the mounting geometry.Four error sources contribute.Camera height, variable with the machine’s setting and the ground relief.Viewing angle, whose drift displaces every projection.Optical distortion, more pronounced at the edge of the field of view.And variable chain delay, whose fluctuation offsets placement in the direction of travel.One observation follows. Those four sources produce a systematic offset that detection quality does not compensate, and a perfectly trained system can treat beside the target, a failure belonging to integration rather than to annotation.The role of pre-annotation
One method appreciably reduces corpus cost in this particular use and it deserves setting out.A system already deployed produces detections that can be reused as an annotation starting point.Three benefits follow.Corrected volume replaces produced volume, an operation several times faster on common configurations.Disagreements between the system and the annotator locate the model’s weaknesses, which directs the composition of the next batch.And the cases where the system failed constitute precisely the most useful corpus, an already mastered example contributing almost nothing to training.One caveat accompanies those three benefits. Pre-annotation induces a confirmation bias, an annotator validating a proposal more readily than producing an annotation, which requires periodically measuring agreement on batches annotated without assistance to verify that quality is not drifting.Measuring the real saving
Three indicators account for it.They do not follow from weed detection metrics.The rate of reduction in volume applied, a direct comparison with a blanket treatment on the same field.The coverage rate of weeds present, which measures what the system actually treated rather than what it detected.And the rate of unnecessary treatment, area sprayed with no weed, which explains part of the saving missed.One important practical consequence follows. Those three indicators are measured in the field and not on a test set, which presupposes a comparison protocol on control strips, and their absence explains the frequent difficulty in economically demonstrating a system that is otherwise technically satisfactory.The control strip strategy
A measurement arrangement conditions the economic demonstration and it is set up before deployment.The principle is to leave strips in the field treated differently in order to compare results.Four strips compose a complete arrangement.The blanket strip, treated in full, which gives the reference for consumption and efficacy.The targeted strip, treated by the system, which gives the real consumption and the efficacy obtained.The untreated strip, which gives the infestation pressure and permits estimating the loss avoided.And the targeted strip at a modified threshold, which permits measuring the effect of the setting rather than supposing it.One important observation follows. This arrangement costs a portion of the field and an operational effort, and it constitutes the only admissible proof of a saving, which makes its absence problematic as soon as an investment must be justified.One practical consequence follows for a weed detection provider. Proposing this arrangement to the client at scoping distinguishes them markedly, since it shifts the conversation from announced performance to demonstrable performance.Hybrid modes
Three arrangements combine targeted and blanket treatment, and they are more widespread than full targeting.Targeting the inter-row alone, the band over the row receiving a continuous application, an arrangement capturing most of the saving without demanding the hardest recognition.The dual circuit, one tank applying a product at reduced rate throughout and a second applying a specific product on detection, which combines agronomic safety and saving.And conditional targeting, the system switching to continuous treatment above a detected density threshold, which avoids performance degradation in heavily infested zones.One important observation follows for a weed detection project. Those three arrangements relax the constraint on the detection system, since an omission is caught by the background treatment, which shifts the cursor towards tolerating false detections and changes the parameter setting expected.What targeted spraying does not solve
Four limits deserve stating.This use attracts high expectations.The saving depends on the infestation rate, a heavily infested field producing a small saving since almost all the area is treated.Resistance is not reduced, the same active substance exerting the same selection pressure on the plants struck.Work rate falls, forward speed being bounded by the decision chain.And the equipment investment is significant, which shifts the break-even point towards large areas or farms under strong regulatory pressure.That first limit deserves emphasis. It means the technology contributes least where the need is greatest, a property that should feature in an honest economic evaluation rather than be discovered after investment.What deployment reveals about the corpus
One loop deserves organising because it is the principal improvement lever after first commissioning.A deployed system meets configurations the initial corpus did not contain, and it fails on them in identifiable ways.Four exploitable sources of information follow.Images recorded during failures flagged by the operator, which point directly at corpus gaps.Low-score detections, which locate configurations the model is unsure about with no error being observed.Differences between fields, a field where performance drops signalling an uncovered context.And operator feedback on missed treatments, qualitative but precise information on the species and stage involved.One practical consequence follows. This loop turns a fixed corpus into an evolving one, it concentrates annotation effort on what is actually missing, and it presupposes a recording arrangement decided before deployment rather than improvised afterwards.What the operator expects of the system
Four ergonomic requirements condition adoption and they escape technical discussions.Predictability, an operator needing to understand why the system treated or did not, failing which they lose confidence and revert to manual.Accessible adjustment, the ability to tune a threshold in the field according to the plot rather than depending on external intervention.Visibility of operation, immediate feedback on what is detected permitting verification that the system is working.And graceful degradation, a system switching to continuous treatment when it meets a condition it cannot handle rather than stopping or treating at random.One practical consequence follows. Those four requirements do not belong to weed detection itself and they nonetheless determine whether the system will be used, which matches the finding the preceding series established on the gap between a correct deliverable and a used one.Approaching a targeted spraying project
Five questions scope this type of project and they are asked in one meeting.Which treatment regime, nozzle or section. That answer determines the annotation precision needed and the mat convention.What forward speed is targeted. That answer bounds model complexity and therefore achievable performance.What is the usual infestation pressure. That answer determines the saving expected and therefore the profitability.Are plants to protect present. That answer changes the trade-off on false detections.And how will the saving be measured. That answer imposes a comparison protocol that is not set up after the fact.Those five answers determine the corpus, the parameter settings and the economic demonstration. Asking them before acquiring images avoids annotating to a precision the treatment regime did not require.What this use case teaches
One cross-cutting observation deserves closing this examination because it matches a finding of the preceding series.Detection performance is only one link in a chain any element of which can cancel the others.Three findings compose it.An excellent model paired with an approximate calibration treats beside the target.An excellent model too slow for the working speed is unusable.And a perfectly functional system with no economic measurement protocol stays unsellable, for want of being able to demonstrate what it contributes.That finding matches exactly the one the series on aerial LiDAR established about power line corridors, in an unrelated domain: a system’s value sits not in its detection capacity but in the complete chain that turns that detection into measurable action.Why the saving is easier to promise than to prove
One commercial pattern deserves naming because it affects every supplier in this market equally.The headline figure for a targeted spraying system is a percentage of herbicide saved. That figure is real, it is achievable, and it is almost never verifiable on a given farm.Three reasons explain the gap.The reference is missing. A saving is a comparison, and the blanket treatment the system replaces was not applied, so there is nothing to compare against unless control strips were planned.Infestation varies between years more than between systems. A season with low pressure produces a large saving that says nothing about the equipment, and the following season can reverse the impression.And the figure quoted comes from someone else’s fields. A percentage measured on a demonstration site carries no commitment about a farm with different pressure, different flora and different spacing.One practical consequence follows for a weed detection provider. The honest position is to quote no percentage and to propose the measurement instead, which is a weaker sales line and a considerably stronger basis for a second contract.Common mistakes
These failures recur often enough that naming them is usually enough to avoid them.- Building a corpus by discarding the blurred images that will be present in production.
- Annotating to a precision the section regime did not require.
- Deciding the treatment regime after starting annotation.
- Choosing a model without checking its latency at the target working speed.
- Setting a single decision threshold for every context.
- Neglecting plants to protect in the corpus.
- Omitting field-of-view edge configurations.
- Attributing to detection a failure that belongs to geometric calibration.
- Promising a saving without knowing the usual infestation pressure.
- Launching a deployment with no measurement protocol on control strips.
