When Looking Again Doesn't Help: View-Dependent Detector Bias in Autonomous Crop Scouting

Drones that scout fields for disease or weeds decide where to fly next by asking which patch of ground would most reduce their uncertainty. This "information gain" logic quietly assumes that the onboard image detector makes independent mistakes, so that flying over a patch twice gives two independent pieces of evidence. Real detectors do not behave this way. A lesion covering about 1% of the frame, a leaf in shadow, or a mite hidden under a fold is missed the same way every time it is seen from the same height and angle. Repeated passes then produce confident, wrong maps.

This note proposes to model the detector as a biased oracle. An oracle here means the black box that answers "is there disease here?". Its answers can carry a bounded, persistent error that depends on the viewpoint, and that error cannot be averaged away. From this model we build a matching hard instance and a scouting policy. The hard instance is a family of infestation maps that no scout can tell apart from a healthy field unless it changes viewpoint. The policy credits a revisit only for what is new about its view, and it descends exactly where the bias floor of the current altitude hides the outbreak sizes that matter. The goal is not a better detector. The goal is a scout that knows what its detector structurally cannot see and spends battery accordingly.

Core Claim

Carbone et al. built a decentralized drone swarm that maps weeds by greedy information gain using an empirical confusion-matrix model of an onboard Faster R-CNN. It reaches occupancy-grid MSE ≈ 0.15 in about one-third of the flight time of uniform coverage. The authors themselves flag the "same perspective problem": repeated views from similar angles do not resolve ambiguity, yet their filter treats them as independent. This paper motivates our work. It found the gap but left it unmodeled.

We write the observation of cell $c$ from viewpoint $v$ (altitude, pitch, ground resolution) as

$$y_t = f(c) + b(c, v) + \xi_t,\qquad |b(c,v)| \le \varepsilon(v,a),$$

where $f(c)$ is the true infestation signal, $b$ is a bias that repeats identically for the same cell and view, $\xi_t$ is independent noise, and $\varepsilon(v,a)$ is the largest bias the detector exhibits at view $v$ for lesions of apparent image area $a$.

This borrows its method from Singer and Vondrák. They prove that bounded non-random oracle error, unlike stochastic noise, cannot be removed by resampling and can make optimization intractable even on affine functions. Their setting is abstract zeroth-order convex optimization. Ours is a spatial search where the error band is a measurable function of viewpoint, and where the searcher can pay to change the band by descending.

The claim follows by an indistinguishability argument in the style of that paper. Suppose an early outbreak has signal amplitude $\delta$, and $\delta \le \varepsilon(v,a)$ at every view a policy uses. Then the adversary can set $b = -f$ on infected cells, making the field look exactly healthy. Coarse passes therefore provide zero pruning for such outbreaks, no matter how many repeats are made.

For $k$ hidden outbreaks placed adversarially among $N$ cells, any policy needs on the order of $N/k$ fine-view inspections. That is the standard needle-in-haystack count, and it holds once coarse evidence is worthless. Coarse views can prune cells for outbreak sizes with $\delta > \varepsilon(v,a)$, and an algorithm that knows $\varepsilon$ can achieve the matching upper bound.

The new content is twofold. First, the pruning power of each altitude is set by its measured bias floor, not by its stochastic noise level. Second, a policy that ignores this distinction can be shown to declare infected cells healthy with arbitrarily high confidence.

Iwazaki motivates the "sharp peak" picture of an early outbreak. That paper uses Dirac-like narrow functions to tighten Gaussian-process bandit regret lower bounds. However, it assumes independent noise and a hypersphere domain. We do not reuse its bound. We use only its lesson that narrow peaks are what hardness constructions exploit.

Research Program

The work has three parts.

Measuring the bias floor. We first measure $\varepsilon(v,a)$ empirically. We take images with expert masks from PlantSeg, which has 7,774 in-the-wild images and reports failure concentrated on lesions near 1% of image area and on low convex-hull-ratio lesions. We also take rice-leaf boxes from IDADP as used by DHLC-DETR.

From these we synthesize "altitudes" by downsampling, which shrinks apparent lesion area $a$. We synthesize "repeat views" by small spatial jitter and photometric perturbation that keep the viewpoint class fixed. For a segmentation model (ConvNeXt-L) and a lightweight edge model (MobileNetV3), we then measure two things. The first is how detection error decomposes into a component that persists across jittered repeats and a component that varies. The second is the worst-case persistent error per bin of $a$.

This empirical decomposition is itself a deliverable. We found no reported measurement of within-viewpoint error correlation for plant-disease detectors.

Building the simulator and policy. Second, we build a scouting simulator. Its field maps are seeded from the Wageningen sugar-beet/volunteer-potato setup in Carbone et al., and from synthetic sparse-outbreak maps with clusters of controlled size. Detector responses are drawn from the measured bias-plus-noise model rather than an independent confusion matrix.

In this simulator we implement a bias-aware information gain policy. It credits a repeated observation only with its independent-noise information, conditioning on the persistent bias. It treats a cell as resolved for outbreak size $\delta$ only once some view with $\varepsilon(v,a) < \delta$ has been used. It chooses descents by trading flight cost against the pruning each altitude can actually certify.

Constructing the hard family. Third, we construct the adversarial infestation family from the lower-bound argument and run every policy against it. This exhibits the failure concretely: the standard filter's posterior "healthy" probability on infected cells goes toward 1 as revisits accumulate.

Baselines

The naive reference point is uniform lawnmower coverage. We run it once at the coarse altitude and once at the fine altitude, under the same battery budget. The fine-altitude lawnmower cannot be fooled by coarse bias, but it covers less area per battery. It sets the floor that any adaptive policy must beat on missed outbreaks per unit flight time.

The strongest existing method is the greedy information-gain swarm of Carbone et al. with its independent confusion-matrix sensor model. Its reported result is reaching MSE ≈ 0.15 in about one-third of uniform-coverage flight time on the Wageningen data. It is the strongest competitor because it is the only cited agricultural scout that already adapts flight to uncertainty.

We reimplement it in our simulator with the same swarm sizes, the same detector, the same flight-cost model, and the same hyperparameter search budget as our policy. We also reproduce its uniform-coverage ratio under its original independent-error model as a sanity check before switching on view-dependent bias. Beating it means two things. Under realistic measured bias, our policy must achieve a clearly lower missed-outbreak rate for sub-floor outbreak sizes at equal flight time. It must also not lose more than a few percent in grid MSE on the original, large-patch weed maps where bias matters little.

We add a Gaussian-process upper-confidence-bound scout, which navigates toward cells with high predicted infestation plus uncertainty. It uses a squared-exponential kernel over the field, independent noise, and the same altitude actions.

The key ablation keeps everything in our policy but replaces the persistent-bias model with a heteroscedastic independent-noise model. That model has a per-altitude variance fitted to the same measurements. The ablation isolates whether the gain comes from modeling persistence or merely from knowing that high altitude is noisier.

Why not just…?

Why not just calibrate a per-altitude confusion matrix? This is the most serious threat, and it is exactly our heteroscedastic ablation. A per-altitude model correctly says high altitude is less reliable. But it still lets ten same-view passes compound into near-certainty, which is the mechanism the lower bound attacks. If the ablation matches our policy on missed outbreaks, then persistence is negligible for these detectors. We would report that as a negative finding. The measurement in the first part of the program tells us in advance which case holds.

Why not just discount repeated observations by a tuned constant? A fixed discount factor on revisits is a cheap stand-in for persistence, and it may recover much of the benefit. We include it as a baseline tuned on held-out maps with the same budget. Our expectation is that a constant discount is miscalibrated, because the persistent share of error depends strongly on lesion area $a$: near zero for large lesions, dominant near 1% of frame. A single constant cannot track that dependence. If the tuned constant matches us across outbreak sizes, the simpler heuristic wins, and the contribution reduces to the measurement and the hardness construction.

Why not just use a bigger detector? PlantSeg reports that ConvNeXt-L reaches only 46.24% mIoU in the wild and MobileNetV3 only 10.22%, with failures concentrated on tiny and fragmented lesions. Even a better model shrinks $\varepsilon$ without removing it, and the edge models drones can carry sit at the bad end. The framework applies to any detector. We run it with both models to show how the optimal descent schedule shifts as $\varepsilon$ shrinks.

Why not just fly low everywhere? The fine-altitude lawnmower is immune to coarse bias. Carbone et al.'s three-fold time advantage of adaptive over uniform coverage shows how costly this is. Our lower bound says low flying is unavoidable only for outbreak sizes below the coarse floor. The policy should therefore fly low only where it must, and the lawnmower baseline measures that saving directly.

Proposed Experiments

Measurement. The first decisive check is the bias decomposition on PlantSeg and IDADP. For each detector and each apparent-area bin, we report the fraction of error that persists across jittered same-view repeats. If that fraction is small everywhere, the proposal's premise fails early and cheaply.

Hard family. The second check runs all policies on the constructed adversarial family. We report the posterior "healthy" confidence assigned to infected cells as a function of revisit count. The prediction is that it rises toward 1 for the information-gain and Gaussian-process baselines and stays bounded for ours.

Realistic maps. The third check uses realistic maps: Wageningen-seeded weed fields and synthetic sparse outbreaks at several cluster sizes straddling the measured floor. Swarm sizes run from 10 to 50, matching Carbone et al.'s range. Metrics are missed-outbreak rate at a fixed battery budget, flight time to first detection of each outbreak, occupancy-grid MSE, and the calibration of "healthy" declarations.

The proposal succeeds if three things hold together. Missed outbreaks drop sharply for sub-floor outbreak sizes. MSE on large-patch maps stays close to Carbone et al.'s level. And the constant-discount and heteroscedastic baselines fail to close the gap.