Separating Glare from Tissue: A Polarization-Based Test of Lab-to-Field Transfer in Plant Disease Vision

Plant disease classifiers are trained mostly on leaves photographed in the lab or greenhouse, and they lose accuracy outdoors. Part of that loss is optical. Sunlight bounces off the waxy leaf surface as glare, and that glare depends on sun angle, leaf tilt and sky conditions rather than on the disease. The disease signal comes from light that has entered the leaf, scattered inside the tissue, and come back out. Physics gives a way to split the two. Glare off a smooth surface keeps much of its polarization, the direction in which the light wave oscillates. Light that has scattered inside tissue loses its polarization. A camera that records brightness through polarizing filters at several angles can therefore estimate, pixel by pixel, how much of the image is surface glare and how much is tissue. This note proposes giving the classifier only the tissue part, to test whether that input stays nearly the same between lab and field. It also proposes predicting in advance where this invariance will fail.

The second half of the idea comes from an unrelated field. In weak decays of B mesons (heavy particles that decay into pairs of spinning particles), angular-momentum conservation plus the handedness of the weak force predict that the daughter spins almost always line up along their flight direction. The fraction with this alignment is called the longitudinal polarization fraction $f_L$, and the prediction is $f_L \approx 1$. The prediction holds in simple "tree-level" decays. It fails badly in "penguin" decays, which proceed through quantum loops and in which the produced particles can scatter off each other again before being detected. The lesson for field validation is methodological: a physically derived invariance holds exactly where its assumptions hold, and it breaks in a predictable place, namely where secondary scattering sets in. For leaves, the analogue of that secondary scattering is light bouncing between overlapping leaves in a dense canopy. We propose to measure that multi-bounce regime directly from the polarization data. Field validation then becomes a stratified test with a stated prediction, rather than a single averaged accuracy number.

Core Claim

We claim that most of the lab-to-field accuracy drop for leaf-level disease recognition under sunlight comes from surface glare, and that a model trained on the depolarized tissue component will transfer from lab to field with at least half of the gap closed. We further claim this holds only in image regions dominated by a single scattering event, such as isolated sunlit leaves. It should fail in regions where the polarization signature indicates multi-leaf bounces or deep canopy shadow, and the failure should be predictable before any field labels are seen.

The particle-physics side supplies the structure of this claim, not its physics. Measurements in tree-dominated channels such as B⁰→ρ⁰ρ⁰ and B⁰→D_s^(*)ρ decays find $f_L$ near unity. Theory for radially excited charmed final states still predicts about 90%. The loop-mediated B⁰→K⁰K̄⁰ stays high at about 0.80. The prediction collapses to 0.4–0.5 in ωK* and related vector–tensor channels and in φK₁ and K₂*(1430) final states. These results play a motivating role. They show the value of stating, before looking at the data, which regime a leading-order invariance should survive in and which it should not. They also show that the size of the correction is itself a finding. We make no claim that leaf optics and quark dynamics share mathematics. The analogy only shapes the validation design.

Research Program

The optical model is the standard mixture of a diffuse and a specular component. A polarizer at angle $\theta$ records brightness

$$I(\theta) = \tfrac{1}{2} I_d + \tfrac{1}{2} I_s \left[1 + p \cos 2(\theta - \phi)\right],$$

where $I_d$ is the depolarized tissue (diffuse) intensity, $I_s$ is the glare intensity, $p \in [0,1]$ is the degree of polarization of the glare, and $\phi$ is its polarization angle. The minimum over angles, $I_{\min} = \tfrac{1}{2}I_d + \tfrac{1}{2}I_s(1-p)$, is a leading-order estimate of the tissue signal. Its error term $I_s(1-p)$ is small near the Brewster angle, where glare is almost fully polarized, and grows away from it. This known, geometry-dependent residual plays the role that the $\mathcal{O}(m_V^2/m_B^2)$ correction plays for $f_L$. It tells us in advance which pixels the method should handle poorly. A second quantity, the pixel's overall degree of linear polarization, provides the stratification signal. Single-bounce glare regions are strongly polarized. Tissue regions are weakly polarized. Multi-leaf interreflection and sky-lit shadow produce intermediate, spatially incoherent patterns, which we use to define a "rescattering index" per image region.

The build has three parts. The first is a paired dataset. We will photograph tomato and bean leaves with a four-angle polarization camera, of the kind that captures 0°, 45°, 90° and 135° in a single shot, at three levels. The levels are excised leaves on a uniform background in the lab (matching the PlantVillage protocol), potted plants in a greenhouse, and open-field plots under varied sun angle and cloud cover. Disease classes will be matched to those in PlantDoc and KUTomaDATA. The second part is the estimator: per-pixel diffuse, glare-angle and degree-of-polarization maps, fed to the same detector and classifier backbones as the RGB baselines. The third part is the stratified evaluation. Field images are binned by rescattering index before labels are inspected. The transfer gap is then reported per bin, where the gap is defined as the accuracy of a field-trained model on field data minus the accuracy of a lab-trained model on the same field data.

The hyperspectral thermal-stress classifier of Osco et al. reaches 99.42% on controlled pot spectra. Its authors name illumination variation and canopy structure as the obstacles to field use. Our dataset includes a spectrometer arm with and without a polarizer, so that we can test whether their diagnostic green (530–570 nm) and red-edge (700–710 nm) bands survive the move to the field once glare is removed. Polarimetric radar work decomposes field backscatter into surface, double-bounce and volume scattering to monitor maize soil moisture and crop signatures. That work supplies the method family: decomposition by scattering mechanism. It does so at the metre scale and for geophysical retrieval, whereas we apply the idea at leaf scale as an input representation for transfer.

Baselines

The naive reference is a standard RGB model trained on lab images with conventional augmentation (flips, affine transforms, noise, colour jitter) and tested on field images. Next to it we report the same backbone trained directly on field images, which bounds the achievable transfer. Every other number is reported as a fraction of the gap between these two.

For detection, the strongest existing competitor is WD-YOLO. It combines dual-scale convolutions with knowledge distillation and reports 65.4% mAP50 on PlantDoc, 9.1 points above YOLOv10n, at 2.78M parameters and 32 fps on a Jetson Nano. It is the strongest competitor because it is the best in-field detector in the file that also meets edge-hardware constraints. For classification, the strongest competitor is Multi-Granularity Feature Aggregation, which aggregates multi-scale attention features and reports 89.75% accuracy and 89.13% F1 on PlantDoc. We also include TomFormer, a CNN–Transformer hybrid, because it is the only work in the file that reports one model across lab, greenhouse and field: 83%, 87% and 81% mAP on PlantVillage, KUTomaDATA and PlantDoc. No field-transfer result is reported on Lacuna for the Osco hyperspectral classifier; we will produce the first one.

Each competitor is retrained from its published recipe on our paired data, first on RGB and then with polarization-derived channels. All arms get an identical hyperparameter search budget, the same GPU-hours and the same augmentation. Beating a baseline means that its polarization-input version closes at least half of the lab-to-field gap in the single-scatter bin, with bootstrap confidence intervals over field plots that exclude its RGB version's gap reduction, while staying within one point on lab data.

Three ablations isolate the key ingredient. The first is a single fixed polarizer, as a photographer would use. The second is a software specular-removal step applied to plain RGB. The third keeps the full diffuse estimate but drops the degree-of-polarization channel. Together they separate the contribution of physical glare separation, per-pixel angle handling, and the stratification signal.

Why not just…?

Why not just rely on augmentation and more data under ordinary ERM? Here ERM means empirical risk minimization, that is, ordinary training that minimizes average loss on the training set. It is a serious option. Kamath et al. show ERM beating a learned-invariance method on their Two-Bit benchmark, with test loss 0.28 against 0.38. Still, the file's own numbers leave a large field residual: YOLOv10n reaches about 56% mAP50 on PlantDoc even with augmentation. Colour jitter also cannot synthesise glare that follows leaf geometry. The naive baseline arm measures this directly. If augmentation alone closes the gap, the core claim is falsified.

Why not just learn the invariance from multiple environments? We could train with IRM (invariant risk minimization, a method that learns features whose predictive relationship holds across several training environments) on lab, greenhouse and field RGB. Kamath et al. show that its practical version learns "pseudo-invariances" and collapses under finite-sample noise. Our proposal fixes the invariance at the sensor, using physics, instead of asking a penalty term to find it. We still run IRM on the multi-environment RGB data as a comparison, so the argument rests on our own data and not only on theirs.

Why not just remove highlights in software from ordinary RGB? If single-image highlight removal matched the polarization camera, the hardware would be unnecessary. This is testable today. We train on PlantVillage and test on PlantDoc's shared classes with and without software highlight removal, then repeat on the paired data. Our expectation is that software removal works on saturated glints and fails on the broad, partially polarized sheen common under diffuse sky. The comparison stratified by rescattering index will show whether that expectation is right.

Why not just put one polarizing filter on the camera? This is the real threat. It is cheap, and near Brewster geometry a single well-oriented filter removes most glare. Outdoors, however, the glare angle $\phi$ changes from leaf to leaf, and a fixed filter suppresses glare on some leaves while leaving it on others. It also cannot produce the degree-of-polarization map that the stratified evaluation depends on. We run it as a full arm. If it matches four-angle capture on transfer, we will report that result: the recommendation then becomes the cheaper hardware, and the four-angle system is justified only for diagnosing where transfer fails.

Why not just collect field data and fine-tune? Fine-tuning on field data is the upper-reference arm, and it will usually win on its own plots. The question is what a lab-trained model can deliver on a new farm before labels exist. That is the setting field-validation work repeatedly flags, from soybean stress phenotyping to greenhouse weed robots. We will also report how many field labels a fine-tuned RGB model needs to match the zero-label polarization model.

Decisive Checks

Three results would settle the proposal. The first is the headline test. In the single-scatter bin, a lab-trained model on diffuse input must close at least half the transfer gap for both WD-YOLO-style detection and MFA-style classification. The fixed-polarizer and software-removal ablations must close measurably less, or else we adopt them as the recommendation.

The second is the pre-registered failure test. The gap must reopen monotonically as the rescattering index rises. The index thresholds are frozen on greenhouse data before any field labels are opened. This is the direct analogue of predicting, from topology alone, that penguin channels will violate $f_L \approx 1$. A flat gap across bins would mean glare is not the dominant mechanism and the invariance claim is wrong.

The third is spectral. The Osco green and red-edge bands, measured through the polarizer on field leaves, must keep their ranking by gain ratio from the lab. If they do not, the tissue signal itself shifts in the field, for example because of compound stresses that isolated chamber experiments cannot produce. Optics would then not be the bottleneck, and that negative result would redirect field-validation effort toward the biology.

References

  1. Measurement of the branching fraction, polarization, and<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>C</mml:mi><mml:mi>P</mml:mi></mml:math>asymmetries in<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>B</mml:mi><mml:mn>0</mml:mn></mml:msup><mml:mo>→</mml:mo><mml:msup><mml:mi>ρ</mml:mi><mml:mn>0</mml:mn></mml:msup><mml:msup><mml:mi>ρ</mml:mi><mml:mn>0</mml:mn></mml:msup></mml:math>decay, and implications for the CKM angle<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>α</mml:mi></mml:math>
  2. Measurement of the Branching Fractions of the Rare Decays $B^0 \to D_s$ (*) + $\pi^{-}$, $B^0 \to D_s$ (*) + $\rho^{-}$, and $B^0 \to D_s$ (*) - $K^{(*)}$ +
  3. <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"> <mml:msub> <mml:mi>D</mml:mi> <mml:mrow> <mml:mo stretchy="false">(</mml:mo> <mml:mi>s</mml:mi> <mml:mo stretchy="false">)</mml:mo> </mml:mrow> </mml:msub> <mml:mo stretchy="false">(</mml:mo> <mml:mn>2</mml:mn> <mml:mi>S</mml:mi> <mml:mo stretchy="false">)</mml:mo> </mml:math> and <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"> <mml:msubsup> <mml:mi>D</mml:mi> <mml:mrow> <mml:mo stretchy="false">(</mml:mo> <mml:mi>s</mml:mi> <mml:mo stretchy="false">)</mml:mo> </mml:mrow> <mml:mo>*</mml:mo> </mml:msubsup> <mml:mo stretchy="false">(</mml:mo> <mml:mn>2</mml:mn> <mml:mi>S</mml:mi> <mml:mo stretchy="false">)</mml:mo> </mml:math> production in nonleptonic <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"> <mml:msub> <mml:mi>B</mml:mi> <mml:mrow> <mml:mo stretchy="false">(</mml:mo> <mml:mi>s</mml:mi> <mml:mo stretchy="false">)</mml:mo> </mml:mrow> </mml:msub> </mml:math> weak decays
  4. Observation of<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>B</mml:mi><mml:mn>0</mml:mn></mml:msup><mml:mo>→</mml:mo><mml:msup><mml:mi>K</mml:mi><mml:mrow><mml:mo>*</mml:mo><mml:mn>0</mml:mn></mml:mrow></mml:msup><mml:msup><mml:mover accent="true"><mml:mi>K</mml:mi><mml:mo>¯</mml:mo></mml:mover><mml:mrow><mml:mo>*</mml:mo><mml:mn>0</mml:mn></mml:mrow></mml:msup></mml:math>and Search for<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>B</mml:mi><mml:mn>0</mml:mn></mml:msup><mml:mo>→</mml:mo><mml:msup><mml:mi>K</mml:mi><mml:mrow><mml:mo>*</mml:mo><mml:mn>0</mml:mn></mml:mrow></mml:msup><mml:msup><mml:mi>K</mml:mi><mml:mrow><mml:mo>*</mml:mo><mml:mn>0</mml:mn></mml:mrow></mml:msup></mml:math>
  5. Observation of<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>B</mml:mi></mml:math>meson decays to<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>ω</mml:mi><mml:msup><mml:mi>K</mml:mi><mml:mo>*</mml:mo></mml:msup></mml:math>and improved measurements for<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>ω</mml:mi><mml:mi>ρ</mml:mi></mml:math>and<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>ω</mml:mi><mml:msub><mml:mi>f</mml:mi><mml:mn>0</mml:mn></mml:msub></mml:math>
  6. Observation and Polarization Measurements of<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>B</mml:mi><mml:mo>±</mml:mo></mml:msup><mml:mo>→</mml:mo><mml:mi>ϕ</mml:mi><mml:msubsup><mml:mi>K</mml:mi><mml:mn>1</mml:mn><mml:mo>±</mml:mo></mml:msubsup></mml:math>and<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>B</mml:mi><mml:mo>±</mml:mo></mml:msup><mml:mo>→</mml:mo><mml:mi>ϕ</mml:mi><mml:msubsup><mml:mi>K</mml:mi><mml:mn>2</mml:mn><mml:mrow><mml:mo>*</mml:mo><mml:mo>±</mml:mo></mml:mrow></mml:msubsup></mml:math>
  7. A Machine Learning Framework for Classifying Thermal Stress in Bean Plants Using Hyperspectral Data
  8. A Soil Moisture Retrieval Method Based on Typical Polarization Decomposition Techniques for a Maize Field from Full-Polarization Radarsat-2 Data
  9. C- and L-band multi-temporal polarimetric signatures of crops
  10. A Lightweight and Efficient Plant Disease Detection Method Integrating Knowledge Distillation and Dual-Scale Weighted Convolutions
  11. Multi-Granularity Feature Aggregation with Self-Attention and Spatial Reasoning for Fine-Grained Crop Disease Classification
  12. Early and Accurate Detection of Tomato Leaf Diseases Using TomFormer
  13. Does Invariant Risk Minimization Capture Invariance?
  14. An explainable deep machine vision framework for plant stress phenotyping
  15. Crop and weed classification based on AutoML