Spending Destructive Assays Where the Harvest Decision Is Uncertain

Deciding when to harvest a medicinal or specialty crop depends on chemistry that can only be measured by destroying the plant. Alkaloids and flavonoids in the same orchid peak two years apart. Sugar in purple waxy corn peaks at a fixed point in kernel development while pigments keep rising. A medicinal root loses half of one active coumarin once the plant bolts, meaning it switches from leafy growth to flowering. Growers and researchers therefore map these curves with a handful of expensive lab assays spaced on a fixed calendar. A camera that measures reflectance across hundreds of wavelengths (hyperspectral imaging) can scan every plot every day at almost no cost, but it only estimates the chemistry indirectly. We propose to use the cheap daily scans to decide which plot, on which day, is worth sacrificing for a lab assay. The selection rule favors assays that would most likely change a plot's harvest date. It does not aim to improve the chemistry prediction everywhere.

The idea borrows from active learning, the family of methods that pick which unlabeled examples to send to an expensive labeler based on how unsure the model is about them. It changes that approach in two ways that the harvest setting forces. First, the uncertainty must hold up when a new field, season, or cultivar looks different from the training data. The standard neural-network uncertainty signal is known to collapse under exactly this kind of shift for plant spectra. Second, the value of an assay should be judged by its effect on a decision, the harvest day, and not by its effect on average prediction error. In short, the daily scans tell us roughly where each plot is on its developmental clock. We spend lab assays only on the plot-days where the model is both genuinely unsure and close enough to a plausible harvest window for the answer to matter.

Core Claim

Under a fixed budget of destructive assays, a decision-focused rule can pick plot-days to assay across many plots that each develop at their own pace. The rule combines two things: a model uncertainty that stays reliable when the field conditions shift, and a weight for how close the plot is to its likely optimal harvest window. We claim this rule achieves lower harvest regret than calendar-grid assays, plain uncertainty sampling, or pure diversity sampling. Harvest regret is the quality lost by harvesting on the model's chosen day instead of the truly best day. We also claim the gain comes mainly from the decision weighting. It does not come from a better uncertainty estimate alone.

The prior work on each side motivates and constrains this claim. The harvest timing literature supplies the motivation. Integration of Transcriptome and Metabolome Provides New Insights to Flavonoids Biosynthesis in Dendrobium huoshanense profiled four growth years and found alkaloids peaking in year 2 (0.62 mg/g) while flavonoids peak in year 4 (8.94 mg/g). Comparative Transcriptomics Unveil the Crucial Genes Involved in Coumarin Biosynthesis in Peucedanum praeruptorum Dunn showed that bolting, not calendar age, cuts praeruptorin B by 52.1%. COMPARATIVE STUDIES ON NUTRITIVE QUALITY AND ANTHOCYANIN SYNTHESIS OF PURPLE SELENIUMENRICHED WAXY CORN AND COMMON PURPLE WAXY CORN DURING KERNEL DEVELOPMENT found a sugar peak at 28 days after pollination while anthocyanins keep accumulating. All three sample on fixed, pre-chosen time points and describe the curves after the fact. None asks which time points were worth measuring.

On the sensing side, Prediction of Tea Polyphenols, Free Amino Acids and Caffeine Content in Tea Leaves during Wilting and Fermentation Using Hyperspectral Imaging showed that hyperspectral regression tracks processing chemistry well, but it used a fixed half-hourly or hourly schedule with no uncertainty estimate at all. The active learning side supplies the method and also the key constraint. Dropout-based Active Learning for Regression (MCDUE) selects regression samples by the spread of predictions when dropout, the random switching-off of network units, is left on at test time. It beats random and max-min distance sampling on tabular benchmarks. However, Uncertainty Assessment in Deep Learning-based Plant Trait Retrievals from Hyperspectral data found that this same dropout signal covered only 5–6% of the actual error range on out-of-domain plant spectra. A spectral-distance score (Dis_UN) calibrated with quantile regression did far better. We treat that result as ruling out dropout variance as the sole acquisition signal in the field.

Research Program

The system has four parts. The first is a chemistry predictor that maps each plot's daily spectrum to target compound levels, fitted to whatever assays have been collected so far. It will be a small ensemble of networks alongside a partial least squares model, the standard linear chemometric regression.

The second is a per-plot developmental trajectory model. It fits a smooth curve of each compound over time for each plot, pooling across plots, and turns it into a posterior over that plot's best harvest day $t_i^*$. The best day is defined by a composite quality score, for example sugar subject to an anthocyanin floor for corn. Composites are used so that the conflicting peaks documented above are part of the decision.

The third is an uncertainty estimate $u(i,t)$ for plot $i$ on day $t$. It takes the larger of the ensemble spread and a Dis_UN-style bound. That bound maps the spectral distance to the already-assayed set into an upper error quantile, as in Uncertainty Assessment in Deep Learning-based Plant Trait Retrievals from Hyperspectral data.

The fourth is the acquisition rule, which scores each candidate plot-day as

$$a(i,t) = P\big(|t - t_i^*| \le \delta\big)\cdot u(i,t),$$

where the probability is taken under the current posterior over plot $i$'s optimal harvest day and $\delta$ is the harvest window width in days. Each day, the top-scoring plots within the day's assay budget have a sister plant sacrificed for lab analysis, and the models are refit.

We add one guard against measurement noise. Repeated scans of the same plot give a direct estimate of scan-to-scan noise, and scans with outlying noise are discounted rather than preferred. This follows the finding in Privacy-preserving and Uncertainty-aware Federated Trajectory Prediction for Connected Autonomous Vehicles that selecting the most uncertain samples mostly selects irreducible noise, while mid-range uncertainty selects learnable cases.

Baselines

The naive reference is the calendar grid used throughout the harvest literature. The same number of assays is spread evenly over days and plots, and each plot is harvested on the day its fitted curve peaks. A second naive reference is the fixed agronomic rule, which harvests every plot on the published date, for example 28 days after pollination for corn. Beating these means lower mean harvest regret at equal assay budget.

The strongest existing active method is MCDUE from Dropout-based Active Learning for Regression. It selects plot-days with the largest dropout-prediction spread and is run with the same predictor, the same budget, and the same refitting schedule. For this pairing of active selection with a harvest decision, we found no reported result on Lacuna. The paper reports only "up to 20%" gains over max-min distance sampling on tabular benchmarks.

For the chemistry predictor itself, the reference number is CARS-PLS at R² = 0.91 for tea polyphenols in Prediction of Tea Polyphenols, Free Amino Acids and Caffeine Content in Tea Leaves during Wilting and Fermentation Using Hyperspectral Imaging. Our predictor should come near this when given the full label set. If it does not, the active learning comparison is uninformative.

We also run a Gaussian process with expected improvement, a standard Bayesian optimization method that searches over time for each plot using spectral features as inputs. It is the obvious sequential-design competitor.

Every method gets the same data pool, the same predictor family, and the same hyperparameter search of 20 configurations tuned on a separate calibration site.

Two ablations isolate the key ingredient. The first uses our full uncertainty estimate without the harvest-window weight. The second uses the window weight with dropout spread only. If the first ablation matches the full method, the decision focus adds nothing and we will say so.

Why not just…?

Why not just assay on a calendar grid, as every study cited here does? A grid ignores that plots develop at different speeds. The Peucedanum loss is triggered by bolting, whose timing varies with site and plant, so a grid spends assays on days far from any plot's transition. This is testable rather than assumed. If plots in our field trial turn out to be developmentally synchronous, the grid will tie with our method, and the regret comparison will show it.

Why not just use plain dropout-uncertainty sampling (MCDUE)? Under site shift, dropout spread has been measured to cover 5–6% of the true error range on plant spectra (Uncertainty Assessment in Deep Learning-based Plant Trait Retrievals from Hyperspectral data). An acquisition rule built on it will be overconfident exactly where a new field looks unfamiliar. On a single in-domain site it may be competitive, which is why the cross-site transfer experiment is the decisive test.

Why not just sample for spectral diversity, which is the distance score alone with no model? Diversity sampling spreads assays over everything that looks different, including plots that are clearly far from harvest. MCDUE's authors found max-min diversity worse than model uncertainty in-domain. Out of domain, the distance score is the better uncertainty signal, and we already use it. What diversity alone lacks is the harvest-window weight, and the first ablation measures exactly that.

Why not just run Bayesian optimization over harvest time per plot? This is the most serious alternative. For one plot with time as the only input, a Gaussian process with expected improvement is essentially the right tool. Our argument is that it treats plots independently and uses spectra only as covariates, so it cannot transfer information about developmental stage across plots or detect when a new site is out of domain. If it matches our method on regret in the multi-plot trial, the contribution reduces to "use Bayesian optimization with a calibrated spectral kernel," and we will report that outcome.

Proposed Experiments

The first experiment is a retrospective simulation on the tea wilting and fermentation data from Prediction of Tea Polyphenols, Free Amino Acids and Caffeine Content in Tea Leaves during Wilting and Fermentation Using Hyperspectral Imaging: 116 samples with four replicates, hourly and half-hourly. Each lab value is hidden until the method "buys" it. The decision is the processing stop time that meets a target polyphenol and amino acid level. We report stop-time regret against budget for 10–50% of the samples. This set is small and comes from a single cultivar, so it checks the basic mechanism only.

The Dendrobium and Bupleurum datasets have three or four time points, too few for any sequential design. We will not use them for acquisition claims.

The main experiment is a prospective field trial with the two waxy corn cultivars from the waxy corn kernel development paper, grown at two sites. Every plot is scanned daily from 18 to 34 days after pollination. The assay budget is 20% of plot-days, covering sugar and anthocyanins by HPLC, the standard liquid chromatography assay. A held-out set of plots is assayed densely every day to provide ground-truth optimal days. Each method is trained on site A and deployed on site B with the same budget.

The decisive checks are threefold. The full method must beat MCDUE and the Gaussian process on site-B regret by a margin larger than the spread across five random seeds and plot splits. The window-weight ablation must lose most of that margin. The reported uncertainty must stay calibrated on site B, measured by expected normalized calibration error, the average mismatch between predicted and observed error. Its value should be near the 0.44–0.53 that Dis_UN reached, rather than the 1.4–17.6 of ensemble and dropout estimates. If the gain appears only in-domain, or disappears without the window weight, the central claim fails.