Farmers and plant biologists usually pick one fertilizer or priming dose per species. Different cultivars of the same crop, however, often peak at different doses, and a dose that helps one cultivar can harm another. Today the only way to find each cultivar's best dose is to grow many plants at many doses and weigh them at harvest, months later. This note proposes a faster test. We would film plants from many angles as they grow and fit a 3D model whose surfaces are allowed to bend and stretch over time. From that model we would read off how fast each individual organ (radicle, leaf blade, stem internode) is stretching. The claim is that these organ-level stretch rates, measured in the first days after treatment, reveal each genotype's best dose long before harvest. They should also do this better than whole-plant summaries such as height or projected leaf area.
The enabling technology comes from surgical video. There, 4D reconstruction methods track soft organs that deform under breathing and instrument contact, and they do it in minutes on a single GPU. We cannot simply port these methods, because they assume no new tissue appears. Plants grow new leaves. The proposal therefore has two parts. The method part is a growth-aware deformation model in which new surface can appear over time. The biological part is a dose-finding study that tests whether the kinematics this model extracts actually move the dosing decision.
Core Claim
Genotype-by-dose interactions often act in opposite directions on different organs of the same plant. Whole-plant summaries average these effects together and hide them. Organ-resolved strain rates, taken from a continuous deformation field, do not.
The clearest evidence is in wheat. Exogenous Melatonin Promotes the Salt Tolerance… found that 300 µM melatonin under 100 mM NaCl promoted germination in the salt-tolerant cultivars SM15 and ZM18. The same dose inhibited germination in the sensitive cultivar JM22 and suppressed germ and radicle elongation in all three. Organ-level effects like these partly cancel in any single whole-plant number. A second example is sugarcane. In A Study on Response of Sugarcane Genotypes to Phosphorus Fertilizer, sucrose yield peaked at 192 kg/hm² P₂O₅ for cultivar YT236 but at 288 kg/hm² for ROC22 and LC5136. That study found the optimum only from a 9-month endpoint. These two papers motivate the work: the effect is real, and endpoint measurement is slow.
The testable prediction is concrete. Suppose we predict each genotype's endpoint-optimal dose from data truncated at day $d$. The prediction built on organ strain rates should reach low error at a smaller $d$ than predictions built on dense time series of whole-plant scalars taken from the same images. If it does not, the conclusion is that simple growth curves suffice, and we will report that.
Research Program
Surgical methods we reuse, and their limitation. Two lines of surgical work supply the machinery. Forplane splits space-time into three 2D spatial feature grids and three space-time feature grids, which cut training to about 10 minutes. Endo-4DGS and HFGS instead represent the scene as 3D Gaussian splats (small colored ellipsoid blobs rendered by rasterization) moved over time by a small deformation network. All of them warp one fixed canonical shape across time. That assumption holds for an organ being pulled by a surgical tool. It fails for a seedling that emerges a new leaf. The file names this topology-conservation tension as an open problem.
Growth-aware representation. We give every Gaussian a learned birth time $b_i$ and place the canonical shape at the final observation, when all tissue exists. Gaussian $i$ is then described by
$$x_i(t) = x_i^{c} + D_\theta(x_i^{c}, t), \qquad \alpha_i(t) = \alpha_i \, \sigma\!\big((t - b_i)/\tau\big),$$
where the terms are defined as follows. $x_i^{c}$ is its canonical position. $D_\theta$ is a deformation field stored in Forplane-style space-time grids. $\alpha_i$ is its opacity. $\sigma$ is the logistic function. $\tau$ is a small temperature that lets the Gaussian fade in smoothly. Tissue that does not yet exist is invisible. The deformation field then only has to explain the stretching of tissue that already exists.
Two regularizers keep this well behaved. A penalty discourages births inside regions that are already visible. A soft monotonicity term discourages organs from shrinking, since growing tissue rarely contracts. We keep HFGS's high-frequency losses because leaf-surface texture carries tracking signal.
Kinematic readout. Each Gaussian inherits an organ label, using segmentation of the kind validated on Pheno4D. From the deformation field we compute the spatial gradient of velocity. Its component along the organ's main axis is the relative elongation rate, and its cross-axis components give expansion. Averaged per organ and per day, these form a kinematic profile for each plant.
Dose-finding model. For each genotype we fit a hierarchical dose-response model. It maps the kinematic profile at a chosen day to the endpoint outcome, and it shares strength across genotypes through a genotype random effect. The model outputs a posterior over each genotype's optimal dose. The same model can drive an adaptive design, in which the next batch's doses go to the genotypes whose optimum is least certain. The primary evaluation, however, is retrospective.
Data. No existing dataset crosses genotypes with dose. SG×P has no nutrient treatments, and Pheno4D and GrowSplat grew plants under uniform conditions. We will therefore collect two sets:
- Salinity-priming set. This mirrors the melatonin study: the cultivars SM15, ZM18 and JM22 under 100 mM NaCl, with six melatonin doses around 300 µM and about 12 replicates per cell. Seedlings are imaged daily for 21 days in a multi-camera chamber comparable to NPEC Maxi-Marvin, then harvested for biomass, proline and antioxidant assays.
- Phosphorus set. This mirrors the five-level sugarcane design: several maize lines at five P levels over a vegetative window. Maize is chosen so that tracking can be validated against Pheno4D.
Baselines
Naive reference. Every genotype receives the dose that is optimal for the pooled species average, estimated from endpoints. Its regret, meaning the endpoint loss relative to each genotype's true optimum, sets the ceiling that any genotype-specific method must beat. The same endpoint data also define the "truth" for every other comparison. Truth is the per-genotype optimum of a curve fitted to held-out replicates' harvest biomass.
Strongest competitor for the biological claim: dense whole-plant scalar trajectories. From exactly the same images, we reconstruct each day as a static 3D Gaussian splat. From each reconstruction we extract height, volume, 3D leaf area and projected area. These trajectories feed the identical hierarchical dose-response model with the same tuning budget. The plant 3D reconstruction survey reports that 3D splats reach R² = 0.953 for leaf area, so this baseline gets strong inputs.
To beat it, we need two results. First, a significantly lower optimal-dose error at matched day $d$, or the same error at least several days earlier. Second, the gap should be largest for the melatonin set, where opposing organ effects are documented.
Strongest competitor for tracking: GrowSplat. GrowSplat fits a static splat at each time step and then aligns consecutive steps with global registration followed by Colored ICP (iterative closest point, which matches points by position and color). We run its pipeline on the same frames and compute strain rates from its point correspondences. No reported tracking-accuracy result was found on Lacuna for GrowSplat. Its reported figure is optimization speed ("3s vs 6s for the first 100 iterations"), so we measure its tracking ourselves on Pheno4D.
Direct surgical port. We also run Endo-4DGS and Forplane unmodified, with no birth times, under the same GPU budget. On the surgical EndoNeRF benchmark, the strongest reported rendering quality is HFGS at 38.14 PSNR (peak signal-to-noise ratio, higher is better), against 37.31 for Forplane. As a sanity check, our growth-aware variant, with birth times switched off, should stay within about 0.5 dB of 38.14 on EndoNeRF.
Ablations of the key ingredient. Two ablations isolate what matters. One replaces organ-resolved strain rates with the same deformation field averaged over the whole plant. This tests whether the benefit comes from organ resolution or merely from dense tracking. The other removes birth times, which tests whether modeling growth matters for the kinematics.
Why not just…?
Why not just fit growth curves to height and leaf area over time? This is the real threat, and we treat it as the main comparison rather than arguing it away. If whole-plant scalars predict the optimal dose as early and as accurately as organ kinematics, the deformation machinery is unnecessary for dosing, and the paper reports that. There is a reason to expect otherwise. The melatonin study shows the same dose promoting germination in two cultivars while suppressing radicle and germ elongation in all three. Mixed-sign organ effects of this kind are exactly what a whole-plant sum hides. The whole-plant-averaged ablation tests this mechanism directly.
Why not just use GrowSplat-style per-day reconstruction plus ICP? It may give adequate tracking for slow, large organs. The GrowSplat authors note, however, that registration without temporal priors struggles when leaves bend non-rigidly and when branching structures look alike. Stretch rate is a derivative, so matching errors are amplified when it is computed. We will compare point-tracking error and leaf-length error on Pheno4D. There, LatticeNet's reported mean leaf-length error is 8.9 mm on tomato and its segmentation mIoU is 98.8% on maize. We will also check whether the ICP-derived strain rates degrade the dose prediction. If they do not degrade it, the simpler pipeline wins on cost and we will say so.
Why not just use 2D overhead imaging? 2D imaging is cheap, but it loses leaf angle and suffers from leaves hiding one another. In SG×P, SNP classification accuracy fell from 70% with RGB+3D fusion to 64% with RGB alone. We will include a 2D projected-area trajectory as a further scalar baseline. Its cost is negligible and it bounds how much 3D buys.
Proposed Experiments
Reconstruction. On Pheno4D we measure four things: 3D tracking error of labeled leaf points across days, leaf-length error against the laser scans (precision better than 0.1 mm), strain-rate agreement with scanner-derived elongation, and emergence-time error for new leaves. We run our model, GrowSplat and the unmodified surgical ports. On EndoNeRF we confirm the growth-aware variant does not lose surgical performance.
Decision. On both new datasets we truncate observations at days 2, 4, 7, 10 and 14 and predict each genotype's endpoint-optimal dose. We report absolute dose error, endpoint-biomass regret and a calibration check on the posterior. We cross-validate over replicates and also hold out a dose level to test interpolation.
Decisive checks. Three checks are decisive:
- The organ-kinematics model must beat the dense whole-plant-scalar model at early days, on the same images and with the same tuning budget.
- The advantage must vanish, or at least shrink, under the whole-plant-averaged ablation.
- The melatonin set must show the largest gain, where it should appear as organ strain rates of opposite sign between JM22 and the tolerant cultivars.
If the gain is uniform across both datasets and does not depend on organ resolution, the proposed mechanism is wrong even if the numbers look good.
References
- Exogenous Melatonin Promotes the Salt Tolerance by Removing Active Oxygen and Maintaining Ion Balance in Wheat (Triticum aestivum L.)
- A Study on Response of Sugarcane Genotypes to Phosphorus Fertilizer in Western Guangdong, China
- Efficient Deformable Tissue Reconstruction via Orthogonal Neural Plane
- Endo-4DGS: Endoscopic Monocular Scene Reconstruction with 4D Gaussian Splatting
- HFGS: 4D Gaussian Splatting with Emphasis on Spatial and Temporal High-Frequency Components for Endoscopic Scene Reconstruction
- Pheno4D: A spatio-temporal dataset of maize and tomato plant point clouds for phenotyping and advanced plant analysis
- A survey on 3D reconstruction techniques in plant phenotyping: From classical methods to Neural Radiance Fields (NeRF), 3D Gaussian Splatting (3DGS), and beyond
- GrowSplat: Constructing Temporal Digital Twins of Plants with Gaussian Splats
- SG×P : A Sorghum Genotype × Phenotype Prediction Dataset and Benchmark