Plant Segmentation Studio (PSS)

• Plant Segmentation Studio (PSS) is a modular, workflow-based platform that integrates controlled acquisition, diverse annotation methods, segmentation algorithms, and phenotyping into a unified framework.
• It accommodates various imaging regimes—from 2D scanners and field RGB to 3D microscopy and point clouds—supporting both deterministic plugins and advanced deep model pipelines.
• PSS enhances annotation efficiency and cross-domain robustness by leveraging synthetic data, human-in-the-loop workflows, and zero-shot as well as supervised segmentation strategies.

Plant Segmentation Studio (PSS) is best understood as a modular plant segmentation and phenotyping environment implied by a broad literature rather than as a single fixed algorithm. In this usage, PSS denotes a software stack that spans acquisition design, annotation, segmentation, post-refinement, and quantitative trait extraction across scanner RGB images, field and laboratory photography, video, 3D microscopy, MRI, and point clouds. The papers associated with this concept range from morphology-plus-CRF root segmentation on flatbed scans (Orlando et al., 2017) and ImageJ-based batch seed segmentation (Vale et al., 2020) to hierarchical crop–weed panoptic parsing (Roggiolani et al., 2022, Nguyen et al., 2023), weakly supervised 3D organ segmentation (Luo et al., 2022), collaborative 3D correction interfaces (Spina et al., 2017), and full-resolution point-cloud organ labeling across modalities (Gilson et al., 25 Sep 2025, Owen et al., 6 Mar 2025).

1. Conceptual architecture

Across the literature, PSS is consistently motivated as a workflow system rather than a monolithic model. The recurring stages are acquisition under controlled or semi-controlled protocols, annotation or weak labeling, segmentation proper, postprocessing or structural grouping, and downstream phenotyping. This pattern is explicit in scanner-based seed analysis, which couples a blue-background acquisition protocol to a batch ImageJ plugin (Vale et al., 2020); in video-based disease analysis, which chains YOLOv8, DeepSORT, ROI selection, and DeepLabV3Plus (Marques et al., 2024); in collaborative 3D microscopy correction, which distributes segmentation tiles in a browser and merges them through consensus (Spina et al., 2017); and in hierarchical crop–weed parsing, where detection, prompting, segmentation, and panoptic assembly are deliberately modular (Nguyen et al., 2023).

This suggests that a coherent PSS architecture has at least five layers. First, an acquisition layer constrains imaging conditions when possible, because several systems achieve reliability by simplifying foreground–background separation at capture time rather than compensating entirely at inference time (Vale et al., 2020). Second, an annotation layer supports dense masks, sparse control points, weak image labels, superpixel-assisted editing, or collaborative correction, depending on modality and organ type (Spina et al., 2017, Yang et al., 2020, Xu et al., 2022, Luo et al., 2022). Third, a segmentation layer hosts classical, deep, and foundation-model backends (Orlando et al., 2017, Tamvakis et al., 2022, Nguyen et al., 2023, Xing et al., 11 Sep 2025). Fourth, a refinement layer handles CRFs, morphology, skeletonization, clustering, or hierarchical grouping (Orlando et al., 2017, Bell et al., 2019, Roggiolani et al., 2022, Xing et al., 11 Sep 2025). Fifth, a phenotyping layer converts masks or labeled points into domain-specific measurements such as root length, leaf width, stem diameter, or wood–leaf separation for ecological analysis (Orlando et al., 2017, Luo et al., 2022, Owen et al., 6 Mar 2025).

The literature also shows two software-organizational idioms that are directly relevant to PSS. One is the plugin model, exemplified by the Java/ImageJ seed segmentation plugin with folder-based batch execution (Vale et al., 2020). The other is the task-queue model, exemplified by SEGMENT3D, where image volumes are tiled, dispatched, corrected, and recombined through consensus (Spina et al., 2017). A plausible implication is that PSS is best implemented as a modular host rather than as a single hard-coded application path.

2. Imaging regimes and biological targets

Plant segmentation in this literature is not a single task family. It spans 2D scanner imagery, field RGB, video, underground imaging, confocal stacks, MRI volumes, and point clouds, with target semantics varying from binary foreground masks to hierarchical plant–leaf structures and 3D organ classes.

Imaging regime	Typical segmentation target	Representative papers
Scanner RGB	Arabidopsis roots; detached seeds	(Orlando et al., 2017, Vale et al., 2020)
Field and lab RGB	Leaf blade/veins; crop plants and leaves; rosette leaves	(Tamvakis et al., 2022, Roggiolani et al., 2022, Ward et al., 2018, Bell et al., 2019, Xing et al., 11 Sep 2025)
Video RGB	Individual tracked leaves and damaged regions	(Marques et al., 2024)
Underground imagery	Root vs soil in minirhizotron images	(Xu et al., 2022)
3D microscopy and MRI	Cells in shoot apical meristem; super-resolved root volumes	(Spina et al., 2017, Uzman et al., 2019)
3D point clouds	Stem/leaf organs, canopy leaf/wood, orchard structures	(Luo et al., 2022, Owen et al., 6 Mar 2025, Gilson et al., 25 Sep 2025)

The biological targets differ accordingly. Scanner-root work focuses on thin, elongated, low-contrast structures below the shoot, with phenotyping goals including primary root length, lateral root counts, and curvature or angles (Orlando et al., 2017). Seed analysis instead assumes small detached organs on a uniform background, intended for morpho-colorimetric measurements across large batches (Vale et al., 2020). Grapevine phenotyping separates leaf blade, leaf veins, and background as a 3-class semantic problem (Tamvakis et al., 2022). Arabidopsis rosette studies target leaf-level instance separation under strong overlap (Ward et al., 2018, Bell et al., 2019). Agricultural field systems target joint semantic, plant-instance, and leaf-instance segmentation, sometimes extending to hierarchical panoptic crop–weed parsing (Roggiolani et al., 2022, Nguyen et al., 2023, Xing et al., 11 Sep 2025).

The 3D regimes broaden the concept further. PRMI defines minirhizotron segmentation as a binary root vs soil problem across 72,568 RGB images, with 63,943 pixel-level masks and 8,625 image-level-only labels (Xu et al., 2022). MRI root work casts segmentation as a mapping from a low-resolution volume $I \in \mathbb{R}^{x\times y\times z}$ to a super-resolved binary segmentation $S \in B^{2x\times 2y\times 2z}$ (Uzman et al., 2019). Point-cloud studies operate at organ or material level: Eff-3DPSeg separates stem vs leaf and then leaf instances (Luo et al., 2022), PointsToWood separates wood vs leaf across complete canopies (Owen et al., 6 Mar 2025), and OmniPlantSeg targets organ and structural semantics across several species and sensors (Gilson et al., 25 Sep 2025).

This diversity means that “segmentation” in PSS cannot be reduced to one output type. It includes binary semantic masks, multiclass semantic maps, instance masks, panoptic hierarchies, skeletons, super-resolved voxel labels, and full-resolution per-point semantics.

3. Algorithmic repertoire

The algorithmic space represented by PSS is unusually heterogeneous. Classical image processing remains central in some regimes. Arabidopsis root segmentation begins with contrast stretching,

$$I_c = 255 \left( \frac{I - \min\{I\}}{\max\{I\} - \min\{I\}} \right),$$

then removes leaves by erosion and dilation with a square structuring element of side 3, enhances thin structures with a multi-scale line detector, and finally solves binary labeling with a fully connected CRF (Orlando et al., 2017). Its CRF energy is

$$E(\mathbf{y}) = \sum_i \psi_u(x_i, y_i) + \sum_{i<j} \psi_p(x_i,x_j,y_i, y_j),$$

with unary terms derived from the line-enhanced image and pairwise terms given by a contrast-sensitive Gaussian kernel with Potts compatibility (Orlando et al., 2017). Seed segmentation under controlled scans is even more acquisition-driven: RGB-to-HSB conversion, blue-background rejection, region filling, and area filtering with $A_i \ge \frac{1}{3}A_{\max}$ are sufficient to process 480 images without reported object detection errors under that acquisition protocol (Vale et al., 2020). Arabidopsis leaf segmentation through edge classification uses Canny edge detection, a shallow CNN that classifies edge pixels into background, plant edge, leaf edge, or internal noise, and a deterministic flood-fill/morphology pipeline for region recovery (Bell et al., 2019).

Deep segmentation appears where semantic differentiation or harder visual clutter makes hand-designed rules insufficient. Vine leaf phenotyping evaluates three supervised U-Net-like variants and an unsupervised FCM-based U-Net; the best supervised system is a U-Net with a MobileNetV2 encoder reaching $PA = 0.95$ and $\text{MeanIoU} = 0.75$ on the controlled grapevine task, although vein IoU remains only $0.31$ in the detailed classwise table (Tamvakis et al., 2022). Joint field segmentation of crops, plants, and leaves uses one encoder and three decoders with hierarchical task-specific skip connections and center-offset instance formulation. Its joint training objective is

$\begin{split} \mathcal{L} = & \, w_1 \, \mathcal{L}_{\mathrm{sem}} + w_2 \, \mathcal{L}^p_{\mathrm{cen}} + w_3 \, \mathcal{L}^l_{\mathrm{cen}} \ & + w_4 \, \mathcal{L}^p_{\mathrm{off}} + w_5 \, \mathcal{L}^l_{\mathrm{off}}, \end{split}$

with $w_1 = 1$, $S \in B^{2x\times 2y\times 2z}$0, and $S \in B^{2x\times 2y\times 2z}$1 (Roggiolani et al., 2022). That architecture explicitly operationalizes the hierarchy semantic $S \in B^{2x\times 2y\times 2z}$2 plant instance $S \in B^{2x\times 2y\times 2z}$3 leaf instance, rather than treating the tasks independently.

Foundation-model pipelines extend this repertoire. One field panoptic solution uses DINO and YOLO-v8 detectors to generate box prompts for HQ-SAM and reaches $S \in B^{2x\times 2y\times 2z}$4 PQ+ on the CVPPA hierarchical panoptic challenge (Nguyen et al., 2023). ZeroPlantSeg instead treats hierarchical plant segmentation as a zero-shot problem: SAM produces leaf instances, OVSeg filters them with text prompts such as "green leaf" and "soil", Grounding DINO cross-attention with prompts "stem" and "petiole" estimates leaf base directions, and greedy clustering plus Mahalanobis outlier reassignment groups leaves into whole-plant instances (Xing et al., 11 Sep 2025). This suggests that PSS can support both training-heavy and training-free segmentation modes, with the latter especially valuable for annotation bootstrapping and cross-domain initialization.

Three-dimensional segmentation introduces additional design patterns. MRI root segmentation adapts RefineNet, uses five neighboring slices mapped to RGB via PCA, and predicts super-resolved binary volumes from very few real examples by training on 384 synthetic train and 384 synthetic validation pairs (Uzman et al., 2019). Eff-3DPSeg adopts Sparse ConvUnet with self-supervised pretraining via the Viewpoint Bottleneck loss,

$S \in B^{2x\times 2y\times 2z}$5

then fine-tunes with only sparse point labels (Luo et al., 2022). OmniPlantSeg does not propose a new network but introduces KD-SS, a resolution-retaining KD-tree partitioning strategy that feeds fixed-size sub-samples into DGCNN and reconstructs full-resolution predictions without global down-sampling (Gilson et al., 25 Sep 2025). PointsToWood adds a gated reflectance integration module to a PointNet++/pointNEXT-derived architecture and targets complete-canopy leaf–wood separation across diverse forests and sensors (Owen et al., 6 Mar 2025).

A plausible implication is that PSS should not privilege one model family as the universal default. The literature instead supports a backend registry: morphology+CRF for low-data grayscale roots, U-Net-like decoders for dense 2D masks, detector-prompted foundation models for rapid hierarchical annotation, and point/voxel pipelines for 3D geometry.

4. Annotation, datasets, and human-in-the-loop workflows

The dataset and annotation literature around PSS is as important as the segmentation models themselves. Several papers explicitly show that the limiting factor is not architecture but label production, label fidelity, and domain-specific capture design.

PRMI is the largest explicitly segmentation-oriented root dataset in this set, with 72,568 RGB minirhizotron images, 63,943 pixel-level binary masks, and 8,625 image-level-only labels, all accompanied by metadata such as species, location, tube number, date, depth, and DPI (Xu et al., 2022). The split is tube-based for most species, and switchgrass is handled specially by placing all manually generated ground truth into the test set while using AI-refined masks in training and validation (Xu et al., 2022). The annotation protocol is heterogeneous: for cotton, papaya, peanut, sesame, and sunflower, masks are reconstructed from multiple rectangular boxes; for switchgrass, 600 masks are fully manual and 3,312 are technician-refined U-Net predictions (Xu et al., 2022). This makes PRMI valuable, but also means that “pixel-level ground truth” is not geometrically uniform across subsets.

At the opposite end of the scale, grapevine leaf phenotyping uses only 24 original labeled images at $S \in B^{2x\times 2y\times 2z}$6 resolution, with manual trimaps produced in OpenCV and GIMP under expert guidance; an unlabeled augmentation set of 240 images supports the unsupervised experiment (Tamvakis et al., 2022). Arabidopsis root segmentation relies on a much smaller dataset of 14 plant photographs, with an internal inconsistency stating both 3 plants of different genetic backgrounds and 9 plants per photograph, and with the first image used for parameter tuning and excluded from evaluation (Orlando et al., 2017). These contrasts make clear that PSS must handle both large benchmark corpora and tiny expert-built datasets.

Annotation efficiency emerges as a recurring theme. Stem segmentation on tomato introduces a control-point-based Point-Generated Ground Truth (PGT) workflow: 4–5 control points define a B-spline centerline,

$S \in B^{2x\times 2y\times 2z}$7

which is then dilated to thickness $S \in B^{2x\times 2y\times 2z}$8 to form a mask (Yang et al., 2020). The average labeling time falls from 224 s/image for Detailed Ground Truth to 27 s/image for PGT (Yang et al., 2020). Eff-3DPSeg pursues an even sparser route: after a $S \in B^{2x\times 2y\times 2z}$9 down-sampling stage, only 50, 100, or 200 labeled points per cloud are retained, corresponding to about 0.5% annotation, while self-supervised pretraining supplies structure-aware features (Luo et al., 2022). SEGMENT3D distributes 3D image tiles to multiple annotators, supports either segmentation from scratch or correction of pre-segmentation, and merges tile results with STAPLE-based consensus (Spina et al., 2017).

Synthetic and AI-assisted supervision are equally prominent. Deep Leaf Segmentation Using Synthetic Data renders 10,000 labeled images per synthetic regime and shows that mixing real and synthetic Arabidopsis data yields the best mean score across the CVPPP sets (Ward et al., 2018). MRI root segmentation generates 384 artificial MRI–ground truth pairs for training and 384 for validation because voxelized labels from real reconstructions do not align well enough with raw MRI for direct supervised learning (Uzman et al., 2019). Switchgrass in PRMI uses technician-refined U-Net masks (Xu et al., 2022). A reasonable interpretation is that PSS should treat synthetic generation, pre-annotation, and human correction as first-class annotation modes rather than as secondary conveniences.

One further lesson is methodological rather than purely operational: controlled acquisition can be annotation reduction. In seed analysis, the blue background and non-overlapping arrangement are so effective that segmentation becomes a rapid deterministic plugin stage (Vale et al., 2020). In Plant Doctor, DeepSORT and ROI-quality scoring select the best crop before damage segmentation (Marques et al., 2024). This suggests that PSS should expose acquisition protocols alongside models, because in several regimes standardized capture contributes more to annotation efficiency and segmentation reliability than architectural sophistication alone.

5. Evaluation regimes and benchmark behavior

Evaluation in the PSS literature is highly task-dependent, and the choice of metric often encodes what counts as a successful segmentation. Root skeleton recovery is measured by

$$I_c = 255 \left( \frac{I - \min\{I\}}{\max\{I\} - \min\{I\}} \right),$$0

where $$I_c = 255 \left( \frac{I - \min\{I\}}{\max\{I\} - \min\{I\}} \right),$$1 is the predicted skeleton and $$I_c = 255 \left( \frac{I - \min\{I\}}{\max\{I\} - \min\{I\}} \right),$$2 the reference contour (Orlando et al., 2017). On the Arabidopsis scanner dataset, the method reaches average $$I_c = 255 \left( \frac{I - \min\{I\}}{\max\{I\} - \min\{I\}} \right),$$3, while the second human observer reaches $$I_c = 255 \left( \frac{I - \min\{I\}}{\max\{I\} - \min\{I\}} \right),$$4, making the method useful as a baseline but clearly below human agreement (Orlando et al., 2017). MRI root segmentation, by contrast, adopts a Distance Tolerant F1-Score because low resolution and annotation misalignment make exact voxel overlap too strict (Uzman et al., 2019). This is a strong indication that PSS should support tolerance-aware and topology-aware evaluation, especially for thin structures.

Speed and workflow metrics are equally important in controlled settings. The ImageJ seed plugin processes images in 0.02 s per image versus 63 s for the manual dual-image method, and acquisition time for about 100 seeds drops from 96.5 s for double-image capture to 59.9 s for the single-image blue-background workflow (Vale et al., 2020). Stem segmentation shows that cheap labels can be practically competitive: Mask R-CNN reaches F1 85.9 and Precision 93.1 on Detailed Ground Truth, while PGT-trained Mask R-CNN attains Precision 96.3 with respect to DGT, despite the approximate annotation mechanism (Yang et al., 2020).

On larger 2D datasets, the metric patterns reveal the difficulty of minority structures. In PRMI, supervised U-Net results vary widely by species, from IoU 4.8%, F1 0.092 on Cotton-150 to IoU 61.9%, F1 0.765 on Peanut-120, while IRNet-based weak supervision collapses across the board, for example IoU 2.2%, F1 0.043 on Switchgrass-300 (Xu et al., 2022). Grapevine leaf phenotyping reports strong global metrics but weak vein recovery: the best supervised model reaches PA 0.95 and MeanIoU 0.75, yet classwise vein IoU peaks at only 0.31, whereas blade IoU reaches 0.93 (Tamvakis et al., 2022). The lesson is not merely that the tasks are hard; it is that global metrics can conceal failure on the biologically salient fine structure.

Field hierarchical segmentation establishes stronger task-specific baselines. The joint semantic–plant–leaf model reaches, on SugarBeets, IoU 79.3, PQ$$I_c = 255 \left( \frac{I - \min\{I\}}{\max\{I\} - \min\{I\}} \right),$$5 76.2, PQ$$I_c = 255 \left( \frac{I - \min\{I\}}{\max\{I\} - \min\{I\}} \right),$$6 63.5, with 2.4M parameters and 26.3 FPS; on GrowliFlower it reaches IoU 80.2, PQ$$I_c = 255 \left( \frac{I - \min\{I\}}{\max\{I\} - \min\{I\}} \right),$$7 89.2, PQ$$I_c = 255 \left( \frac{I - \min\{I\}}{\max\{I\} - \min\{I\}} \right),$$8 71.0, at 20.7 FPS (Roggiolani et al., 2022). The SAM-based hierarchical panoptic system obtains 81.33 PQ+ on the CVPPA competition (Nguyen et al., 2023). ZeroPlantSeg then shows that zero-shot hierarchy construction can outperform supervised cross-domain transfer on GrowliFlower and SB20, even if in-domain supervised methods remain competitive on PhenoBench (Xing et al., 11 Sep 2025). This sequence suggests that PSS evaluation should distinguish in-domain optimization from cross-domain robustness, because the best trained model and the best deployment model need not coincide.

Three-dimensional point-cloud evaluation adds another dimension. PointsToWood reports strong transfer across ecosystems and sensors, for example balanced accuracy improving from 0.840 to 0.952 on the Cameroonian external dataset and from 0.800 to 0.929 on the German dataset relative to retrained FSCT, with especially strong gains in recall (Owen et al., 6 Mar 2025). OmniPlantSeg shows that full-resolution KD-SS + DGCNN can be competitive—pepper reaches MIoU 95.6, Acc 98.4—but also reveals hard failure modes such as wheat stem recall 0.02 and stem F1 0.05 (Gilson et al., 25 Sep 2025). A fair reading is that retaining full resolution does not by itself solve class imbalance or weak modality cues, but it preserves the geometric detail needed for downstream phenotyping even when metrics are only competitive rather than dominant.

Finally, fuzzy supervision for UHR remote sensing introduces a different evaluation regime. On SixP, classical hard-mask segmentation gives OA 91.49% but only F1-score 41.55 and $$I_c = 255 \left( \frac{I - \min\{I\}}{\max\{I\} - \min\{I\}} \right),$$9, reflecting strong background dominance (Pande et al., 2024). Under Gaussian-refined soft targets, cosine-similarity loss gives the best SixP fuzzy scores, while cross-entropy is best on the Weed dataset (Pande et al., 2024). The evidence is more qualitative than unified, but it reinforces a general principle: when labels are spatially uncertain, PSS should not assume that hard-mask IoU alone is a meaningful training or validation objective.

6. Applications, constraints, and future directions

The practical value of PSS lies in what segmentation enables after the mask or point labels are produced. In roots, downstream goals include main-root length, number of lateral roots, curvature, distances between initiation points, and gravitropic response angles (Orlando et al., 2017). In seed analysis, segmentation is explicitly a prerequisite for morpho-colorimetric measurements used in plant systematics and germplasm characterization (Vale et al., 2020). In grapevine leaves, the intended outputs include leaf width, leaf length, number of superior lateral veins, blade area surface, and vein-to-area ratio, while broader ampelographic traits are identified as plausible extensions (Tamvakis et al., 2022). Stem masks serve as a stable reference structure for wilting analysis in tomato (Yang et al., 2020). Leaf damage masks in Plant Doctor are converted into per-leaf damage ratios through a two-pass workflow estimating total leaf area and damaged area (Marques et al., 2024). In 3D phenotyping, Eff-3DPSeg extracts stem diameter, leaf width, and leaf length from organ-level point-cloud segmentation (Luo et al., 2022), and PointsToWood frames wood–leaf separation as upstream of tree architecture, productivity, and biomass analysis (Owen et al., 6 Mar 2025).

The same literature is equally clear that no single segmentation engine is universally adequate. Controlled scanner seeds can be solved by background design and color rules (Vale et al., 2020); roots in flatbed scans benefit from morphology and CRF refinement (Orlando et al., 2017); grape veins remain difficult even with supervised U-Nets (Tamvakis et al., 2022); generic CAM-based weak supervision is poor for sparse minirhizotron roots (Xu et al., 2022); heavy leaf overlap defeats simple instance logic and motivates hierarchical skips or explicit edge semantics (Bell et al., 2019, Roggiolani et al., 2022); and point-cloud organ labels depend strongly on sensor modality, class balance, and whether full-resolution geometry is preserved (Owen et al., 6 Mar 2025, Gilson et al., 25 Sep 2025). PSS should therefore be understood not as a single “best model,” but as a policy layer for selecting, composing, and validating methods under regime-specific assumptions.

Several controversies and common misconceptions emerge. One is the belief that high aggregate accuracy implies useful phenotyping. The vine-leaf results show the opposite: blade segmentation can be reliable while vein segmentation remains weak, limiting the trustworthiness of vein-based traits (Tamvakis et al., 2022). Another is the assumption that more automation always dominates careful acquisition design. The seed plugin shows that a controlled blue background and batch capture protocol can eliminate entire classes of segmentation error without needing a more complex model (Vale et al., 2020). A third is the idea that exact labels are always necessary. Sparse points, spline control points, AI-refined masks, synthetic scenes, and collaborative correction all contradict that view, though the quality of the resulting supervision remains task- and organ-dependent (Spina et al., 2017, Ward et al., 2018, Uzman et al., 2019, Yang et al., 2020, Luo et al., 2022).

Future directions are strongly implied by the most papers. Zero-shot hierarchical segmentation via SAM, OVSeg, and Grounding DINO indicates that PSS can use foundation models for annotation bootstrapping and cross-domain deployment when labeled data are absent (Xing et al., 11 Sep 2025). The SAM-based crop–weed system shows that prompt-based panoptic pipelines can also be high-performing in supervised competition settings (Nguyen et al., 2023). Species-agnostic, full-resolution point-cloud segmentation across modalities suggests that preprocessing modules like KD-SS may become generic infrastructure for 3D phenotyping, especially where point loss is unacceptable (Gilson et al., 25 Sep 2025). Complete-canopy wood–leaf separation across European forests shows that reflectance-aware but reflectance-robust architectures are now practical at ecological scale (Owen et al., 6 Mar 2025). Fuzzy-loss training for UHR remote sensing implies that PSS should support both hard and uncertainty-aware supervision modes when masks are derived from field geometry rather than exact tracing (Pande et al., 2024).

Taken together, these works imply that Plant Segmentation Studio is most credible when framed as a modality-aware, annotation-aware, and hierarchy-aware research platform. Its core function is not merely to draw boundaries, but to translate heterogeneous plant observations into validated structural representations suitable for phenotyping, benchmarking, and iterative model improvement across 2D, 3D, and temporally evolving plant datasets.