TomatoMAP: Advanced Tools in Tomato Analysis

• TomatoMAP is a suite of advanced tools for high-fidelity tomato phenotyping, field-scale crop mapping, and topological clustering, uniting diverse methodologies under one framework.
• It employs state-of-the-art techniques such as IoT-enabled multi-angle imaging, YOLOv11, Mask R-CNN, and U-Net architectures with composite loss functions to achieve robust, automated analysis.
• The framework supports scalable analysis in agriculture, enabling rapid, interpretable insights for breeding programs, crop system inventories, and multi-modal data clustering.

TomatoMAP is a designation shared by multiple advanced resources and methodologies in the domains of plant phenotyping, remote sensing, and topological data analysis, each focused on high-fidelity, large-scale mapping and modeling of tomato (Solanum lycopersicum) systems. Distinct implementations exist under this name, including a multi-purpose fine-grained imaging dataset for phenotyping, a field-scale crop mapping pipeline based on geospatial deep learning, and a multi-parameter topological clustering algorithm for complex structured data. The term "TomatoMAP" is therefore not a single algorithm or dataset but a family of leading-edge tools underpinning state-of-the-art research in computer vision, geospatial analytics, and topological machine learning.

1. TomatoMAP for Fine-Grained Phenotyping

TomatoMAP (Zhang et al., 15 Jul 2025) is an extensive dataset explicitly designed for fine-grained tomato phenotyping using multi-angle, multi-pose imaging and supports advanced model development for automated crop analysis. The acquisition protocol involves an IoT-driven imaging station equipped with a turntable and four synchronized color CMOS 5 MP camera modules (three 90° FOV and one 170° fisheye), capturing 12 discrete plant poses per camera covering under, side, three-quarter, and top views. Over 163 days, 32 imaging sessions, and 101 plants, this results in 64,464 RGB images for analysis.

Three nested annotation schemas are provided:

• Classification (TomatoMAP-Cls): Every image is labeled using a 50-class modified BBCH growth stage scale, ultimately grouped into three coarser classes for model development, but retaining all original 50 labels for validation.
• Object Detection (TomatoMAP-Det): Seven biologically important regions of interest are annotated with axis-aligned bounding boxes: whole plant, leaf, panicle, batch of flowers, batch of fruits, shoot, and axillary shoot. Annotation was performed through a progressive, AI-assisted approach using YOLOv11 models iteratively refined via expert corrections.
• Semantic and Instance Segmentation (TomatoMAP-Seg): A 3,616-image high-resolution subset is provided with pixel-wise semantic and instance segmentation masks for ten flower and fruit development categories, enabling precise object boundary modeling.

Rigorous stratified splitting ensures robust training, validation, and external testing (80% train, 10% validation, 10% test), with comprehensive metadata tracked per image.

2. Model Baselines and Assessment

A cascading deep learning framework is established using state-of-the-art architectures:

• Classification: MobileNetV3-Large, trained with cross-entropy loss and cosine learning rate decay, achieves 79.19% accuracy. Errors predominantly occur between adjacent BBCH stages, consistent with phenotypic ambiguity at stage boundaries.
• Object Detection: YOLOv11-Large (CSPDarknet backbone), 640×640 inputs, attains mAP₀.₅ = 0.92 on validation, with per-class AP spanning 0.88 (fruit clusters) to 0.96 (leaf, whole plant).
• Semantic/Instance Segmentation: Mask R-CNN (ResNet-50-FPN backbone), after a hyperparameter sweep, yields best AP₅₀ = 63.59.

Evaluation employs standard metrics:

• Classification: $$\mathrm{Accuracy} = \frac{TP + TN}{TP + TN + FP + FN}$$
• Detection: $\mathrm{mAP}_{0.5} = \frac{1}{C} \sum_{c=1}^C AP_c$
• Segmentation: $\mathrm{IoU}(A,B) = \frac{|A \cap B|}{|A \cup B|}$
• Consistency: Cohen's Kappa $\kappa = \frac{p_o - p_e}{1-p_e}$

Extensive AI-vs-human analysis with five experts demonstrated "almost perfect" agreement (κ ≈ 0.83–0.88 for AI vs. human; 0.85–0.92 for human vs. human) and that automated inference is three orders of magnitude faster than manual annotation, with comparable quality.

3. Field-Scale Tomato Crop Mapping via Geospatial Embeddings

The TomatoMAP system in (Narimani et al., 20 May 2026) implements deep-learning-based, field-scale mapping of processing tomato systems using Google DeepMind's AlphaEarth 64-band geospatial embeddings at 10 m resolution. The workflow is as follows:

• Polygon labels: 9,484 fields are extracted from LandIQ 2018 (4,742 tomato, 4,742 non-tomato across multiple crop types).
• Each polygon's bounding box is clipped from the annual AlphaEarth embedding image, associated with a precise binary mask, and spatially partitioned into train/val/test splits to ensure independence.

A U-Net segmentation model (64×H×W input, base width 32 channels, 4-level encoder/decoder with skip connections, per-pixel sigmoid) is trained to optimize a composite loss: $L = 0.5 \,L_{BCE} + 0.5\,L_{Dice}$, with $L_{BCE}$ the masked binary cross-entropy and $L_{Dice}$ the soft Dice loss. All training and inference leverage masked valid pixels only.

Monte Carlo (MC) dropout is activated during inference (T=100 forward passes), yielding:

• Pixelwise predictive mean: $\hat{p}(x) = \frac{1}{T} \sum_{t=1}^T p^{(t)}(x)$
• Predictive variance: $$Var(x) = \frac{1}{T} \sum_{t=1}^T [p^{(t)}(x) - \hat{p}(x)]^2$$

Performance on the 1,424-field test set:

• Pixel accuracy: 99.19%
• Precision: 98.69%
• Recall: 99.40%
• F1: 99.04%
• IoU: 98.11%
• Chip accuracy: 99.02%

Uncertainty is highest at field boundaries, as expected from label ambiguity and mixed pixel composition. No manual feature engineering or time-series construction is required—AlphaEarth embeddings encode crop-phenological and management context directly, allowing robust year-to-year transfer.

4. ToMAToMP: Multi-Parameter Topological Clustering

ToMAToMP (Andrianirina et al., 14 May 2026) (sometimes stylized as "TomatoMAP" in this context) is a topological data analysis (TDA) algorithm for clustering based on persistent homology of multiple functions applied to an underlying dataset $X$ (finite point cloud or abstract space). Key concepts include:

• Single-parameter ToMATo: Computes the 0-dimensional persistence diagram $\mathrm{mAP}_{0.5} = \frac{1}{C} \sum_{c=1}^C AP_c$0 for a function $\mathrm{mAP}_{0.5} = \frac{1}{C} \sum_{c=1}^C AP_c$1, clusters connected components by mode prominence.
• Multi-parameter extension: Multiple functions $\mathrm{mAP}_{0.5} = \frac{1}{C} \sum_{c=1}^C AP_c$2 are combined, with filtrations computed along diagonal slices $\mathrm{mAP}_{0.5} = \frac{1}{C} \sum_{c=1}^C AP_c$3 in $\mathrm{mAP}_{0.5} = \frac{1}{C} \sum_{c=1}^C AP_c$4, yielding slice functions $\mathrm{mAP}_{0.5} = \frac{1}{C} \sum_{c=1}^C AP_c$5.
• MMA decomposition: Chains of persistence intervals ("bars") across slices are matched using vineyard algorithms, each chain corresponding to an axis-aligned box in $\mathrm{mAP}_{0.5} = \frac{1}{C} \sum_{c=1}^C AP_c$6; these boxes define the clusters.

The algorithm builds the MMA decomposition, applies classical ToMATo to each slice, aggregates clusterings via majority voting, and achieves robustness without manual neighborhood graph tuning. Theoretical results guarantee stability under perturbations and resilience to outliers given suitable function separation.

Empirical results show that ToMAToMP achieves superior clustering quality (e.g., Adjusted Mutual Information up to 0.93 in synthetic benchmarks) and outlier robustness compared to k-means, spectral, and single-function topological baselines, at the cost of increased computation in the MMA step.

5. Methodological Innovations and Interpretability

The TomatoMAP resources exemplify methodological innovation across diverse domains:

• Dataset and Annotation Protocols: The TomatoMAP phenotyping resource demonstrates the efficacy of integrated hardware (multi-angle, multi-pose IoT imaging) and AI-assisted, iterative annotation that significantly reduces labor while improving accuracy and inter-rater agreement.
• Modeling Approaches: The use of composite loss functions (e.g., BCE + Dice), MC dropout for epistemic uncertainty estimation, and multi-stage deep learning pipelines enables both fine-grained object delineation and interpretable uncertainty quantification.
• Scalable Geospatial Mapping: Leveraging AlphaEarth embeddings for crop mapping removes the need for hand-crafted vegetation indices, supports spatially explicit cross-validation, and substantially improves inference robustness and scalability for agricultural applications.
• Topological Insights: ToMAToMP and its MMA decomposition enable principled multi-modal clustering, capturing complex joint structure in multi-feature spaces without the need for arbitrary parameter blending or graph tuning.

A plausible implication is that these methods enable automated, high-throughput phenotyping, robust spatial crop analytics, and structure-preserving clustering for datasets where traditional heuristics are limited by noise, high dimensionality, or ill-posed hyperparameter tuning.

6. Applications and Future Directions

TomatoMAP datasets and algorithms are deployed in:

• Automated phenotyping pipelines for breeding programs, assessing traits across developmental stages, and streamlining high-throughput, reproducible measurements (Zhang et al., 15 Jul 2025).
• Field-scale crop system inventories supporting supply-chain forecasts, policy, and management decisions; robust to annual variability without manual preprocessing (Narimani et al., 20 May 2026).
• Clustering of complex biological phenotypes, multi-modal gene expression or spatial data, and other applications requiring theoretically grounded, multi-feature structure discovery (Andrianirina et al., 14 May 2026).

Ongoing and future work includes:

• Extension of AlphaEarth-based mapping to new geographies and crop types, and to multi-year temporal transfer.
• Incorporation of RGB-D data, anchor redesign, and focal loss in robotic detection deployments (Magalhães et al., 2021).
• Acceleration of MMA computation and extension of ToMAToMP to very high-dimensional, multi-omic, or multi-view biological data.

The convergence of robust dataset engineering, advanced modeling architectures, and topologically informed unsupervised learning embodied in the TomatoMAP suite positions it as a critical resource suite for modern plant science, computer vision, and data-driven agriculture.