FloraSyntropy Archive: AI Plant Disease Benchmark

• FloraSyntropy Archive is a comprehensive, heterogeneous benchmark integrating 178,922 images of 35 plant species and 97 disease classes from 13 public repositories.
• The archive enforces a strict hold-out evaluation with balanced data splits, ensuring robust multiclass plant disease diagnosis and validation of federated learning methods.
• Beyond image data, the archive conceptually extends to cross-modal flora information, supporting scalable integration of LiDAR, 3D reconstructions, and network-based ecological data.

FloraSyntropy Archive denotes, in its most explicit formulation, a large-scale benchmark for AI-powered plant disease diagnosis comprising 178,922 images across 35 plant species, annotated with 97 distinct disease classes, including healthy and diseased states, and aggregated from 13 public repositories published between 2018 and 2023 (Khan et al., 25 Aug 2025). It was introduced to address the generalization gap between laboratory-scale image classification and heterogeneous agricultural reality by standardizing multi-source imagery, enforcing a strict hold-out evaluation regime, and providing a common benchmark for multiclass diagnosis. In related literature summaries, the same name is also invoked as an archival or data-warehouse concept for harmonized forest LiDAR products, 3D plant structural reconstructions, metabolite-species networks, and topology-aware ecological group discovery. This suggests a broader conception of FloraSyntropy Archive as a cross-modal flora data infrastructure rather than only an image corpus (Vautier et al., 30 Jun 2026, Rudskiy et al., 2012, Takemoto, 2011, Chen et al., 2024).

1. Core identity and dataset composition

The archive was introduced as a large-scale dataset for plant disease diagnosis with 178,922 images, 35 plant species, and 97 classes. It integrates healthy and diseased conditions and includes classes such as Apple-BlackRot, Rice-Blast, Soybean-Caterpillar, Tomato-SpiderMites, and Tea-Anthracnose. The source material was compiled from 13 public repositories, including datasets such as Plant Village, Plant Village V2, Cassava, BananaLSD, Coffee, Soybean, Tea, and Sugarcane (Khan et al., 25 Aug 2025).

Aspect	Value	Notes
Total images	178,922	Large-scale benchmark
Plant species	35	Multi-crop coverage
Disease classes	97	Healthy and diseased classes
Source datasets	13 public repositories	Published 2018–2023
Image size	224 × 224	Standardized input
Data split	70:10:20	Train, validation, test

The archive is characterized as extremely heterogeneous. It contains images from multiple plant organs such as leaves and fruit, spans fungal, bacterial, viral, pest-related, and healthy conditions, and includes both real field and lab photos with varied backgrounds and acquisition conditions. The stated aim is robust global representation and diversity, with variation in geographies, lighting conditions, species, and phenotypes rather than restriction to a single crop or controlled imaging environment (Khan et al., 25 Aug 2025).

A defining feature is its balancing strategy. The training split generally contains about 4,700 images per class, and validation sets about 500 images per class, with balancing achieved via augmentation. This design is central to the archive’s benchmark role, because the dataset is intended not merely as a collection of images but as a standardized evaluation substrate for models that claim cross-species and cross-condition generalization (Khan et al., 25 Aug 2025).

2. Curation, annotation, and partitioning

Dataset construction proceeded by aggregating and standardizing images from open datasets published between 2018 and 2023. The archive uses the original labels provided by the source datasets, with each image mapped to one of 97 classes encoding disease type and plant species. All images were resized to 224 × 224 pixels to ensure model-input consistency, and augmentation procedures, including rotation and flipping, were applied to improve class balance and visual diversity (Khan et al., 25 Aug 2025).

The partitioning protocol is explicit: 70% train, 10% validation, 20% test. The test set is described as a strict hold-out, remaining completely untouched during development in order to provide an unbiased final evaluation. This is a consequential design choice, because many earlier plant disease studies were evaluated on small or homogeneous datasets under conditions that can inflate performance estimates. FloraSyntropy Archive formalizes itself as a corrective to that pattern by embedding the evaluation split into the archive definition itself (Khan et al., 25 Aug 2025).

The curation logic also reflects an explicit rejection of narrow dataset scope. Earlier datasets are described as often single-species or single-crop focused, frequently collected under controlled conditions or uniform backgrounds, and therefore not representative of global agricultural practice. By contrast, FloraSyntropy Archive is described as supporting true assessment of whether a model can generalize across crops, disease types, image conditions, and world regions (Khan et al., 25 Aug 2025).

3. Benchmarking function and evaluation protocol

FloraSyntropy Archive serves as the definitive testbed for multiclass plant disease detection models. All models are trained and evaluated under a consistent protocol comprising 224 × 224 input resolution, the 70:10:20 split, and a fixed set of metrics: Accuracy, Precision, Recall, F1-score, together with ROC, PR curves, and t-SNE for feature-space separability (Khan et al., 25 Aug 2025).

The benchmark is accompanied by an explicit formalization of the dataset-generalization problem:

$$\text{Performance}_E = f(A, Parm, DS), \quad DS \sim P_{DS}$$

and

$f(A, Parm, S_{DS}) \gg f(A, Parm, P_{DS})$

The first relation defines expected performance as a function of architecture, parameters, and dataset sampled from the true data distribution; the second encodes the claim that performance on a small dataset can substantially exceed performance on the real-world distribution. Within the archive’s framing, these expressions justify why heterogeneous, globally sourced data are required for credible benchmarking rather than optional dataset enlargement (Khan et al., 25 Aug 2025).

A wide range of baseline architectures was evaluated, including DenseNet121, DenseNet169, DenseNet201, InceptionV3, MobileNetV1/V2, ResNet50V2/101V2/152V2, VGG16/19, as well as SVM and a hybrid autoencoder. The best previous models, particularly DenseNets and ResNets, achieved accuracy of around 93–94% on the benchmark. FloraSyntropy-Net achieved 96.38% accuracy, establishing the reported state of the art on this archive, while out-of-domain evaluation on the unrelated multiclass Pest 2022 dataset yielded 99.84% accuracy (Khan et al., 25 Aug 2025).

An important clarification follows from these numbers: the archive is the benchmark and resource, whereas 96.38% and 99.84% are model results obtained on and beyond that benchmark. The paper’s contribution therefore has two layers: a new archive that exposes a performance gap, and a specific learning framework that partially closes it (Khan et al., 25 Aug 2025).

4. Relationship to FloraSyntropy-Net and federated learning

The archive is tightly coupled to FloraSyntropy-Net, a federated learning framework introduced alongside it. The model integrates a Memetic Algorithm (MAO) for optimal base-model selection, selecting DenseNet201, together with a novel Deep Block for enhanced feature representation and a client-cloning strategy for scalable, privacy-preserving training. In this formulation, the archive is not only a benchmark but also the substrate on which a particular federated systems design is validated (Khan et al., 25 Aug 2025).

Federated benchmarking was conducted using five clients, each trained on unique, balanced subsets of the archive. Client models were then aggregated with weighted FedAvg, formalized as

$$w^{(t+1)} = \sum_{k=1}^{m} \frac{n_k}{n} \cdot w_k^{(t+1)}.$$

This makes the archive relevant to two technical questions at once: centralized multiclass diagnosis and distributed training under heterogeneous but balanced partitions. The study presents this as the first use of federated learning at this scale in plant disease diagnosis (Khan et al., 25 Aug 2025).

The archive also supports interpretability and diagnostic analysis. Reported tools include GRADCAM visualizations, confusion matrices, and t-SNE projections to inspect discriminative capacity and feature learning. Thus, the benchmark is not restricted to scalar leaderboard comparisons; it is designed to expose representation quality, class confusions, and cross-domain transfer behavior (Khan et al., 25 Aug 2025).

5. Archival extension beyond plant disease images

Related studies use the phrase FloraSyntropy Archive to denote, or to anticipate, a broader archival infrastructure into which non-image plant and ecosystem data could be integrated. In the national forest monitoring setting, FLORA—Forest LiDAR Octree Regression with Auxiliary Data—is described as compatible with data warehouses like a FloraSyntropy Archive, thereby facilitating reproducible, open, and harmonized national forest attribute time series and maps. FLORA predicts dominant height, total volume, deciduous volume, coniferous volume, basal area, and stem density from heterogeneous airborne LiDAR point clouds and auxiliary variables, and its modular design is described as allowing seamless updates with new contextual layers such as climate, disturbance, and time series (Vautier et al., 30 Jun 2026).

In plant structural biology, the 3D reconstruction protocol of Rudskiy and Khodorova is explicitly described as incorporable into archival infrastructures like the FloraSyntropy Archive. The relevant properties are device and software independence, preservation of data in standardized formats such as .tiff, .svg, and .ai, and support for scalable annotation, measurement, and metadata addition. The same summary proposes archive-ready representations including vector contours, genealogy trees as graphs, and mesh topologies, extending the notion of an archive from pixel data to spatial, topological, and developmental information (Rudskiy et al., 2012).

In computational chemotaxonomy, the study of flavonoid distributions across species identifies practical applications for a FloraSyntropy Archive in network-based curation, predictive modeling of missing or unobserved metabolite-species associations, automated clustering, hierarchical taxonomic organization, evolutionary tracing, and hub identification. Here the archive is not an image repository but a structured network resource centered on nested structure, modular structure, and heterogeneous connectivity in species–metabolite bipartite graphs (Takemoto, 2011).

In soil microbiome analysis, gFlora is presented as a topology-aware method whose explicit treatment of network structure would allow an archive to catalog interaction-rich functional teams across ecosystems and soil types. The method models the soil community as an ecological co-occurrence network, applies graph convolution, and outputs not only selected taxa but also a network of important taxa with node-level and pairwise importance. This extends the archive idea toward storing group-level functional response patterns rather than only sample-level abundance tables (Chen et al., 2024).

A plausible implication is that FloraSyntropy Archive functions, across these summaries, as a unifying archival concept for heterogeneous flora-associated modalities: 2D RGB images, LiDAR point clouds, 3D anatomical reconstructions, bipartite metabolite networks, and ecological co-occurrence graphs. The cited works do not define a single finalized schema, but they consistently point toward interoperable, metadata-rich, and analysis-ready plant data storage.

6. Significance, limitations, and prospective directions

The central significance of FloraSyntropy Archive lies in its role as an explicit answer to the generalization gap in agricultural AI. The archive is designed to distinguish models that perform well on small, homogeneous datasets from those that remain robust under large-scale, heterogeneous real-world diversity. Its benchmark role is therefore methodological as much as empirical: it changes what counts as a credible result in plant disease diagnosis by making cross-crop, cross-condition, and cross-domain variation part of the default evaluation problem (Khan et al., 25 Aug 2025).

Its anticipated impact is correspondingly broad. The archive is presented as a platform for building and benchmarking next-generation plant disease diagnosis models, for developing trustworthy, interpretable, and explainable AI for agriculture, and for advancing federated learning research under privacy and data sovereignty constraints. The paper also frames it as a resource for global food security through rapid, automated identification of plant diseases in highly diverse field conditions (Khan et al., 25 Aug 2025).

At the same time, the current record distinguishes clearly between what is already instantiated and what remains prospective. The fully specified implementation is the plant disease image archive and its benchmark protocol. By contrast, the broader archival vision—forest LiDAR warehousing, community-curated 3D reconstruction, metabolite-network completion, and topology-aware microbiome group cataloguing—appears in related summaries as compatibility claims, proposed integrations, or natural extensions rather than as a single deployed multimodal system (Vautier et al., 30 Jun 2026, Rudskiy et al., 2012, Takemoto, 2011, Chen et al., 2024).

Future development is therefore best understood in two layers. First, within plant disease diagnosis, the archive supports continued work on generalization, interpretability, and resistance to domain shift. Second, across the wider flora-informatics landscape, the existing proposals suggest expansion toward archives that preserve not only labels and images but also provenance, topology, geometry, developmental history, ecological relationships, and time series. If realized, that expansion would convert FloraSyntropy Archive from a benchmark-centered dataset into a genuinely integrative flora data infrastructure.