FloraSyntropy-Net: Plant-Centered AI

• FloraSyntropy-Net is a plant-centered distributed system that integrates federated deep learning with bio-digital communication for scalable plant disease diagnosis.
• Its innovative design leverages DenseNet201, optimized via Memetic Algorithm and a Novel Deep Block, to achieve significant accuracy gains and strong cross-domain generalization.
• The framework supports privacy-preserving training across heterogeneous datasets while paving the way for future integration with energy-harvesting, plant-inspired cyber-physical systems.

Searching arXiv for the cited FloraSyntropy-Net papers and related entries. FloraSyntropy-Net denotes a plant-centered systems concept that appears in two closely related but technically distinct forms in recent literature. In one formulation, it is a bio-digital infrastructure built atop the “Internet of Plants,” in which agent plants interface with biological communication channels and harvest energy by mimicking mycorrhizal symbiosis (Bilgen et al., 2024). In a later and more explicit machine-learning formulation, FloraSyntropy-Net is a federated deep learning framework for large-scale plant disease diagnosis, built around the FloraSyntropy Archive, DenseNet201 selected by Memetic Algorithm Optimization, a Novel Deep Block, and client-cloning with weighted Federated Averaging (Khan et al., 25 Aug 2025). Taken together, these usages position FloraSyntropy-Net at the intersection of plant communication networks, bio-inspired cyber-physical systems, and scalable agricultural AI.

1. Terminological scope and conceptual lineage

The term FloraSyntropy-Net is used most directly in the 2025 work “FloraSyntropy-Net: Scalable Deep Learning with Novel FloraSyntropy Archive for Large-Scale Plant Disease Diagnosis” (Khan et al., 25 Aug 2025). There it refers to a federated learning framework designed to address a specific generalization problem in plant disease recognition: models trained on small, homogeneous datasets often achieve high reported accuracy on narrow benchmarks while degrading substantially on the true real-world distribution of plant species and disease states. This concern is formalized as

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

with the additional observation that

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

The framework is therefore explicitly motivated by heterogeneous data, privacy constraints, and deployment across many plant species and institutions (Khan et al., 25 Aug 2025).

A broader systems interpretation is developed in the 2024 paper “Mycorrhizal Fungi and Plant Symbiosis for Energy Harvesting in the Internet of Plants” (Bilgen et al., 2024). Although that paper does not title its contribution FloraSyntropy-Net, the detailed exposition treats the “Internet of Plants” as a naturally occurring distributed information-and-resource network and describes an agent plant architecture that can monitor plants, communicate through multiple modalities, store nutrients and signaling compounds, and harvest energy by extracting glucose through mycorrhizal-like interaction (Bilgen et al., 2024). Within that interpretation, FloraSyntropy-Net is the intersection of a biological layer and a cyber-physical agent layer.

The coexistence of these two meanings is not a contradiction. Rather, it indicates that FloraSyntropy-Net has evolved from a bio-digital networking concept into a named learning framework. A plausible implication is that the later machine-learning usage inherits some of the earlier emphasis on heterogeneity, distributed operation, and multi-node coordination, but the two papers address different technical problems: one concerns plant-network communication and energy autonomy, while the other concerns federated visual diagnosis at scale (Bilgen et al., 2024, Khan et al., 25 Aug 2025).

2. FloraSyntropy-Net as a large-scale plant disease learning framework

In its 2025 formulation, FloraSyntropy-Net is a federated learning framework for large-scale, privacy-preserving plant disease diagnosis across heterogeneous data sources (Khan et al., 25 Aug 2025). Its stated purpose is to close the gap between performance on small-scale datasets and performance on globally diverse agricultural imagery. The framework is coupled to the FloraSyntropy Archive, a large-scale heterogeneous dataset assembled from 13 public repositories covering years 2018–2023, with 178,922 images across 35 plant species and 97 distinct disease classes, including healthy and diseased leaves (Khan et al., 25 Aug 2025).

The dataset design is central to the framework. Images are standardized to 224 × 224 pixels to ensure consistent feature extraction across the 13 source datasets, and class balancing is achieved through “advanced augmentation techniques including geometric transformations (balance dataset)” (Khan et al., 25 Aug 2025). The reported split is 70:10:20 for train, validation, and test, with 20% reserved as a hold-out test set and validation carved from the remaining data (Khan et al., 25 Aug 2025). The class structure explicitly distinguishes healthy from diseased states, with examples including Apple-BlackRot, Apple-Healthy, Rice-Blast, Rice-Healthy, Tomato-LateBlight, Tomato-Healthy, Tea-Healthy, and multiple classes for Cassava, Cotton, Corn, Grapes, Soybean, and Sugarcane (Khan et al., 25 Aug 2025).

The core claim of the framework is that generalization requires both a more representative dataset and an explicitly distributed training protocol. The architecture is therefore not presented as a standalone image classifier, but as a federated system in which multiple clients train local clones of a common global model under data heterogeneity and privacy constraints (Khan et al., 25 Aug 2025). This framing is particularly relevant for agricultural settings in which farms or institutions may not wish to share raw images.

3. Architectural composition: MAO, DenseNet201, Deep Block, and federated aggregation

FloraSyntropy-Net is built from four tightly coupled components: optimal backbone selection through Memetic Algorithm Optimization (MAO), a DenseNet201 backbone, a Novel Deep Block inserted after feature extraction, and client-cloning with weighted FedAvg aggregation (Khan et al., 25 Aug 2025).

MAO is used to choose the base architecture rather than fixing it a priori. The candidate population is defined as

$MP_0 = \{ mp_1, mp_2, mp_3, \dots, mp_{11} \},$

covering DenseNet variants, ResNet variants, VGG16/19, MobileNet, and InceptionV3 (Khan et al., 25 Aug 2025). The optimization objective is

$mp^* = \arg\max_{i} F(mp_i),$

where the fitness function is validation accuracy after a rapid fine-tuning cycle: $F(mp_i) = \frac{1}{D_{val} \sum_{j=1}^{D_{val}} I(\hat{c}_j = C_j).$ Selection uses tournament sampling,

$\text{Parent} = \arg\max_{m \in T} F(mp),$

while crossover and mutation are defined over architectural hyperparameters by

$hp_{\Phi} = M \Theta hp_1 + (1 - M) \Theta hp_2$

and

$$hp'_{\Phi}[i] = \begin{cases} hp_{\Phi}[i] + \epsilon & \text{with probability } p_m \ hp_{\Phi}[i] & \text{otherwise}. \end{cases}$$

After $G$ generations, the selected GlobalNet is

$\text{GlobalNet} = \arg\max_{m \in P_G} F(mp).$

The reported outcome is that DenseNet201 is the best-performing backbone on the FloraSyntropy Archive (Khan et al., 25 Aug 2025).

The DenseNet201 baseline is then augmented with a Novel Deep Block placed after feature extraction and before flattening or final classification. The described sequence is: Dense layer, ReLU activation, concatenation of original and transformed features, and a RepeatVector operation that converts a 1D feature vector into a sequence-like representation suitable for subsequent 1D layers (Khan et al., 25 Aug 2025). The block can be “wrapped in a recursive loop,” meaning the pattern may be applied multiple times. Formally, its core operations are expressed as

$f(A, Parm, S_{DS}) \gg f(A, Parm, P_{DS}).$0

$f(A, Parm, S_{DS}) \gg f(A, Parm, P_{DS}).$1

$f(A, Parm, S_{DS}) \gg f(A, Parm, P_{DS}).$2

$f(A, Parm, S_{DS}) \gg f(A, Parm, P_{DS}).$3

The intended effect is to preserve original features, enrich them through nonlinear transformation, and restructure them into a pseudo-sequence for deeper feature processing (Khan et al., 25 Aug 2025).

The federated component uses a central server maintaining global weights $f(A, Parm, S_{DS}) \gg f(A, Parm, P_{DS}).$4, with $f(A, Parm, S_{DS}) \gg f(A, Parm, P_{DS}).$5 clients in the reported experiments (Khan et al., 25 Aug 2025). Each client receives a clone of the global model, trains locally on its own data, and returns updated weights. Aggregation is weighted by local data volume: $f(A, Parm, S_{DS}) \gg f(A, Parm, P_{DS}).$6 The same principle is restated through the scaling-weight formulation

$f(A, Parm, S_{DS}) \gg f(A, Parm, P_{DS}).$7

and through an alternate gradient-form notation for global update (Khan et al., 25 Aug 2025). This explicit weighting is intended to handle non-IID client data and unequal client dataset sizes.

4. Training protocol, metrics, and empirical results

The reported training setup uses 30 epochs, Adam optimizer, learning rate $f(A, Parm, S_{DS}) \gg f(A, Parm, P_{DS}).$8, batch size 32, and categorical cross-entropy loss (Khan et al., 25 Aug 2025). The local objective is written as

$f(A, Parm, S_{DS}) \gg f(A, Parm, P_{DS}).$9

with update rule

$MP_0 = \{ mp_1, mp_2, mp_3, \dots, mp_{11} \},$0

Evaluation uses the standard definitions of accuracy, precision, recall, and F1 score: $MP_0 = \{ mp_1, mp_2, mp_3, \dots, mp_{11} \},$1

$MP_0 = \{ mp_1, mp_2, mp_3, \dots, mp_{11} \},$2

$MP_0 = \{ mp_1, mp_2, mp_3, \dots, mp_{11} \},$3

$MP_0 = \{ mp_1, mp_2, mp_3, \dots, mp_{11} \},$4

These metrics are computed per class and overall (Khan et al., 25 Aug 2025).

On the FloraSyntropy Archive, the full framework—DenseNet201 selected by MAO, Novel Deep Block, federated client-cloning, and weighted FedAvg—achieves 96.38% accuracy, with macro precision 0.9525, recall 0.9409, and F1-score 0.9504 (Khan et al., 25 Aug 2025). This exceeds the centralized DenseNet201 baseline, which reports accuracy 93.73%, precision 0.8986, recall 0.8864, and F1 0.8838 (Khan et al., 25 Aug 2025). The framework also outperforms other centralized baselines, including DenseNet121 at 93.06%, DenseNet169 at 93.18%, InceptionV3 at 93.34%, MobileNetV1 at 93.06%, MobileNetV2 at 92.83%, ResNet50V2 at 93.48%, ResNet101V2 at 92.48%, ResNet152V2 at 92.55%, VGG16 at 86.92%, and VGG19 at 85.28% (Khan et al., 25 Aug 2025).

The contribution of the Deep Block is reported explicitly. Without the Deep Block, FloraSyntropy-Net reaches 94.74% accuracy; with the Deep Block, it reaches 96.38%, an absolute gain of 1.64 percentage points (Khan et al., 25 Aug 2025). Class-level examples include Arjun-Diseased improving from 0.8261 to 0.9070 F1, Bael-Diseased from 0.8627 to 0.9200, and Potato-LateBlight from 0.8544 to 0.9879 (Khan et al., 25 Aug 2025). The ROC and precision-recall curves also improve, with AUC from approximately 0.999 to 1.000 and average precision from approximately 0.999 to 1.000 (Khan et al., 25 Aug 2025).

The reported cross-domain evaluation on the multiclass Pest dataset is especially notable. Without the Deep Block, overall accuracy is 0.9778; with the Deep Block, overall accuracy is 0.9984, or 99.84% (Khan et al., 25 Aug 2025). For Aphids, Armyworm, Beetle, Bollworm, Grasshopper, Mites, Mosquito, and Stem_borer, the paper reports precision, recall, and F1 of 1.0000 for all but Sawfly, which has F1 = 0.9929; the confusion matrix contains only one misclassification, Mosquito → Mites (Khan et al., 25 Aug 2025). The paper interprets this as exceptional cross-domain generalization from plant disease imagery to insect morphology.

5. Ablation structure, comparative interpretation, and limitations

The ablation analysis isolates both backbone selection and block design (Khan et al., 25 Aug 2025). For backbone choice, DenseNet201 is the strongest baseline, with reported accuracy 0.9373, while competing models range from 0.8692 to 0.9306 (Khan et al., 25 Aug 2025). This supports the MAO outcome that DenseNet201 should serve as the Global-Net backbone.

For feature-module comparison, the Novel Deep Block is evaluated against Residual, Inception, SE, CBAM, Naïve Inception, and Coordinate Attention blocks (Khan et al., 25 Aug 2025). The reported results are summarized below.

Block	Accuracy	F1
Residual Block	94.31%	0.8559
Inception Block	94.27%	—
SE	94.20%	—
CBAM	94.39%	—
Naïve Inception	94.48%	—
CA	94.40%	—
Novel Deep Block	94.74%	0.8704

The Novel Deep Block also reports precision 0.8925, recall 0.8709, and Cohen’s Kappa 0.9463, exceeding the corresponding values listed for the alternative blocks (Khan et al., 25 Aug 2025). Within the bounds of the reported study, this supports the claim that the block provides more effective feature enhancement than the compared generic residual, inception, and attention modules.

The paper also identifies clear limitations. First, federated training imposes computational and coordination overhead, especially when clients differ in hardware and network capability (Khan et al., 25 Aug 2025). Second, the framework depends on client data quality: if multiple clients contribute noisy or mislabeled data, weighted aggregation cannot fully mitigate the negative impact, and the paper states that there is currently no explicit robust data-quality filtering (Khan et al., 25 Aug 2025). Proposed future directions include a lightweight asynchronous federated protocol, robust data validation and quality control at client level, edge deployment through pruning and quantization, temporal models for disease progression, multimodal sensing, and unsupervised or few-shot learning for emerging diseases (Khan et al., 25 Aug 2025).

A common misconception would be to read the reported 96.38% and 99.84% as directly comparable to older single-crop studies reporting very high accuracy. The paper explicitly argues against that interpretation by emphasizing that many earlier methods were validated on small, homogeneous datasets, whereas FloraSyntropy-Net is evaluated on a 97-class global dataset and marked as “Large-Scale validation: Present” (Khan et al., 25 Aug 2025).

6. Bio-digital and energy-harvesting interpretation in the Internet of Plants

A separate but related usage of FloraSyntropy-Net arises from the 2024 Internet of Plants study (Bilgen et al., 2024). There, the relevant substrate is a naturally occurring distributed network formed by plants and symbiotic partners, especially mycorrhizal fungi. The network comprises multiple communication modalities: electrical signaling, acoustic signaling, VOC-based molecular signaling, and belowground chemical signaling and resource-sharing via common mycorrhizal networks (CMNs) (Bilgen et al., 2024).

These modalities are treated as communication channels. The electrical channel uses soil and leaf connections; the acoustic channel uses airborne sound and vibration; the molecular VOC channel uses diffusion and advection of chemical signals; and the mycorrhizal or CMN channel enables transfer of carbon, nitrogen, phosphorus, and defense-related signals (Bilgen et al., 2024). This yields a conceptual two-layer architecture. The biological layer is

$MP_0 = \{ mp_1, mp_2, mp_3, \dots, mp_{11} \},$5

where $MP_0 = \{ mp_1, mp_2, mp_3, \dots, mp_{11} \},$6 are biological nodes such as plants and fungi, and $MP_0 = \{ mp_1, mp_2, mp_3, \dots, mp_{11} \},$7 are biological communication links including hyphae, soil pathways, and air diffusion fields. The cyber-physical agent layer is

$MP_0 = \{ mp_1, mp_2, mp_3, \dots, mp_{11} \},$8

where $MP_0 = \{ mp_1, mp_2, mp_3, \dots, mp_{11} \},$9 are agent plants and $mp^* = \arg\max_{i} F(mp_i),$0 are electromagnetic, wired, and bio-chemical links (Bilgen et al., 2024).

The agent plant itself is defined as a synthetic, plant-like node whose mission is to communicate with surrounding plants, monitor health and stress, form “agent plant–plant highways” for signal and molecule delivery, store and distribute nutrients, drugs, and infochemicals, mimic mycorrhizal fungi–plant symbiosis to extract glucose for energy harvesting, and communicate with a central agent plant using EM-based methods (Bilgen et al., 2024). The architecture is divided into processor, communication, storage, and power units. The processor maintains a database of plants in its identified area and can be modeled by

$mp^* = \arg\max_{i} F(mp_i),$1

where $mp^* = \arg\max_{i} F(mp_i),$2 denotes interventions and $mp^* = \arg\max_{i} F(mp_i),$3 denotes observed electrical, acoustic, or VOC signals (Bilgen et al., 2024).

The power unit is especially significant in this interpretation. The paper proposes a respiration-based biocell inspired by mitochondrial respiration, supplied with glucose extracted from plants through mycorrhizal-like interaction (Bilgen et al., 2024). The chemical picture is summarized as

$mp^* = \arg\max_{i} F(mp_i),$4

with electrons routed through an external circuit and protons through nanochannels. Power density is expressed as

$mp^* = \arg\max_{i} F(mp_i),$5

The cited PET nanochannel biocell reports current density approximately $mp^* = \arg\max_{i} F(mp_i),$6, maximum power density approximately $mp^* = \arg\max_{i} F(mp_i),$7 at 0.35 V, and lifetime greater than 60 h (Bilgen et al., 2024). The cited SPEEK nanochannel biocell reports current density $mp^* = \arg\max_{i} F(mp_i),$8, maximum power density $mp^* = \arg\max_{i} F(mp_i),$9, and 8 days continuous operation (Bilgen et al., 2024). The paper states that these mW/cm$F(mp_i) = \frac{1}{D_{val} \sum_{j=1}^{D_{val}} I(\hat{c}_j = C_j).$0-scale figures are sufficient for low-power sensor and communication modules in the broader FloraSyntropy-Net concept.

The same paper develops a glucose-driven power model using plant-network parameters $F(mp_i) = \frac{1}{D_{val} \sum_{j=1}^{D_{val}} I(\hat{c}_j = C_j).$1, $F(mp_i) = \frac{1}{D_{val} \sum_{j=1}^{D_{val}} I(\hat{c}_j = C_j).$2, $F(mp_i) = \frac{1}{D_{val} \sum_{j=1}^{D_{val}} I(\hat{c}_j = C_j).$3, and $F(mp_i) = \frac{1}{D_{val} \sum_{j=1}^{D_{val}} I(\hat{c}_j = C_j).$4, together with biocell parameters $F(mp_i) = \frac{1}{D_{val} \sum_{j=1}^{D_{val}} I(\hat{c}_j = C_j).$5, $F(mp_i) = \frac{1}{D_{val} \sum_{j=1}^{D_{val}} I(\hat{c}_j = C_j).$6, and $F(mp_i) = \frac{1}{D_{val} \sum_{j=1}^{D_{val}} I(\hat{c}_j = C_j).$7 (Bilgen et al., 2024). Expected glucose extraction per interval is

$F(mp_i) = \frac{1}{D_{val} \sum_{j=1}^{D_{val}} I(\hat{c}_j = C_j).$8

and the expected enzymatic rate follows Michaelis–Menten kinetics: $F(mp_i) = \frac{1}{D_{val} \sum_{j=1}^{D_{val}} I(\hat{c}_j = C_j).$9 The expected glucose concentration is then

$\text{Parent} = \arg\max_{m \in T} F(mp),$0

and expected power density is modeled as

$\text{Parent} = \arg\max_{m \in T} F(mp),$1

Using experimental-aligned constants $\text{Parent} = \arg\max_{m \in T} F(mp),$2, $\text{Parent} = \arg\max_{m \in T} F(mp),$3, and $\text{Parent} = \arg\max_{m \in T} F(mp),$4, the analysis shows that increasing successful extraction probability $\text{Parent} = \arg\max_{m \in T} F(mp),$5, number of connected plants $\text{Parent} = \arg\max_{m \in T} F(mp),$6, and amount per event $\text{Parent} = \arg\max_{m \in T} F(mp),$7, while decreasing signaling interval $\text{Parent} = \arg\max_{m \in T} F(mp),$8, increases expected power output, subject to ecological constraints and plant health (Bilgen et al., 2024).

7. Synthesis, significance, and research directions

Across the two papers, FloraSyntropy-Net can be understood as a family of plant-centered distributed systems with two dominant instantiations. The first is a federated visual learning framework for heterogeneous, privacy-sensitive agricultural diagnosis at scale (Khan et al., 25 Aug 2025). The second is a bio-digital infrastructure in which agent plants exploit multi-modal plant communication and symbiotic glucose extraction to sustain sensing, signaling, and computation (Bilgen et al., 2024).

The 2025 formulation establishes FloraSyntropy-Net as a concrete machine-learning system with a defined dataset, training procedure, architecture search strategy, Deep Block enhancement, and data-proportional federated aggregation (Khan et al., 25 Aug 2025). Its principal technical significance lies in combining large-scale heterogeneous benchmarking with privacy-preserving collaborative learning and reporting strong results on both the 97-class FloraSyntropy benchmark and an unrelated pest dataset (Khan et al., 25 Aug 2025).

The 2024 formulation provides a broader cyber-biological systems context. It suggests a network model in which information flows through electrical, acoustic, VOC, and CMN channels, while energy flows from sunlight to plant photosynthesis, from plant glucose to respiration-based biocells, and then into sensor and communication modules (Bilgen et al., 2024). This suggests a possible long-term convergence between distributed agricultural AI and plant-integrated hardware platforms, although that convergence is not yet realized in a single unified implementation.

The combined literature also defines the main open problems. On the learning side, the stated priorities are asynchronous federated protocols, robust client-side data validation, edge deployment, multimodal sensing, and learning regimes for emerging diseases with limited labels (Khan et al., 25 Aug 2025). On the bio-digital side, the core issues are long-term stability and durability of respiration-based biocells, ecological limits on glucose extraction, and deployment models in which agent plants act as cooperators rather than unilateral consumers of plant resources (Bilgen et al., 2024). In that sense, FloraSyntropy-Net names not only a specific federated classifier, but also a wider research trajectory linking plant communication, distributed intelligence, and sustainable agricultural cyber-physical systems.