CoAtNet-DeepMoE: A Convolution-Attention Hybrid with DeepSeek Mixture-of-Experts for Parameter-Efficient Tomato Disease Classification

Md Nadim Mahamood; Md Arif Shahriar; Md Shafi Ud Doula; Kamrul Hasan

cs.CVSep 16, 2026

CoAtNet-DeepMoE: A Convolution-Attention Hybrid with DeepSeek Mixture-of-Experts for Parameter-Efficient Tomato Disease Classification

Authors: Md Nadim Mahamood, Md Arif Shahriar, Md Shafi Ud Doula, Kamrul Hasan

Organizations: Department of Computer Science and Engineering Begum Rokeya University Rangpur, Bangladesh · Department of Computer Science Texas State University San Marcos, Texas, USA · Department of Information and Communication Technologies Asian Institute of Technology Pathum Thani, Thailand

Abstract

The world population is growing rapidly, and technology is improving in parallel. Meeting the huge demand for food for these 7 billion people not only depends on increasing food production but also on reducing food loss. Crop losses due to disease affect both the food supply and the financial and economic stability of a country. Tomatoes are among the top food-producing crops globally, and a significant portion of this production is lost due to disease. People have used Machine Learning techniques for feature extraction and early diagnosis of tomato diseases, and nowadays, Deep Learning-based models are widely used for disease recognition. However, most existing models are highly parameter-intensive, which increases the time required for training and inference. As a result, while lightweight models are more suitable for user-friendly applications, they often show a reduction in performance. To balance performance and model size, we propose CoAtNet-DeepMoE, a Convolution-Attention hybrid architecture for rich feature extraction, further enhanced with a DeepSeek Mixture of Experts to substantially reduce the number of parameters without sacrificing accuracy. We evaluate our model on both balanced and imbalanced datasets from Kaggle and PlantVillage, demonstrating robustness and achieving 99.80% accuracy, 99.80% precision, 99.80% recall, and 99.80% F1-score on Kaggle, and 99.83% accuracy, 99.85% precision, 99.76% recall, and 99.80% F1-score on PlantVillage, representing state-of-the-art performance with only 2.47M parameters. The source code will be available at https://github.com/nadimbrur/CoAt-MoE.

Figures & tables

Figure 1: Overview of the proposed CoAtNet-DeepMoE architecture. The model integrates convolutional layers for local feature extraction, MbConv and SE blocks for channel-wise enhancement, and attention layers for global feature modeling. A parameter-efficient DeepSeek Mixture-of-Experts replaces the traditional MLP, where the router selects the top-k experts (

2

in this case) along with a shared expert for feature processing. The selected experts specialize in capturing disease-specific patterns, enabling the model to focus on the most relevant features for each input. The shared expert, on the other hand, preserves common information across all classes (e.g., common characteristics), ensuring stable and generalizable representations. This selective expert activation enhances feature representation while greatly reducing computational cost and overall parameter count (

26.67\,M

2.47\,M

\approx 90.73\%

Module	CoAt-Base	CoAt-DeepMoE
Stem	19,360	19,360
MbConv-0	236,768	83,648
MbConv-1	1,410,720	208,416
Transformer2D-2	12,145,374	437,174
Transformer2D-3	12,852,804	1,715,492
Classifier Head	9,226	9,226

Table 1: Block-wise parameter comparison between CoAt-Base and CoAt-DeepMoE.

Dataset	BS	EB	LB	LM	SLS	SE	TS	YLCV	MV	Healthy	Total
Kaggle	1100 (0.10)	1100 (0.10)	1100 (0.10)	1100 (0.10)	1100 (0.10)	1100 (0.10)	1100 (0.10)	1100 (0.10)	1100 (0.10)	1100 (0.10)	11000 (1.00)
PlantVillage	2127 (0.12)	1000 (0.06)	1909 (0.11)	952 (0.05)	1771 (0.10)	1676 (0.09)	1404 (0.08)	5357 (0.30)	373 (0.02)	1591 (0.09)	18160 (1.00)
Total	3227 (0.11)	2100 (0.07)	3009 (0.10)	2052 (0.07)	2871 (0.09)	2776 (0.10)	2504 (0.09)	6457 (0.22)	1473 (0.05)	2691 (0.09)	29160 (1.00)
Bacterial spot: BS; Leaf Mold: LM; Early blight: EB; Late blight: LB; Septoria leaf spot: SLS; Spider mites: SE; Target Spot: TS; Yellow Leaf Curl Virus: YLCV; Mosaic virus: MV

Table 2: Corrected class distribution comparison between Kaggle and PlantVillage tomato datasets (instances and ratios).

Panel of Figure 2 — Figure 2: Examples of representative images from each class in the tomato leaf disease datasets, illustrating the visual diversity and key characteristics.

Parameter	Value
Learning Rate (LR)	$1\times 10^{-5}$
Loss Function	CE (label_smoothing = 0.1)
Optimizer	AdamW (weight_decay = $1\times 10^{-4}$ )
LR Scheduler	CosineAnnealingLR ( $T_{\text{max}}=300$ )
Stopping Patience	50 epochs
Batch Size	32

Table 3: Training configuration details used for model training.

Metric	Expression
Accuracy	$\displaystyle\frac{\sum_{i=1}^{K}\mathrm{TP}_{i}}{\sum_{i=1}^{K}(\mathrm{TP}_{i}+\mathrm{FP}_{i}+\mathrm{FN}_{i})}$
Recall	$\displaystyle\frac{1}{K}\sum_{i=1}^{K}\frac{\mathrm{TP}_{i}}{\mathrm{TP}_{i}+\mathrm{FN}_{i}}$
Precision	$\displaystyle\frac{1}{K}\sum_{i=1}^{K}\frac{\mathrm{TP}_{i}}{\mathrm{TP}_{i}+\mathrm{FP}_{i}}$
F1-score	$\displaystyle\frac{1}{K}\sum_{i=1}^{K}\frac{2\,\mathrm{TP}_{i}}{2\,\mathrm{TP}_{i}+\mathrm{FP}_{i}+\mathrm{FN}_{i}}$

Table 4: Macro-averaged metrics used to evaluate the proposed CoAtNet-DeepMoE

SL	Reference	Parameters	Model Name	Accuracy	Precision	Recall	F1-score
1	[ 37 ]	11.00 M	YOLOv8s	93.2	90.3	–	–
2	[ 9 ]	32.29 M	DenseNet	95.4	95.7	95.4	95.4
3	[ 23 ]	23.50 M	ResNet50	92.4	–	–	–
4	[ 23 ]	23.56 M	SDE-ResNet50	98.0	98.0	–	98.0
5	[ 23 ]	54.32 M	Inception-ResNet	98.2	98.23	–	98.2
6	[ 49 ]	200.00 M	AlexNet + ResNet50 + VGG16 + QSSVM	98.3	–	–	–

Table 5: Performance comparison of state-of-the-art (SOTA) models on the Kaggle tomato disease dataset. “M” denotes the number of trainable parameters in millions. Accuracy, Precision, Recall, and F1-score are reported in percentage (%) form. Bold values indicate the best performance among the compared models.

Panel of Figure 3 — Figure 3: 2D comparison of SOTA models on (a) the Kaggle dataset and (b) the PlantVillage dataset. Each plot presents model accuracy against parameter count (in millions), illustrating how different architectures balance predictive performance and model complexity.

Panel of Figure 4 — Figure 4: Normalized confusion matrices of the proposed CoAtNet-DeepMoE model on (a) the Kaggle tomato leaf disease dataset and (b) the PlantVillage tomato dataset.

SL	Reference	Parameters	Model Name	Accuracy	Precision	Recall	F1-score
1	[ 52 ]	138.36 M	VGG16	96.46	–	–	–
2	[ 52 ]	3.50 M	MobileNetV2	95.76	–	–	–
3	[ 52 ]	27.53 M	Swin Transformer	99.10	–	–	–
4	[ 52 ]	8.06 M	Vision Transformer	99.45	–	–	–
5	[ 41 ]	25.60 M	ResNet50	99.80	–	–	–
6	[ 51 ]	5.70 M	ViT	96.50	93.90	96.70	94.20

Table 6: Performance comparison of state-of-the-art (SOTA) models on the PlantVillage dataset. “M” denotes the number of parameters in millions. Accuracy, Precision, Recall, and F1-score are reported in percentage (%) form. Bold values indicate the best performance among the compared models.

Panel of Figure 5 — Figure 5: Normalized confusion matrices for cross-dataset evaluation.

	PlantVillage $\rightarrow$ Kaggle				Kaggle $\rightarrow$ PlantVillage
Class	Accuracy	Precision	Recall	F1-score	Accuracy	Precision	Recall	F1-score
BS	1.0000	1.0000	1.0000	1.0000	0.9994	1.0000	0.9951	0.9975
EB	1.0000	1.0000	1.0000	1.0000	0.9997	1.0000	0.9953	0.9977
LB	1.0000	1.0000	1.0000	1.0000	0.9994	0.9949	1.0000	0.9974
LM	1.0000	1.0000	1.0000	1.0000	1.0000	1.0000	1.0000	1.0000
SLS	0.9990	0.9901	1.0000	0.9950	0.9997	1.0000	0.9971	0.9986

Table 7: Class-wise performance of the proposed model under cross-dataset evaluation. The model is trained on one dataset and evaluated on the other. Accuracy, Precision, Recall, and F1-score are reported as decimal values. The Average row represents the macro-average across all ten classes.

Figure 6: Cost–latency trade-off among CoAtNet-Base, ResNet50, ViT-Tiny, and the proposed CoAtNet-DeepMoE. The x-axis shows computational cost (GFLOPs), the y-axis shows inference latency (ms/img), and bubble size represents the number of parameters (M). The dashed line marks the Pareto frontier of non-dominated models.

Model	Params (M) $\downarrow$	GFLOPs $\downarrow$	Time/img (ms) $\downarrow$	p95 (ms) $\downarrow$	p99 (ms) $\downarrow$	FPS $\uparrow$
CoAtNet-Base	26.67	8.42	6.79	6.90	6.98	147.3
ResNet50	23.53	8.26	4.47	4.65	4.70	223.6
ViT-Tiny	5.53	2.15	3.78	4.93	5.02	264.8
CoAtNet-DeepMoE	2.47	1.34	4.04	4.19	4.24	247.6

Table 8: Cost analysis of different models in terms of parameter count (M), computational complexity (GFLOPs), and inference latency (ms/img), measured on the same hardware. p95/p99 denote the 95th/99th-percentile per-image latency over repeated runs. All metrics follow a lower-is-better (

\downarrow

) convention, except FPS, which is higher-is-better (

\uparrow

). Bold values indicate the most efficient performance among the compared models.

Plant Disease Classification

Plant disease classification uses image and sensor data to identify crop diseases and support timely, targeted agricultural intervention.

See all 60 papers

Cross-Architectural Mixture-of-Experts with Adaptive Soft Routing for Plant Leaf Disease Classification
Jun 22, 2026Phi-Hung Hoang, Thi-Thu-Hong PhanPlant Disease Classification Mixture-of-Experts Models
AIT laboratory, Faculty of Artificial Intelligence, FPT University, Da Nang, 550000, Viet Nam
U-FedTomAtt: Ultra-lightweight Federated Learning with Attention for Tomato Disease Recognition
Feb 18, 2026Romiyal George, Sathiyamohan Nishankar, Selvarajah Thuseethan +4Lightweight Neural Networks Federated Learning
Department of Computer Engineering, University of Peradeniya, Sri Lanka. · Faculty of Science and Technology, Charles Darwin University, Australia. · Canadian Institute for Cybersecurity, University of New Brunswick, Canada. +1
AgroVisNet: A lightweight Convolutional Network and the BD-PlantDX Expert-Validated Benchmark for Radish, Potato and Pointed Gourd Disease Classification
Sep 9, 2026Md. Abdullah Mandal, Saad Ahmed, Md. Khalid SyfullahPlant Disease Classification Lightweight CNNs
Department of Computer Science and Engineering, Rajshahi University of Engineering & Technology, Rajshahi, 6204, Bangladesh · Department of Computer Science and Engineering, Bangladesh Army University of Science & Technology, Saidpur, Bangladesh

Jun 22, 2026J/K move · Enter open · S save

Cross-Architectural Mixture-of-Experts with Adaptive Soft Routing for Plant Leaf Disease Classification

Phi-Hung Hoang, Thi-Thu-Hong Phan

AIT laboratory, Faculty of Artificial Intelligence, FPT University, Da Nang, 550000, Viet Nam

Plant Disease Classification Mixture-of-Experts Models

Open paper AskPDF arXiv

Plant leaf disease classification is crucial for crop protection and precision agriculture but remains challenging under complex backgrounds, illumination variations, and severe class imbalance. Moreover, single-architecture models often fail to effectively capture both local and global representations. To address these challenges, this study proposes an adaptive soft Mixture-of-Experts (MoE) framework with cross-architectural routing that integrates EfficientNet-B0, DenseNet-121, and Swin-Tiny to exploit complementary multi-scale, local, and global features. A soft gating mechanism dynamically assigns input-dependent expert weights, while a two-stage refinement training strategy improves optimization stability and generalization. Experiments on a highly imbalanced potato leaf disease dataset achieve 91.68% recall and 92.62% F1-score, surpassing the strongest individual expert by 5.91% and 5.03%, respectively. Additional evaluations on durian and sesame leaf disease datasets yield F1-scores of 94.03% and 97.04%, demonstrating robust cross-dataset generalization and the potential of the proposed framework for reliable real-world crop health monitoring

Abstract

Plant Disease Classification

Explore similar work

Cross-Architectural Mixture-of-Experts with Adaptive Soft Routing for Plant Leaf Disease Classification

U-FedTomAtt: Ultra-lightweight Federated Learning with Attention for Tomato Disease Recognition

AgroVisNet: A lightweight Convolutional Network and the BD-PlantDX Expert-Validated Benchmark for Radish, Potato and Pointed Gourd Disease Classification