Dynamic Meta-Ensemble Framework for Efficient and Accurate Deep Learning in Plant Leaf Disease Detection on Resource-Constrained Edge Devices

WELODAY FIKADU MOGES 1*, JIANMEI SU 1, AMIN WAQAS 2 · 2026

Read Dynamic Meta-Ensemble Framework for Efficient and Accurate Deep Learning in Plant Leaf Disease Detection on Resource-Constrained Edge Devices by WELODAY FIKADU MOGES 1*, JIANMEI SU 1, AMIN WAQAS 2 (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-17290.txt.

Introduction

Weloday Fikadu Moges, Jianmei Su, and Amin Waqas built a system to spot plant leaf disease on small devices. Phones, sensors, and farm gear have tight limits on compute and power. Big deep learning models often do not fit those limits. Yet farmers need fast, exact disease checks in the field. The team's Dynamic Meta-Ensemble Framework aims to meet both needs.

It joins models in a meta-ensemble that can adapt as it runs. The design chases high accuracy without heavy cost. Early disease finds can save crops and cut spray waste. A tool that runs on cheap gear spreads that gain widely. The work targets real farms, not just lab tests.

Catalog record

CREATORWELODAY FIKADU MOGES 1*, JIANMEI SU 1, AMIN WAQAS 2
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE49 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-17290

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Dynamic Meta-Ensemble Framework for Efficient and Accurate Deep Learning in Plant Leaf Disease Detection on Resource-Constrained Edge Devices WELODAY FIKADU MOGES 1*, JIANMEI SU 1, AMIN WAQAS 2 Abstract Deploying deep learning models for plant disease detection on edge devices such as IoT sensors, smartphones, and embedded systems is severely constrained by limited computational resources and energy budgets. To address this challenge, we introduce a novel Dynamic Meta-Ensemble Framework (DMEF) for high accuracy plant disease diagnosis under resource constraints. DMEF employs an adaptive weighting mechanism that dynamically combines the predictions of three lightweight convolution neural networks ( MobileNetV2 , NASNetMobile , and InceptionV3 ) by optimizing a trade-off between accuracy improvements ( Δ \Delta Acc) and computational efficiency (model size). During training, the ensemble weights are updated iteratively, favoring models exhibiting high performance and low complexity. Extensive experiments on benchmark datasets for potato and maize diseases demonstrate state-of-the-art classification accuracies of 99.53% and 96.61%, respectively, surpassing standalone models and static ensembles by 2.1–6.3%. With computationally efficient inference latency (<75 ms ) and a compact footprint (<1 million parameters), DMEF shows strong potential for edge-based agricultural monitoring, suggesting viability for scalable crop disease management. This bridges the gap between high-accuracy AI and practical field applications. † † address: South West University Science and Technology, School of Computer Science and Technology, Mianyang, China † † address: South West University Science and Technology, School of Computer Science and Technology, Mianyang, China † † address: South West University Science and Technology, School of Control and Information Engineering Technology, Mianyang, China † † corresponding: Weloday Fikadu Moges ([email protected]) I Introduction The timely detection of plant leaf diseases represents a critical challenge in precision agriculture, serving as the first line of defense against catastrophic crop losses [ 17 ] . In particular, in economically vital crops such as potatoes and maize, undiagnosed pathologies can trigger production declines that exceed 20-40% [ 18 ] , exacerbating global food insecurity. Although the conventional diagnosis of plant diseases predominantly depends on visual inspection performed by trained agronomists, a method that is labor-intensive, time-consuming, and prone to human error, it often results in inconsistent and inaccurate disease identification. Such limitations hinder timely intervention and can exacerbate crop losses, highlighting the need for more reliable automated detection techniques [ 23 ] . Contemporary deep learning architectures, particularly ResNet [ 1 ] , Inception [ 2 ] , and VGG [ 20 ] variants have demonstrated remarkable success in plant disease classification, achieving > 95% accuracy in controlled laboratory settings. However, their translation to practical edge deployment faces three fundamental barriers: (1) the computational intractability of billion-parameter models on IoT-grade hardware [ 24 ] , (2) memory constraints in embedded systems [ 24 ] , and (3) unsustainable power consumption for continuous field operation [ 25 ] . These limitations make state-of-the-art models challenging for real-world agricultural deployment. Emerging compact architectures (MobileNet [ 4 ] , EfficientNet [ 5 ] , SqueezeNet [ 6 ] ) to reconcile this tension through architectural pruning and quantization, but have drawbacks. For instance, SqueezeNet shows very fast inference and small model size but suffers from poor accuracy on complex datasets. At the same time, EfficientNetV2 achieves high accuracy but has larger size and slower inference, limiting its use in resource-constrained settings [ 27 ] . Recent research also explores low-cost ensemble methods to reduce overhead but notes that achieving the same performance as classical ensembles often still requires multiple forward passes or complex model architectures, thus maintaining non-trivial computation costs [ 26 ] . In this paper, we propose a Dynamic Meta-Ensemble Framework for efficient and accurate plant leaf disease detection on resource-constrained edge devices. The framework dynamically adjusts the contribution of each model in an ensemble based on two key factors: (1) the accuracy improvement during training and (2) the model size during inference. By prioritizing accuracy during training and emphasizing smaller, efficient models during inference,the framework ensures optimal performance on edge devices. Our approach is designed to balance accuracy and computational efficiency, making it suitable for real-time disease detection in agricultural applications. I-A Motivation and Contribution I-A 1 Motivation Food security is increasingly threatened by crop diseases, yet traditional detection methods are slow and unreliable. While AI solutions offer promise, they often fail to balance accuracy with the computational constraints of edge devices used in real-world farming. Our work addresses this gap by introducing a dynamic lightweight AI system that delivers lab-grade disease detection to farmers via affordable, edge-compatible devices. I-A 2 Contribution We propose a dynamic meta ensemble framework that adapts to agricultural challenges by: (1) introducing a novel weighting mechanism to balance accuracy and efficiency for edge deployment; (2) designing an architecture optimized for real-time processing; (3) demonstrating consistent, superior performance across various crops. This work bridges the gap between high-accuracy AI and practical field applications, setting a new benchmark for intelligent crop disease monitoring. II RELATED WORKS Deep learning has revolutionized computer vision, achieving state-of-the-art performance in image classification tasks. In agriculture, deep learning models have been widely adopted for plant disease detection. Early work by Mohanty et al. [ 16 ] demonstrated the potential of convolutional neural networks (CNNs), achieving 99.35% accuracy on lab-captured images of 26 diseases across 14 plant species. However, these models struggled in real-world field conditions due to variability in lighting, occlusion, and background noise. Subsequent studies, such as Sladojevic et al. [ 21 ] , achieved high accuracy (up to 96.3%) in classifying 13 leaf diseases across 5 plant species using deep neural networks; however, their results were based on a relatively small and limited dataset, which constrains the generalizability of their approach to broader, real-world scenarios. Ferentinos [ 12 ] further advanced the field by developing a deep CNN-based system that achieved up to 99.53% accuracy in classifying 58 different plant disease classes using the PlantVillage dataset. While this work demonstrated state-of-the-art performance, the computational demands of such deep learning models remain a significant barrier for real-time or edge deployment, as these models typically require substantial processing power and memory. These studies highlight the potential of deep learning for automating plant disease detection. However, most existing approaches rely on large, high-accuracy models such as ResNet and Inception, which are computationally intensive and unsuitable for deployment on resource-constrained edge devices. Although these models achieve high accuracy, their computational cost limits their practicality in real-world agricultural applications. Recent research has introduced innovative architectures such as depthwise CNNs with squeeze-and-excitation (SE) blocks (EfficientNet [ 5 ] ) and residual connections (Xception [ 14 ] ), achieving high accuracy and enhancing computational efficiency and real-time applicability. Transfer learning approaches, notably those leveraging Xception [ 14 ] and advanced data augmentation, have further improved robustness and generalization across multiple plant species and disease types [ 15 , 9 ] . Despite these advances, challenges persist, including data imbalance, model interpretability, and generalization to novel diseases in diverse field conditions [ 9 , 13 ] . State-of-the-art architectures like ResNet [ 1 ] and Inception [ 2 ] deliver high accuracy but are computationally intensive (e.g., VGG-16 requires 15.5 GFLOPs) [ 3 ] . Lightweight alternatives such as MobileNet [ 4 ] and EfficientNet [ 5 ] reduce parameters

Read the complete text