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

Weloday Fikadu Moges, Jianmei Su+1 more|arXiv (Cornell University)|Jan 24, 2026

Smart Agriculture and AI0 citations

TL;DR

The paper introduces Dynamic Meta-Ensemble Framework (DMEF) that dynamically weights three lightweight CNNs (MobileNetV2, NASNetMobile, InceptionV3) to balance accuracy and efficiency for edge deployment, achieving 99.53% potato and 96.61% maize disease classification with under 75 ms latency and <1M parameters.

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 convolutional neural networks (MobileNetV2, NASNetMobile, and InceptionV3) by optimizing a trade-off between accuracy improvements (DeltaAcc) 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% and 6.3%. With computationally efficient inference latency (&lt;75ms) and a compact footprint (&lt;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.

Motivation & Objective

• Address how to detect plant leaf diseases accurately on edge devices with limited compute and energy.
• Develop a dynamic ensemble that balances accuracy with model size for real-time inference.
• Leverage lightweight base models and transfer learning to maximize generalization across crops (potato and maize).
• Demonstrate state-of-the-art performance while maintaining a compact model footprint suitable for edge deployment.

Proposed method

• Preprocess PlantVillage data with resizing to 128x128 and normalization to 0-1; apply data augmentation (random flips, rotations, zoom, contrast).
• Fine-tune three lightweight pretrained models: MobileNetV2, NASNetMobile, InceptionV3, replacing top layers with global pooling and softmax; unfreeze limited deeper layers for task adaptation.
• Define a dynamic ensemble weighting w_i = lambda_i * alpha_i + (1 - lambda_i) * beta_i, where alpha_i is the accuracy proportion and beta_i is the model size proportion.
• Compute lambda_i(t) via adaptive updates lambda_i(t) = clip(lambda_i(t-1) + delta * DeltaA_i(t) / sum_j DeltaA_j(t), lambda_min, lambda_max).
• Train base models and update ensemble weights epoch-by-epoch to favor accurate yet compact models; inference uses weighted sum of softmax outputs across models.

Figure 1 : Dynamic Meta-Ensemble Framework for Plant Disease Classification

Experimental results

Research questions

• RQ1 Can a dynamic meta-ensemble improve accuracy versus individual lightweight models on edge devices?
• RQ2 Does balancing accuracy with model size during training yield better edge-implementation performance than static ensembles?
• RQ3 What are the trade-offs between accuracy gains and latency when using a dynamic meta-ensemble on resource-constrained hardware?
• RQ4 Does the framework generalize across different crops/diseases in real-world field conditions?

Key findings

• potato dataset: accuracy 99.53% with near-perfect class-wise metrics and strong robustness across Early/ Late Blight and Healthy classes.
• maize dataset: accuracy 96.61%; Common rust and Healthy leaves achieved precision/recall/F1 of 1.00; some variance for Gray leaf spot (recall 0.78).
• ensemble total parameters: 994,012 across MobileNetV2, NASNetMobile, InceptionV3 (sub-1M scale); latency < 75 ms for inference.
• dynamic weights evolve to favor InceptionV3 on Potato (final ~0.384) and on Maize scenarios (InceptionV3 ~0.4013; NASNetMobile ~0.3102; MobileNetV2 ~0.2885).
• accuracy-efficiency frontier shows Pareto-optimal positioning with up to 5.8% accuracy gain at 5.2× latency for Potato and 2.3% gain at 6.9× latency for Maize.

Figure 2 : Flow chart of the Dynamic Meta-Ensemble Framework