Optimizing federated learning for maize leaf disease classification using internet of agricultural things

doi:10.1007/s10586-025-05888-8

Optimizing federated learning for maize leaf disease classification using internet of agricultural things

2026-01-06

0

PRE

AI

Umesh Kumar Lilhore *

DOI：10.1007/s10586-025-05888-8

Original

Original request for help

Save

Abstract

En 中文

The effective identification and management of agricultural diseases is critical for increasing crop yields, lowering costs, and promoting sustainable farming practices. Although traditional machine learning and deep learning methodologies have significantly advanced automated disease detection, issues related to centralized data collection, such as security, privacy, and transfer costs, limit their effectiveness. This study introduces a new Federated Learning (FL) framework for classifying maize leaf diseases using the Internet of Agricultural Things (IoAT). The study introduces a novel method for building lightweight and efficient models for use on edge devices by uniting federated transfer learning with federated feature extraction. The framework uses convolutional neural networks (CNNs) and attention-based models that are trained on the PlantVillage dataset. This dataset includes 4,585 images of maize leaves, showing different diseases like common rust, Turcicum leaf blight, maydis leaf blight, grey leaf spot, and also healthy leaves. The evaluation of the model’s performance includes measures like precision, recall, accuracy, loss, and area under the curve (AUC). It shows that DenseNet121 is the best option as the base model, reaching a very high validation accuracy of 99.2% in the federated learning model. The proposed FL plus transfer learning model performed even better, achieving 99.89% accuracy on non-IID data, which shows strong generalization and makes it suitable for use on edge devices. The experimental setup was built using publicly available federated learning code from GitHub. The code and dataset are available at: https://github.com/prabhnoorbachhal-a11y/FL . Details about the hardware and experimental setup are explained in the Methods section. These results support the insight made in the paper: (i) federated learning can reach accuracy similar to a centralized system while keeping data private; (ii) using pre-trained features allows models to work on devices with limited resources; and (iii) DenseNet-121 offers the best balance among the tested model types. This framework also helps with safeguarding data and making it easier to access, while being scalable and efficient for detecting plant diseases. The research utilizes federated learning to establish a standard for secure, decentralized, and high-performance IoT and agricultural applications, thereby enhancing smart agriculture technology.

Keywords:

Internet of agricultural things

Deep learning

Federated learning

Maize leaf disease detection

Transfer learning

Classification

Optimizing federated learning for maize leaf disease classification using internet of agricultural things

Abstract

Journal

Organization