Decentralized federated learning using validation loss for model sharing in crop disease classification
Agriculture plays an essential role in the economies of many countries, as it provides numerous livelihoods. However, managing crop diseases is one of the major challenges in modern agriculture. Using artificial intelligence (AI) for the early detection and diagnosis of crop diseases is an interesti...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-12-01
|
| Series: | Ecological Informatics |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S1574954125002146 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Agriculture plays an essential role in the economies of many countries, as it provides numerous livelihoods. However, managing crop diseases is one of the major challenges in modern agriculture. Using artificial intelligence (AI) for the early detection and diagnosis of crop diseases is an interesting approach to tackle this problem. Several AI methods have been employed for this purpose, but despite achieving good results, many challenges remain, such as protecting farmers’ data, using machine learning on edge devices, and employing collaborative learning. In this context, federated learning (FL) has emerged as a promising machine learning approach that enables to build efficient models with a collaborative manner, while preserving data privacy and security. There exist two types of FL: centralized and decentralized. In this paper we employ the approach of decentralized FL for crop disease image classification that utilizes peer-to-peer communication for updating models for each client. To address the problem of the robustness of shared models, we propose a new strategy based on validation loss, where the aggregated models should satisfy a certain criterion of performances. We implemented and tested two types of deep learning architectures, convolutional neural networks (CNNs) and vision transformers (ViTs). The evaluation of model performance was based on four metrics: Accuracy, F1-Score, Precision, and Recall. However, for the presentation of results in this paper, we focus on Accuracy and F1-Score to highlight key aspects of model performance. We evaluated the impact of the number of shared models, communication cycles, number of clients involved, local iterations, and training data size on model performance. The results show that decentralized FL offers significant advantages over centralized FL approaches, improving rapid convergence to high and stable performance. These results highlight the potential of decentralized FL to advance crop disease management, thereby contributing to agricultural resilience and productivity. |
|---|---|
| ISSN: | 1574-9541 |