Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Loss-Guided Model Sharing and Local Learning Correction in Decentralized Federated Learning for Crop Disease Classification

Denis Mamba Kabala, Adel Hafiane, Laurent Bobelin, Raphael Canals

TL;DR

Addressing privacy and scalability in crop disease classification, the paper proposes a Loss_val-guided decentralized federated learning framework with a $\lambda$-weighted local loss to balance local data representativeness and global generalization. By sharing only the best neighbor models based on $Loss_{val}$ and updating local learning with an adaptive loss that incorporates the received models, the approach enables serverless, peer-to-peer collaboration. Evaluations on PlantVillage with ResNet50, VGG16, and ViT_B16 show faster convergence and higher accuracy, with ResNet50 benefiting most from external guidance while VGG16 and ViT_B16 exhibit robust generalization and stability across settings. The results demonstrate reduced communication costs and improved robustness to data heterogeneity, outperforming centralized FL and prior CNN/ViT-based methods in this domain.

Abstract

Crop disease detection and classification is a critical challenge in agriculture, with major implications for productivity, food security, and environmental sustainability. While deep learning models such as CNN and ViT have shown excellent performance in classifying plant diseases from images, their large-scale deployment is often limited by data privacy concerns. Federated Learning (FL) addresses this issue, but centralized FL remains vulnerable to single-point failures and scalability limits. In this paper, we introduce a novel Decentralized Federated Learning (DFL) framework that uses validation loss (Loss_val) both to guide model sharing between peers and to correct local training via an adaptive loss function controlled by weighting parameter. We conduct extensive experiments using PlantVillage datasets with three deep learning architectures (ResNet50, VGG16, and ViT_B16), analyzing the impact of weighting parameter, the number of shared models, the number of clients, and the use of Loss_val versus Loss_train of other clients. Results demonstrate that our DFL approach not only improves accuracy and convergence speed, but also ensures better generalization and robustness across heterogeneous data environments making it particularly well-suited for privacy-preserving agricultural applications.

Loss-Guided Model Sharing and Local Learning Correction in Decentralized Federated Learning for Crop Disease Classification

TL;DR

Addressing privacy and scalability in crop disease classification, the paper proposes a Loss_val-guided decentralized federated learning framework with a -weighted local loss to balance local data representativeness and global generalization. By sharing only the best neighbor models based on and updating local learning with an adaptive loss that incorporates the received models, the approach enables serverless, peer-to-peer collaboration. Evaluations on PlantVillage with ResNet50, VGG16, and ViT_B16 show faster convergence and higher accuracy, with ResNet50 benefiting most from external guidance while VGG16 and ViT_B16 exhibit robust generalization and stability across settings. The results demonstrate reduced communication costs and improved robustness to data heterogeneity, outperforming centralized FL and prior CNN/ViT-based methods in this domain.

Abstract

Crop disease detection and classification is a critical challenge in agriculture, with major implications for productivity, food security, and environmental sustainability. While deep learning models such as CNN and ViT have shown excellent performance in classifying plant diseases from images, their large-scale deployment is often limited by data privacy concerns. Federated Learning (FL) addresses this issue, but centralized FL remains vulnerable to single-point failures and scalability limits. In this paper, we introduce a novel Decentralized Federated Learning (DFL) framework that uses validation loss (Loss_val) both to guide model sharing between peers and to correct local training via an adaptive loss function controlled by weighting parameter. We conduct extensive experiments using PlantVillage datasets with three deep learning architectures (ResNet50, VGG16, and ViT_B16), analyzing the impact of weighting parameter, the number of shared models, the number of clients, and the use of Loss_val versus Loss_train of other clients. Results demonstrate that our DFL approach not only improves accuracy and convergence speed, but also ensures better generalization and robustness across heterogeneous data environments making it particularly well-suited for privacy-preserving agricultural applications.

Paper Structure

This paper contains 20 sections, 5 figures, 7 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related work
  3. Detection of crop diseases using deep learning models
  4. Federated Learning (FL)
  5. Decentralized Federated Learning (DFL)
  6. Methodology
  7. Proposed approach
  8. Deep learning architectures
  9. Experiments and results
  10. Experimental procedure
  11. Description of the datasets
  12. Hardware setup
  13. Configuration of learning clients
  14. Performance evaluation measures
  15. Results and discussion
  16. ...and 5 more sections

Figures (5)

  • Figure 1: Architecture of the proposed approach.
  • Figure 2: Example of images from our datasets.
  • Figure 3: Impact of the $\lambda$ parameter, in 6-client configuration, one best shared model, with $\lambda = 0.25$.
  • Figure 4: Impact of the number of shared models, in 6-client configuration, five best shared models, with $\lambda = 0.25$.
  • Figure 5: Impact of using $Loss_{val}$ vs $Loss_{train}$ in a 18-client configuration, five best shared models.