Federated learning with dynamic weighted aggregation for multi-crop disease detection: a hybrid CNN-transformer approach.

Abstract

Crop diseases pose a significant threat to agriculture globally, causing a loss of 220 billion dollars annually. The problem can be solved through traditional machine learning methods which run on central servers but face two main obstacles: farmers refuse to provide their farming information and the uneven distribution of crop diseases across different regions. This paper introduces a federated learning framework that addresses these challenges through a novel approach that combines hybrid CNN-Transformer architectures with dynamic weighted aggregation. Our system applies EfficientNet-B0 and MobileNetV2 lightweight models which use MobileViT blocks together with CBAM and ESA attention mechanisms to extract detailed features from crop images. The innovation lies in the dynamic aggregation strategy, "AdaClass"-Adaptive Class-Aware aggregation, that identifies the underperforming classes in each round using class-wise F1 score and emphasizes the clients that are having a stronger performance on challenging disease classes.This approach helps in promoting a balanced performance across different disease classes. This particularly benefits the disease classes that are difficult to detect. Extensive experiments on two standard datasets demonstrate strong results: the proposed EfficientNet-B0 hybrid model achieves 99.32% accuracy on PlantVillage and 92.5% on CCMT datasets, while the MobileNetV2 achieves 99.17% and 91.4% respectively. Importantly, these models remain efficient enough for deployment on edge devices, with the MobileNetV2 hybrid requiring only 7.63 MB storage and processing images in 41.2 milliseconds.

Plant phenotyping relevance

植物画像から病害状態を推定するフェデレーテッド学習手法を開発し、複数データセットで性能評価しており、病害フェノタイピング手法が研究の中心である。

Code and data availability

The paper uses public PlantVillage and CCMT datasets, but these are cited third-party benchmarks, not paper-specific deposits. No author code, models, or data availability URLs are provided; the data availability statement only directs inquiries to the corresponding author, and no allowed URLs exist to cite.

No evidence-backed public reproduction asset is currently recorded.