Scholarlink

<p>Federated Learning (FL) enables privacy-preserving collaborative model training across distributed agricultural IoT networks, yet existing aggregation strategies such as Federated Averaging (FedAvg), Trimmed Mean, and Krum either assume data homogeneity or discard informative outlier updates, limiting performance in heterogeneous soil environments. This study introduces the Hybrid Federated Averaging–Weighted Winsorized Aggregation (WWA) framework for multiclass soil quality prediction across five distributed farms. The method employs FedAvg for warm-up stabilisation, then applies metadata-driven trust weighting and winsorization from round six onward, attenuating rather than discarding divergent client updates to preserve rare but meaningful soil patterns. Formal definitions of the trust weight computation and sensitivity analyses of weighting hyperparameters confirm the stability and fairness of the aggregation mechanism. Experiments on both a synthetic federated soil dataset and the real-world Crop Recommendation Dataset demonstrate consistent superiority of the proposed approach: on the primary dataset, Hybrid FedAvg–WWA achieves 96.9% accuracy and 96.2% F1-score, outperforming FedAvg (94.1%), FedProx (95.0%), and TrimmedFL (95.4%), while reducing communication rounds by 25% and improving the weakest client’s accuracy by up to 3.5 percentage points.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Hybrid Winsorized Federated Averaging for robust soil quality classification across distributed agricultural IoT networks

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月20日

查看全文
  • Rahul Haripriya,
  • Gurleen Kaur Bhatia,
  • Manish Pandey,
  • Nilay Khare,
  • H. S. Hota,
  • Lalit Kumar

摘要

Federated Learning (FL) enables privacy-preserving collaborative model training across distributed agricultural IoT networks, yet existing aggregation strategies such as Federated Averaging (FedAvg), Trimmed Mean, and Krum either assume data homogeneity or discard informative outlier updates, limiting performance in heterogeneous soil environments. This study introduces the Hybrid Federated Averaging–Weighted Winsorized Aggregation (WWA) framework for multiclass soil quality prediction across five distributed farms. The method employs FedAvg for warm-up stabilisation, then applies metadata-driven trust weighting and winsorization from round six onward, attenuating rather than discarding divergent client updates to preserve rare but meaningful soil patterns. Formal definitions of the trust weight computation and sensitivity analyses of weighting hyperparameters confirm the stability and fairness of the aggregation mechanism. Experiments on both a synthetic federated soil dataset and the real-world Crop Recommendation Dataset demonstrate consistent superiority of the proposed approach: on the primary dataset, Hybrid FedAvg–WWA achieves 96.9% accuracy and 96.2% F1-score, outperforming FedAvg (94.1%), FedProx (95.0%), and TrimmedFL (95.4%), while reducing communication rounds by 25% and improving the weakest client’s accuracy by up to 3.5 percentage points.