Scholarlink

In agriculture, delays in crop transport and poor health classification often result in large-scale food wastage and economic loss. Centralised machine learning systems rely on raw data sharing, which raises privacy concerns and limits participation from farmers. We propose a federated learning framework using LSTM, trained across decentralised clients without sharing raw data to solve this. The model analyses NDVI, weather, and soil data, ensuring privacy, scalability, and performance. We integrate SHAP and Grad-CAM to explain which features and time periods most influenced predictions. Our system was benchmarked against a centralised LSTM and achieved comparable results (MSE = 0.039, R2 = 0.89) after 20 communication rounds. The explainability outcomes revealed that NDVI patterns and rainfall had the highest impact. This work introduces a novel combination of privacy-preserving federated training and explainable AI, supporting timely agro-logistics decisions and real-world deployment across diverse farming regions.

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

Fed-AI: Explainable Federated Learning for Sustainable Crop Forecasting and Agro-Logistics

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Aswin Kumaran Mahesh Kannan,
  • Raja Guru Ramaraj,
  • Jaya Swaruba Lakshmi Mahesh Kannan

摘要

In agriculture, delays in crop transport and poor health classification often result in large-scale food wastage and economic loss. Centralised machine learning systems rely on raw data sharing, which raises privacy concerns and limits participation from farmers. We propose a federated learning framework using LSTM, trained across decentralised clients without sharing raw data to solve this. The model analyses NDVI, weather, and soil data, ensuring privacy, scalability, and performance. We integrate SHAP and Grad-CAM to explain which features and time periods most influenced predictions. Our system was benchmarked against a centralised LSTM and achieved comparable results (MSE = 0.039, R2 = 0.89) after 20 communication rounds. The explainability outcomes revealed that NDVI patterns and rainfall had the highest impact. This work introduces a novel combination of privacy-preserving federated training and explainable AI, supporting timely agro-logistics decisions and real-world deployment across diverse farming regions.