ASABE Technical Library - Abstract
Member and Access Notice
Privacy-Preserving Communication-Efficient Federated Transfer Learning with Vision-Language Model for Smart Agricultural Image Classification
Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.org
Citation: 2026 ASABE Annual International Meeting 2600401.(doi:10.13031/aim.202600401)Authors: Long Li, Yanbo Huang, Jiajia Li, Dong Chen, Lina Pu, Haibo Yao
Keywords: Federated learning, Vision language model, Privacy-preserving, Smart agriculture, Transfer learning
Abstract. Federated learning is attractive for agricultural intelligence because it avoids centralizing farm data, but conventional federated optimization often performs poorly when clients hold non-independent and identically distributed (non-IID) data. This problem is particularly severe for agricultural image classification, where crop species, weed species, and imaging conditions can vary markedly across sites. This paper presents a federated transfer learning framework that combines a frozen Contrastive Language–Image Pre-training (CLIP) vision encoder with a lightweight transformer classifier and a privacy-preserving feature replay mechanism. Instead of exchanging raw images, each client contributes a very small subset (1%) of CLIP-extracted feature embeddings to a shared replay pool. These non-reversible embeddings provide a global class reference during local training while preserving privacy. Experimental results show that the proposed method achieves 86.6% accuracy, compared with 52.8% for a standard federated baseline and 20.3% for federated prompt generation baseline. The results demonstrate that feature replay is an effective and practical remedy for privacy-preserving and scalable agricultural intelligence.
(Download PDF) (Export to EndNotes)