Smart agriculture relies heavily on visual data captured using low-cost mobile devices in rural environments. Farmers often face challenges in identifying crop diseases, pest attacks, and soil conditions due to the lack of expert assistance and limited access to reliable information. Image retrieval systems can support decision-making by retrieving visually and semantically similar agricultural images from a reference database. However, traditional classification-based approaches are not well suited for retrieval tasks and often require large computational resources. This paper introduces a deep metric learning-based image retrieval framework for smart agriculture applications, optimized for resource-constrained rural environments. A light convolutional neural network is used for obtaining discriminative embeddings, and a triplet loss function is used to learn an embedding space that is aware of similarities. The proposed approach is validated on agricultural image datasets, with comparisons to existing CNN models for image retrieval tasks in the literature. The results have shown that the proposed approach can highly enhance the retrieval accuracy with lower computational complexity, meeting the requirements for practical rural applications.