This paper explores advanced methodologies for automating the classification of fruits and vegetables, that is a critical task in agricultural grading and exportation processes. Starting from a baseline automatic agricultural products recognition system which was built in a previous work, we compare various dimensionality reduction techniques to enhance the recognition performance. The baseline system was based on the Principal Component Analysis dimensionality reduction technique and a classification by Support Vector Machines; it achieves a recognition accuracy of 90.5%. To improve the recognition performance, we investigate in this paper other dimensionality reduction techniques that are: t-Distributed Stochastic Neighbor Embedding, Isometric mapping, Autoencoders, Multidimensional Scaling, Locally Linear Embedding, Uniform Manifold Approximation and Projection (UMAP) and Factor Analysis. The experimental results demonstrate the strengths and the limitations of each technique, emphasizing the role of dimensionality reduction in improving the classification models. They reveal that UMAP emerged as the most efficient technique, leading to a classification accuracy of about 93.0% while preserving both local and global data structures. The UMAP’s scalability and computational efficiency make it particularly suitable for handling large and complex datasets, outperforming both traditional and other modern methods. This work contributes to the ongoing development of robust automated systems in agriculture, enhancing efficiency in quality control and export logistics.