Timely durian leaf disease detection is critical for Vietnam’s agricultural productivity, yet traditional methods remainlabor-intensive and error-prone. This study proposes a hybrid pipeline integrating deep transfer learning with binary ParticleSwarm Optimization (PSO) for efficient disease classification. Three lightweight backbones, such as MobileNetV3-Large,EfficientNet-B0, and EfficientNetV2-B0 to extract 128-dimensional features from a real-world Vietnamese durian dataset(2595 images, 6 classes), which PSO then prunes u sing five-fold support vector machine (SVM) cross-validation fitness.The optimized subsets were evaluated across five machine learning (ML) classifiers, achieving up to 92.6% test accuracywith a modest improvement over the baseline while reducing dimensionality. PSO-selected features demonstrated thepotential for accelerated inference and interpretable agricultural diagnostics on resource-constrained devices.