PERTANIKA JOURNAL OF SCIENCE AND TECHNOLOGY

Early identification of durian cultivars at the seedling stage remains challenging because conventional methods rely on fruit characteristics that appear only after several years, often leading to misdistribution and inefficiencies in certification. This study proposes a real-time, non-destructive deep learning framework for intra-species durian cultivar identification using leaf images. A dataset of 1,788 leaf images from five cultivars (Bawor, Kani, Monthong, Musang King, and Petruk) was collected under real-world conditions. Unlike existing studies that formulate plant varietal identification solely as an image classification task, this work introduces a detection-based formulation that explicitly models cultivar-specific leaf features at the object level. This formulation enables the model to localise and learn fine-grained morphological differences that are often overlooked in global classification approaches. To validate this approach, MobileNetV2, Xception, and YOLOv11 were systematically evaluated using K-Fold Cross Validation under the same experimental setting. The key contributions of this study are: (1) the formulation of intra-species plant cultivar identification as an object detection problem rather than a pure classification task, (2) the development of a practical, real-time framework for early-stage durian cultivar identification using leaf images in unconstrained environments, and (3) a comparative analysis demonstrating the effectiveness of detection-based models in capturing subtle inter-varietal differences. Experimental results show that YOLOv11 achieves over 99% validation accuracy, with precision and recall above 0.95, while maintaining an inference time of 0.2 ms per image. These findings highlight the potential of detection-based approaches for AI-assisted seedling selection in horticulture.