<![CDATA[Purpose – This study investigates the optimal number of training epochs for the YOLO26 model in detecting graves and free burial slots from UAV photogrammetry imagery, with particular attention to model convergence, generalization, and detection performance. Design/methods/approach – A publicly available cemetery dataset containing two object classes, namely graves and free burial slots, was preprocessed using auto-orientation, 2×2 tiling, resizing to 640 × 640 pixels, grayscale conversion, and data augmentation. The YOLO26 model was trained using transfer learning under six epoch configurations: 50, 100, 150, 200, 250, and 300 epochs. Performance was evaluated using precision, recall, F1-score, mAP@50, mAP@50–95, confusion matrices, and training and validation loss curves. Findings – Model performance improved substantially as training progressed and began to stabilize after approximately 200 epochs. The highest observed performance occurred at epoch 289, with a precision of 98.78%, recall of 98.26%, F1-score of 99%, mAP@50 of 99.40%, and mAP@50–95 of 90.94%. Although the 300-epoch configuration produced similarly strong results, the additional gains were marginal, indicating diminishing returns after convergence. Research implications/limitations – The findings provide practical guidance for selecting an appropriate training duration in UAV-based cemetery mapping and small-object detection. However, the study relies on a relatively small, single-source dataset, which may limit generalizability across different cemetery layouts, environmental conditions, and UAV imaging configurations. Originality/value – This study provides a domain-specific multi-epoch benchmark for YOLO26 and demonstrates the importance of metric- and convergence-based checkpoint selection rather than relying solely on the maximum predefined number of epochs.]]>
