<![CDATA[Purpose – This study develops and evaluates a UAV-based automated surveillance approach using the YOLO26 architecture to detect visible symptoms associated with Ganoderma boninense infection in oil palm canopies. The study addresses the limitations of conventional manual inspection and multi-stage detection systems by applying a unified single-stage object-detection framework. Design/methods/approach – The model was developed using a publicly available dataset containing 1,133 annotated UAV images of oil palm canopies. Images were preprocessed, augmented, and partitioned using a plantation-block-aware strategy to reduce spatial data leakage. YOLO26s was trained using the Ultralytics framework on an NVIDIA Tesla T4 GPU. Model performance was evaluated using precision, recall, mAP@50, mAP@50–95, confidence-threshold sensitivity analysis, precision–recall curves, and five-fold block-aware cross-validation. Findings – On the independent block-aware test set of 118 images, the model achieved an mAP@50 of 77.14%, mAP@50–95 of 41.98%, precision of 66.03%, and recall of 76.32%. Five-fold block-aware cross-validation produced a mean mAP@50 of 74.12% ± 5.51% and a mean mAP@50–95 of 41.31% ± 4.00%. The relatively high recall indicates that the model can identify most visible infection instances, although its moderate precision shows that false-positive detections remain a practical concern. Research implications/limitations – The findings demonstrate the potential of YOLO26 to support UAV-based oil palm disease surveillance and targeted field inspection. However, the study relies on a single public dataset with inherited annotation procedures, lacks geographically independent external validation, and does not include direct benchmarking on onboard UAV or embedded edge devices. Originality/value – This study provides an early empirical evaluation of YOLO26 for UAV-based detection of visible Ganoderma symptoms in oil palm canopies. Its contribution lies in combining a single-stage detection architecture with plantation-block-aware evaluation, threshold-sensitivity analysis, and cross-validation to provide a more leakage-controlled assessment of model performance.]]>
