Enhanced Small Lesion Segmentation in Apple Leaves: A Multimodal Prior‐Guided Approach With PEMA‐UNet++

ABSTRACT Apple leaf diseases are widespread and reduce yield and fruit quality. Orchard images often contain cluttered backgrounds, varying illumination, and small, low‐contrast lesions. As a result, existing segmentation networks still miss lesions and produce false positives. This work focuses on accurate segmentation of small apple leaf lesions in complex field scenes. We propose PEMA‐UNet++, a UNet++ architecture that integrates a lesion prior, global context modeling and edge‐aware attention. A DINOv3‐derived prior is fused with RGB inputs through a PriorGate to guide the encoder toward lesion regions, while a Mamba2D state‐space module at the bottleneck models long‐range dependencies and injects global semantics into decoder stages via AlphaMapFusion. An edge pyramid drives edge‐aware attention to refine lesion contours and suppress background responses. On the ATLDSD dataset, PEMA‐UNet++ achieves a Boundary F1 of 94.32%, a Dice of 92.12%, an mIoU of 92.55%, and an AUPRC of 97.32%, while running at 347.81 FPS with an average inference time of 2.88 ms/image. On the Apple Disease Dataset, it attains a Boundary F1 of 87.51%, a Dice of 82.24%, an mIoU of 84.32%, and an AUPRC of 90.63%, while running at 357.96 FPS with an average inference time of 2.79 ms/image. The parameter count and computational cost are 30.35M and 23.88G FLOPs, respectively. These results indicate that PEMA‐UNet++ provides efficient and stable segmentation of apple leaf lesions in orchards and is a practical tool for automated disease monitoring and decision support in smart agriculture.