Wild arrowhead (Sagittaria trifolia L.) often occurs as isolated plants or compact clusters in paddy fields, yet these states are difficult to distinguish in unmanned aerial vehicle (UAV) imagery because they share similar color, texture, and leaf morphology. This study presents a two-stage framework in which an unchanged YOLOv8n detector localizes candidate targets and a dedicated Patch-cls network refines the single- or multiple-plant label. The classifier combines multi-level features, local multi-scale enhancement, and channel attention; a GhostConv variant is also evaluated to examine the efficiency trade-off. Annotation-box and detector-generated-box results are reported separately, followed by a complete-system evaluation that retains missed targets, false positives, duplicate detections, localization errors, and classification errors. Across three random seeds, the proposed Patch-cls obtained 94.00 ± 0.34% accuracy, 83.81 ± 1.19% Macro-F1, and 73.18 ± 3.91% multiple-class Recall. In the complete test pipeline, Macro-F1 increased from 0.5459 to 0.5539 and multiple-class F1 from 0.4224 to 0.4384, while mean average precision at an intersection over union (IoU) of 0.50 (mAP50) decreased slightly from 0.7023 to 0.7016. The optimized pipeline achieved 88.89 frames per second (FPS) on an NVIDIA RTX A4000 with approximately 1.62 GB peak allocated graphics processing unit (GPU) memory. The results indicate that Patch-cls can improve category balance under detector-generated crops, although the overall gain is modest and does not replace the need for stronger localization and dense-target separation.