Abstract: [Purposes]Well-facilitated farmland infrastructure currently suffers from inadequate post-construction maintenance.While UAV inspection combined with AI analysis offers a viable solution,it is hindered by issues such as insufficient scenario adaptability and suboptimal detection performance.[Methods]This paper proposes a scenario-based,multi-level visual detection framework.For common defects,taking field roads as a case study,we employ the YOLO11 object detection model and introduce the Wise-IoU loss function to enhance detection accuracy.For complex anomalies in specific facilities,taking outlet protection piers as an example,we integrate the precise localization capability of YOLO11 with the semantic understanding of the vision-language model (VLM)Qwen3-VL to construct a few-shot learning detection method.[Findings]Experiments demonstrate that on the field road defect dataset,the proposed method achieves an 8.3 percentage point improvement in mAP₅₀ over the baseline model.On the outlet protection pier dataset,the precision and recall rates reach 86% and 96%,respectively.[Conclusions]This research provides a flexible and reliable analysis method for UAV inspections of well-facilited farmland infrastructure,holding significant promise for engineering applications.

Key words: well-facilited farmland, post-construction maintenance, UAV inspection, object detection, vision-language model