Abstract: In order to improve the inspection efficiency of large domestic waste sanitary landfill for diseases such as puncturing and tearing of the covering film, the Xingfeng landfill in Guangzhou was taken as an example. The selection and cruise parameters of unmanned aerial vehicles (UAV) were investigated based on UAV aerial photography and depth learning model, the disease samples collected by UAV were trained by oversampling strategy in the disease image recognition model. A subtle target recognition layer was added to the YOLOv5 model, which achieved high recognition accuracy and recall rate. The test results showed that the oversampling strategy improved the representativeness and balance of the samples and significantly improved the accuracy of disease identification of the model under the condition that the number of samples in the disease map database was relatively limited. UAV combined with high-precision RTK positioning technology could accurately locate the coordinates of aerial photos, solved the problem of difficult positioning of film coating diseases without obvious reference objects. By querying the geographic information of disease defect photos, the location of the disease could be quickly found and repair operations could be carried out in a timely manner.

Key words: landfill, HDPE film, patrol inspection, disease identification, transfer learning, unmanned aerial vehicles