| Pine wilt nematode disease is a devastating disease of pine trees,which can cause the death of a large number of pine trees in a short period of time.Timely and accurate identification and location of diseased pine dead trees,as well as their harmless disposal,are crucial measures to prevent the spread of pine wilt disease.The combination of unmanned aerial vehicle(UAV)remote sensing technology and deep learning technology can quickly and effectively identify and locate dead pine trees in the field,which has become a hot topic for the current research on pine dead tree monitoring.The existing deep learning models have high accuracy in High-precision drone aerial image for identifying dead pine trees,but due to the small aerial photography area and limited datasets,it is not possible to cope with the monitoring scene of large-scale pine dead trees efficiently and accurately.Based on this,this study collected large-scale UAV remote sensing images for practical application scenarios in the field,The total flying area is 344.71km~2.Established a large-scale wild pine dead tree data set,and then used a deep learning model to train it.While optimizing the model parameters,it improved the issues of missed detections and false detections,so as to develop a set that can be applied to production practice.The monitoring and identification technology of pine dead trees provides technical support for the prevention and control of pine nematode disease.The main research results are as follows:1.Pine dead tree identification model and its optimal parameter screeningA large sample of dead pine tree dataset was collected and established,and pine dead trees were boxed and verified in the field using manual labeling.Six deep learning models were used to train the pine dead tree dataset,and it was found that the YOLOv5 object detection model had the best performance.The three parameters of model learning rate,confidence threshold and training epochs are further optimized.It was found that the optimal performance of the model could be achieved with a learning rate of 0.0001,a confidence threshold of 0.4,and training for 60 epochs.Using data sets of different data volumes to train the model,it is found that with the increase of data volume,the performance of the model is gradually improved,indicating that the increase of data volume effectively improved the recognition accuracy of the model.2.Construction of pine dead tree recognition method based on different color spacesTo improve the issue of missed detections in the process of identifying pine dead tree recognition,a pine dead tree recognition method based on multi-color space fusion was proposed.The RGB,Lab and HSV color space results are integrated,and the non-maximum suppression algorithm is used to remove the output results of repeated selection boxes.Compared with the single color space results,the recall rate and AP were improved,the recall rate increased by 4.80%,and the AP increased by 1.20%.Indicating that the multi-color space fusion model could reduce missed detection.The performance of the model was upgraded,but the false detection of the model increased.Visual error analysis showed that the main sources of errors are roofs and features that are close to the color or texture characteristics of pine dead tree samples.3.Pine dead tree identification error screening and application evaluationAiming at the problem of false detections of dead pine trees and the main error sources of roofs and objects,a method of using edge detection algorithm to find the difference between error samples and samples of pine dead trees was further proposed.By using the Canny edge detection operator,the low threshold is set to 175 and the high threshold is set to255 to distinguish the pine dead trees from the error samples.Further use of the straight-line detection method can screen the error sample,so as to improve the recognition accuracy of the pine dead tree.After removing the error sample,multiple data sets were used to test the application performance of the model,the results showed that multi-color space fusion can effectively improve the recall rate of the model,and edge detection and linear detection have significantly improved the precision rate of the test area.The UAV aerial image pine dead tree recognition method established in this study can be applied to production practice.The recognition accuracy is higher than 80%,and low rates of false negatives and false positives,which has a good application prospect. |
