| Pine nematode disease is a devastating disease,which seriously threatens the resource security of pine forests in China.Timely detection and treatment of infected trees is an effective way to control the spread of disease.Therefore,the monitoring technology of pine nematode is very important to the success of disease control.Traditional manual monitoring methods are time-consuming,laborious and inefficient.Satellite based remote sensing methods are susceptible to climate disturbance and low spatial resolution;Drone surveillance methods are vulnerable to weather and cannot realize continuously monitor.To this end,this thesis uses the high platform(e.g.,iron tower)based video surveillance equipment to monitor the forest area extensively.Through deep learning,incremental learning and edge computing,the automatic monitoring system is constructed,which is helpful to improve the monitoring efficiency of pine nematode tree and to control pine nematode disease effectively for the forest protection department.In general,the main tasks of this thesis are given as follows:(1)Two-stage pine nematode tree detection method based on improved YOLOv5sAs vast data will be collected to monitor the wide-area forest areas,a large number of which are redundant without diseased images,and all these images are sent back for further analysis will occupy excessive network and computing resources,a two-stage pine nematode detection method based on improved YOLOv5 s is proposed to solve these problems.In the first stage,as the embedded edge computing equipment cannot efficiently support the traditional detection model,the model can be lightweight by introducing mixed network module and Stem module,so that it can run efficiently in the embedded edge computing equipment and filter the redundant data quickly.In the second stage,CE-Block module is adopted and a multi-dimensional scaling coefficient is introduced to improve the accuracy.Based on the improved YOLOv5 s two-stage pine nematode tree detection method,the redundant data can be effectively filtered,the network,storage and computing resource utilization can be improved,and the detection of pine nematode tree can be achieved in time.(2)Pine nematode tree detection based on incremental learningDue to the different characteristics caused by spatial and temporal divergence of pine nematode trees,the models detection performance is reduced.The traditional methods of updating the model tend to lead to catastrophic forgetfulness or data integration.This thesis combines important parameter consolidation method,representative sample construction method and knowledge distillation technology to update the model incrementally.(3)Design and implementation of automatic monitoring systemThe system is composed of high platform based cloud camera and embedded edge computing equipment.The programs for the edge computing platform and user end are both written by Py Qt development framework,based on which the automatic high platform based monitoring system of pine nematode tree is set up.Experience shows that the system can automatically complete the functions of data collection,identification,transmission,display and management,which can meet the actual monitoring needs and realize efficient monitoring.The high platform based automatic monitoring system proposed in this thesis can realize the real-time monitor of the wide-area forest in a long distance and time manner,which can improve the detection accuracy and efficiency,and provide technical support for the forest protection department to acquire the dynamics of forest areas and control the occurrence of pine nematode disease effectively. |
