Detection Of Pine Trees Infected With Pine Wood Nematode Disease Based On Improved SSD Algorithm

Pine wood nematode disease is a forest disease that causes the rapid death of pine trees,which seriously threatens the pine forest resources in China.Regular detection of diseased pine trees in disease-affected areas to achieve the purpose of dynamic monitoring is of great significance for the development of pine wood nematode epidemic prevention and control.At present,the detection of diseased pine trees has problems of low accuracy and efficiency.In this thesis,taking a forest farm in Yangshan Country,Guangdong Province as an example,the optimization and improvement of deep learning algorithms are carried out to improve the detection level of pine trees infected with pine wood nematode.Aiming at the problem of missed detection of diseased pine trees by the algorithm,a method combining feature fusion module(FFM)and channel attention mechanism module(Squeeze and extraction module,SEM)was proposed to strengthen the detection of smaller diseased pine trees.In addition,by comparing with a variety of algorithms,the feasibility of the improved SSD algorithm to replace the object-oriented method to detect individual diseased pine trees is verified.The main work and achievements of this thesis are as follows:1.In this thesis,Faster R-CNN,SSD and YOLO v4 are used to detect the diseased pine trees in the epidemic area and carry out comparative analysis.Firstly,three algorithms were used to train the model of the self-made infected pine dataset to obtain the best weight file,and then the best weight file of different algorithms was used to evaluate the accuracy of the algorithm for detecting the infected pine tree: the average accuracy and F1 of the SSD algorithm were 79.59% and 77.00%.It is 5.61% and 9.77%higher than Faster R-CNN algorithm.it is 2.15% and 3.89% higher than YOLO v4 algorithm respectively.The SSD algorithm is more suitable for detecting infected pine trees than the Faster R-CNN algorithm and the YOLO v4 algorithm.2.Aiming at the problem of missed detection when the SSD algorithm detects small target objects,a method of combining feature fusion module and attention module is proposed.Feature fusion module integrates deep feature semantic information into shallow features,enriches the semantic information of shallow features,and strengthens the detection ability of small target objects.The attention module adjusts the channel weight of shallow features,strengthens the channel weight of rich information,inhibits the channel weight of other information,and improves the recognition ability of small target objects.Using the improved SSD algorithm to perform performance test on the self-made diseased pine data set,the test results: the average precision and F1 of the improved SSD algorithm are 81.83% and 77.00% respectively,which are 2.24% and 2%higher than the SSD algorithm,respectively.The improved method can effectively improve the average accuracy and F1 of the SSD algorithm for detecting diseased pine trees.3.Ablation experiments were carried out to improve the detection performance of infected pine trees by using different modules to improve SSD algorithm.Compared with the SSD algorithm,the SSD algorithm is improved by using a single module: the average accuracy of the SSD algorithm improved by the feature fusion module is increased by0.85 % and the F1 is increased by 1.00%;the average accuracy of the SSD algorithm improved by the attention module is decreased by 0.95 % and the F1 is increased by 3 %.Compared with the SSD algorithm proposed in this thesis,the average accuracy is increased by 2.24% and F1 by 2.00%,which shows that the improved method proposed in this thesis can effectively combine the advantages of feature fusion module and attention module,and further improve the average accuracy and F1 of SSD algorithm to detect infected pine trees compared with a single module.4.Compared with a variety of algorithms,the improved SSD algorithm proposed in this thesis is verified to be effective in detecting infected pine trees in epidemic areas.Faster R-CNN,SSD,YOLO v4,VEG-OSTU and improved SSD were used to detect infected pine trees in the verification area,and the number of infected pine trees was counted.92 infected pine trees on the ground in the verification area,81 infected pine trees were correctly detected by SSD algorithm,59 infected pine trees were correctly detected by YOLO v4 algorithm,70 infected pine trees were correctly detected by Faster R-CNN algorithm,50 infected pine trees were correctly detected by VEG-OSTU method,and 87 infected pine trees were correctly detected by improved SSD algorithm.The results show that the improved SSD algorithm proposed in this thesis has the highest accuracy for single tree detection of infected pine trees in the epidemic area,and can fully replace the object-oriented method used previously.