基于U-Net网络和无人机影像的松材线虫病变色木识别

松材线虫病是由松墨天牛等媒介昆虫快速传播的一种针对松树的毁灭性流行病。及时对染病变色木进行识别、定位，并砍伐清除是当前控制该病扩散蔓延的主要手段。该研究使用无人机航拍获取大区域染病松林可见光影像，结合深度学习分割网络U-Net开展染病松材变色木图像分割研究。使用制作的样本数据训练U-Net网络，得到训练精度和验证精度分别为98.74%和97.76%。使用混淆矩阵评估U-Net网络分割精度，表明变色木图像分割的用户精度和生产者精度分别达到93.51%和97.30%，误报率6.49%，漏报率2.70%。总体上，U-Net网络变色木识别精度95.17%，Kappa系数0.90，达到较高精度。U-Net分割网络运用于松材线虫病变色木图像识别较随机森林方法能更有效地降低误报，减少分割噪音。

Recognition of wilt wood caused by pine wilt nematode based on U-Net network and unmanned aerial vehicle images

Pine wilt nematode is a species of nematode that infects pine trees and causes the pine wilt disease. The affected pine will infect other pines in the short term, hence, identification and removal of the affected pines are important to control the disease. In this study, a fix-wing Unmanned Aerial Vehicle (UAV) equipped with a professional true-color camera was used to collect images for the study area, and a deep learning network, the U-Net model, was adopted to segment the images of wilt pines. Several 3 799 wilt pine images in this study were collected by the UAV, of which, 1 567 images for the Heihu mountain area and 2 232 images for Yangkou tunnel. The ground survey was conducted to collect the true wilt pines, and 45 samples were obtained. The collected images were identified by visual interpretation, and compared with the ground true samples, the accuracy of the visual interpretations was 95.24%. Label images were generated manually according to the visual interpretation results, and the label images amplification method was adopted to get more training samples. The training samples of 1 3120 were obtained in the study, and the training was performed using a GTX 2080 Ti GPU with a memory of 11GB. The training accuracy gradually increased as the number of iterations increased, and the training loss value gradually decreased as the number of iterations increased. The training and validation accuracies of the U-Net model were 98.74% and 97.76% respectively, and the problems such as over-fitting or under-fitting and gradient disappearance were not observed in the training. The confusion matrix was used to evaluate the segmentation accuracy of the U-Net model, and 145 images were randomly selected to perform the confusion matrix. The confusion matrix showed that the user and producer accuracies for the wilt wood segmentation were 93.51% and 97.30% respectively, the false positive rate was 6.49%, and the false-negative rate was 2.70%. This indicated the accuracy of the U-Net model achieved the accuracy requirements of segmentation wilt pine which required both a low false-positive rate and false-negative rate but emphasized more on the lower false-negative rate. In general, the overall segmentation accuracy of the U-Net model is 95.17% with the Kappa coefficient of 0.9. The accuracy evaluation indicated the segmentation model achieved a reasonable accuracy. A comparison with the Random Forest (RF) algorithm was also conducted to evaluate the performance difference between the RF and the U-Net model, the accuracy analysis showed that the U-Net network greatly reduced the false positive rate and the noise of segmentation compared with the random forest method. The analysis also indicated the U-Net model achieved higher accuracy than the RF model. The classification results of the RF model without using the texture information which includes Local Binary Pattern (LBP), angular second moment, contrast, correlation, variance, entropy, mean, and homogeneity, presented heavy salt and pepper noise. Although the texture information was useful for the RF method, it still showed low accuracy compared with the U-Net model. To analyze the effects of the number of samples and the quality of samples on model accuracy, a comparative analysis was conducted by reducing the training samples and degradation of the quality of the training samples. The number of training samples was reduced by 2/3, that is, 4 373 samples were randomly selected to train the model. On the other hand, a wilt wood mask with a size of 1-20 pixels was randomly added to the 50% labels for using to train the model. The results indicated that the quality of the samples was more important for the U-Net network to identify the wilt pine woods. In general, the U-Net deep learning segmentation network is suitable for monitoring wilt pine woods in large areas.