Abstract: The invasion of Spartina alterniflora has caused great losses to ecosystem diversity in China. How to accurately identify Spartina alterniflora in scattered patches is important for early monitoring and early warning. In this paper, lowaltitude UAV remote sensing technology was used to study Spartina alterniflora in spring mangroves in Beihai area of Guangxi. Five convolutional neural networks including AlexNet, VGG16, GoogleNet, ResNet50 and EfficientNetB0 were used. Model training, validation and testing were carried out on image data sets containing Spartina alterniflora, mangrove and other ground object backgrounds respectively, then the cultivated five network models were used to identify and mark Spartina alterniflora and ground object backgrounds in the whole experiment area, obtaining the distribution map of Spartina alterniflora. The comprehensive quantitative evaluation results based on confusion matrix and operation time showed that ResNet50 network model was superior to the other four network models in general, with the highest recognition accuracy of 96.96%, and the test set took only 5.47 s. The identification results were compared with the actual distribution of Spartina alterniflora, and the identification results of ResNet50 network model basically coincided with the actual distribution of Spartina alterniflora.

Key words: UAV remote sensing, Spartina alterniflora, convolutional neural network, scattered patches

摘要： 互花米草的侵入对我国的生态系统多样性造成了巨大损失，如何准确地识别零散斑块的互花米草对其早期监测及预警具有重要意义。采用低空无人机遥感技术，以广西北海地区春季时期红树林中的互花米草为研究对象，利用AlexNet、VGG16、GoogleNet、ResNet50、EfficientNetB0五种卷积神经网络，分别对包含互花米草、红树林及其他地物背景的图像数据集进行模型的训练、验证与测试，然后将训练好的五种网络模型对整个试验区域的互花米草及地物背景进行识别并标记，得到互花米草的分布图。基于混淆矩阵和运算时间的综合定量评估结果表明，ResNet50网络模型总体上优于另外四种网络模型，识别准确率最高，达到了96.96%，且在测试集上耗时仅为5.47 s。将识别结果图与互花米草实际分布图进行对比，ResNet50网络模型的识别结果与互花米草的实际分布基本重合。

关键词: 无人机遥感, 互花米草, 卷积神经网络, 零散斑块