【摘要】 针对互花米草（Spartina alterniflora）传统实地调查方法工作量大、成本高、效率低等问题，该研究基于遥感技术，探讨适用于林业基层互花米草调查工作的方法。研究基于高分辨率卫星影像，采用ArcGIS Pro软件中集成的深度学习工具对台山市遭受互花米草入侵的海岸带进行语义分割，在缺少编程先验知识的情况下快速识别互花米草图斑。测试集结果显示，U-Net和DeepLabV3模型皆能在拥有更高空间分辨率的北京二号影像上取得良好的互花米草分类精度，Accuracy分别为97.24%、97.72%,F1 Score分别为0.68、0.79，综合表现DeepLabV3模型更优；而两者在PlanetScope卫星影像上表现均不理想。为分类结果设置适当的缓冲区，Recall提升，对实际作业更有指导意义。该研究方法识别互花米草的精度高，技术易于推广，可有效提高互花米草调查工作效率，为互花米草治理工作提供数据支持。更多还原

【Abstract】 In view of the problems of large workload, high cost, and low efficiency of the traditional field survey method of Spartina alterniflora, this research is based on remote sensing technology to explore methods suitable for the survey of Spartina alterniflora at the forestry grassroots level. Based on high-resolution satellite images, the study used the deep learning tools integrated in ArcGIS Pro software to semantically segment the coastal zone of Taishan City invaded by Spartina alterniflora, and quickly identify the sketch spots of Spartina alterniflora without prior knowledge of programming. The test set results show that both U-Net and DeepLabV3 models can achieve good classification accuracy of Spartina alterniflora on the Beijing-2 image with higher spatial resolution. Accuracy is 97.24% and 97.72% respectively, and F1 Score is 0.68 and 0.79 respectively. The comprehensive performance of DeepLabV3 model is better; both of them perform poorly on PlanetScope satellite images. Set up appropriate buffers for classification results and improve Recall, which is more instructive for actual operations. This research method has high accuracy in identifying Spartina alterniflora and is easy to popularize. It can effectively improve the efficiency of Spartina alterniflora investigation and provide data support for Spartina alterniflora management work.更多还原