欢迎访问林业科学,今天是

林业科学 ›› 2020, Vol. 56 ›› Issue (8): 98-106.doi: 10.11707/j.1001-7488.20200812

• 论文与研究报告 • 上一篇    下一篇

基于FCN与面向对象的滨海湿地植被分类

谢锦莹1,2,3,丁丽霞1,2,3,*,王志辉4,刘丽娟1,2,3   

  1. 1. 浙江农林大学省部共建亚热带森林培育国家重点实验室 杭州 311300
    2. 浙江省森林生态系统碳循环与固碳减排重点实验室 杭州 311300
    3. 浙江农林大学环境与资源学院 杭州 311300
    4. 浙江远卓科技有限公司 杭州 310012
  • 收稿日期:2018-04-25 出版日期:2020-08-25 发布日期:2020-09-15
  • 通讯作者: 丁丽霞
  • 基金资助:
    浙江省自然科学基金项目(LY18D010002)

Classification of Coastal Wetland Vegetation Utilizing FCN and Object-Oriented Methods

Jinying Xie1,2,3,Lixia Ding1,2,3,*,Zhihui Wang4,Lijuan Liu1,2,3   

  1. 1. State Key Laboratory of Subtropical Silviculture, Zhejiang Agriculture and Forestry University Hangzhou 311300
    2. Key Laboratory of Forest Ecosystem Carbon Cycle, Sequestration and Emission Reduction in Zhejiang Province Hangzhou 311300
    3. School of Environmental and Resource Sciences, Zhejiang Agriculture and Forestry University Hangzhou 311300
    4. Zhejiang Yuanzhuo Technology Co. Ltd Hangzhou 310012
  • Received:2018-04-25 Online:2020-08-25 Published:2020-09-15
  • Contact: Lixia Ding

摘要:

目的: 提出一种基于全卷积神经网络(FCN)与面向对象的滨海湿地植被分类方法,以提高滨海湿地植被监测效果。方法: 以浙江慈溪部分杭州湾滨海湿地为研究区,基于高分辨率QuickBird影像,采用FCN与面向对象相结合的方法监测滨海湿地植被:1)融合QuickBird影像的多光谱数据和全色数据提高影像空间分辨率,运用目视判读制作标签图,以100×100窗口选取样本后进行翻转、旋转等操作,获得训练样本4 904对、测试样本544对,采用FCN对样本完成训练后得到相应的模型参数用于整幅影像,获得全图分类结果;2)对原始影像进行多尺度分割,利用平均全局评分指数法确定最优分割尺度为170,以最优分割结果对FCN分类结果进行边界约束,得到最终分类结果并制作滨海湿地植被分类图;3)基于混淆矩阵对仅采用FCN处理的结果影像及采用FCN与面向对象相结合处理的结果影像进行精度评价。结果: 1)采用FCN处理的影像分类总体精度达94.39%,典型滨海湿地植被精度均在85%以上,但分类结果存在少量椒盐现象,分类误差产生的主要原因是滨海湿地下垫面背景复杂,不同植被类型空间分布杂乱;2)将面向对象与FCN相结合处理的结果影像可消除椒盐现象,总体精度达97.56%,典型滨海湿地植被精度均在90%以上。结论: 基于FCN的滨海湿地植被分类方法能够有效从高分辨率影像中提取典型滨海湿地植被信息,在此基础上结合面向对象的多尺度分割方法可有效消除椒盐现象,弥补基于像元分类的缺陷,优化滨海湿地植被分类结果,在滨海湿地植被监测方面值得推广和运用。

关键词: 滨海湿地植被, 遥感监测, 全卷积神经网络, 面向对象, 高分辨率

Abstract:

Objective: Monitoring wetland vegetation is of great significance for the protection and management of coastal wetlands. In order to improve the monitoring effect of coastal wetland vegetation, this paper attempted to propose a more effective method. Method: A part of Hangzhou Bay coastal wetland in Cixi, Ningbo was taken as the study area. Based on high spatial resolution satellite images of QuickBirds, the method of combined fully convolutional networks (FCN) and object oriented was applied to monitor coastal wetland vegetation. The images were improved the spatial resolution by integrating the multispectral data and panchromatic data, the label maps were made by the way of artificial visual interpretation, the training samples were selected with a 100×100 window, then were flipped and rotated, finally 4 904 pairs of training samples and 544 pairs of test samples were prepared. The corresponding model parameters were applied to the whole image after the FCN model training is completed with samples, and the classification results of the whole image were obtained. Multiscale segmentation was performed on the original image, and the optimal segmentation scale was determined to be 170 by using the average global score index method. The classification results of FCN were bounded by the optimal segmentation results, and the final classification results were obtained and the coastal wetland vegetation classification map was made. The confusion matrix was used to evaluate the final accuracy. Accuracy evaluation was performed on the results from FCN and the results processed by the FCN combined with the object-oriented method. Result: The overall accuracy of image classification results from the FCN is 94.39%, the accuracy of the typical wetland vegetation is more than 85%. However, there is a phenomenon of "salt and pepper", the main reason for the error is that the background of the coastal wetland is complex, and the spatial distribution of different vegetation types is chaotic. The result image after combining the two methods reduces the salt and pepper, the overall precision is 97.56%, the accuracy of typical wetland vegetation is above 90%. Conclusion: The coastal wetland vegetation monitoring method based on the FCN could effectively extract the typical wetland vegetation information from the high resolution images. On the basis of this, combined with the object-oriented multi-scale segmentation method, it could effectively eliminate the salt and pepper, make up for the defects based on pixel classification, optimize the classification results of coastal wetland vegetation, and it is worthy of promotion and application in the monitoring of coastal wetland vegetation.

Key words: coastal wetland vegetation, remote sensing monitoring, fully convolutional networks(FCN), object-oriented, high resolution

中图分类号: