• •

### 基于卷积神经网络和高分辨率影像的湿地群落遥感分类——以洪河湿地为例

1. 1.哈尔滨师范大学地理科学学院,黑龙江 哈尔滨 150025
2.长春师范大学城市与环境科学学院,吉林 长春 130032
3.中国科学院东北地理与农业生态研究所,吉林 长春 130102
• 收稿日期:2018-01-02 修回日期:2018-09-06 出版日期:2018-11-20 发布日期:2018-11-20
• 作者简介:

作者简介：孟祥锐（1981-）,女,吉林梨树人,讲师,博士研究生,主要研究方向为地理信息系统算法研究。E-mail: 32965786@qq.com

• 基金资助:
国家自然科学基金项目（41571199）资助

### Remote Sensing Classification of Wetland Communities Based on Convolutional Neural Networks and High Resolution Images: A Case Study of the Honghe Wetland

Xiangrui Meng1,2(), Shuqing Zhang3, Shuying Zang1()

1. 1.College of Geographical Sciences, Harbin Normal University, Harbin 150025, Heilongjiang,China
2. College of Urban and Environmental Sciences ,Changchun Normal University,Changchun 130032, Jilin, China
3. Northeast Institute of Geography and Agricultural Ecology, Chinese Academy of Sciences, Changchun 130102, Jilin, China
• Received:2018-01-02 Revised:2018-09-06 Online:2018-11-20 Published:2018-11-20
• Supported by:
National Natural Science Foundation of China (41571199)

Abstract:

On the high-resolution remote sensing imagery, the feature of the ground object spectra is very rich. The spectral differences are increased within the same type, differences between the categories are decreased, and the phenomenon is obvious that the same object with different spectra and the different objects with the same spectrum. Under high-resolution remote sensing imagery, the application of shallow classification models for classifying wetland remote sensing cannot unearth sufficient information, while the adoption of deep-structure models for classification of high-resolution wetland remote sensing images can fully utilize the spatial structure information of images. Deep-featured representations allow layered learning from simple concepts to higher-level complex patterns such as textures, segments, and objects. Convolutional neural network is a very representative network structure in deep learning, and it has been gradually applied to the classification of remote sensing images in the past two years. This article constructs a proper framework for the study of area image classification by selecting the size of convolution kernels and pooling kernels, classifiers and activation functions, and optimizing the number of convolutional layers. A study on the classification of different vegetation types in Sanjiang Plain Honghe National Nature Reserve is carried out through experiments. A comparison of accuracy is made between support vector machines method based on the spectrum and the support vector machines method based on the texture and spectrum. Experiments show that: 1) When adding the texture spectral features to the classification method, the classifying accuracy by SVM method based on Pixel is slightly improved compared with the SVM method based on spectrum. The CNN classification method based on patch with spectrum, texture, size, color and other semantic features, of which classification accuracy is generally higher than about 4% of the TSP-SVM method. 2) Pixel based classification methods can produce salt and pepper noise effects for misclassified pixels, and it’s hard to deal with the abundant Information from high-resolution data. The key feature of CNN’s algorithm is that it does not require manual pre-extraction of features, and the processing and classifying of images is performed in an end-to-end manner, so it has a stronger generalization ability. 3) However, in the experiment, it is also found that , on the one hand, the accuracy of the CNN classifier on the boundary is not as good as that of the SVM classifier, and there is a phenomenon of uncertainty along the boundary of the object in the classification, which leads to excessive smoothness to a certain extent. On the other hand, in the use of the CNN classifier, although the spectral features are significant, there is little spatial information and it may be misclassified. The experimental results conclude that the CNN as a deep structure classifier explores the complex spatial patterns hidden in high-resolution remote sensing images, and it can extract more abundant semantic features of the ground. Therefore, the use of the CNN method is more accurate to classify high-resolution wetland remote sensing images, and it can provide decision support for real-time monitoring, management, and protection of wetlands.

• F129.9