基于BP神经网络与ETM+遥感数据的盐城滨海自然湿地覆被分类

高效而精确的湿地遥感分类是大范围湿地资源动态监测与管理的必要保障。本研究使用ETM 遥感数据，借助Matlab神经网络工具箱，构建了基于BP神经网络的滨海湿地覆被分类模型，并将其应用于江苏盐城沿海湿地珍禽国家级自然保护区的核心区的自然湿地覆被分类研究中。本研究选择3、4、7、8波段作为输入层变量，单隐藏层设为10个节点，输出层变量对应待划分的8种覆被类型，构建三层式BP神经网络滨海湿地覆被分类模型。结果显示，BP分类总精度为85.91%，Kappa系数为0.8328，与最小距离法和极大似然法的分类总精度相比，分别提高了7.99%和6.08%，Kappa系数也相比提高。研究结果表明，BP神经网络分类法是一种较为有效的湿地遥感影像分类技术，能够提高分类精度。

Land cover classification of Yancheng Coastal Natural Wetlands based on BP neural network and ETM+ remote sensing data

It is necessary to classify wetland remote sensing efficiently and accurately for monitoring and management of the wetland resources. In this study we used ETM+ (Enhanced Thematic Mapper) remote sensing data from the United States' Landsat-7 satellite, after strip processing, to build a coastal wetland classification model. This was based on a back-propagation (BP) neural network using the Matlab neural network toolbox (late 2010 version). The model was applied to natural wetland cover classification research in the core area of the Yancheng National Natural Reserve for Coastal Rare Birds. The natural cover of the study area can be divided into eight types: Spartina alterniflora, Suaeda glauca, Imperata cylindrica, Phragmites australis, Sandy beach, Muddy beach, Pond water and Shallow water.The choice of input layer variables for the BP neural network, the hidden layer set and the optimization algorithms, were quite different from previous studies and this impacted directly on the efficiency and accuracy of classification. In this study we conducted the following analysis. First, by the analysis of single-band information quantity and the correlation among bands, band 3, band 4, band 7 and band 8 were chosen as input layer variables for the BP neural network and then fused with each other. This achieved a remote sensing image resolution of 15m×15m. Second, by comparing the training accuracies of the BP neural network with 2 to 17 single hidden-layer nodes, 10 single hidden layer nodes were defined for the model. Third, the output layer variables of the BP neural network were matched to the 8 natural wetland cover types into which the area is to be divided. Roughly equal numbers of training samples were chosen for each type, with the total number of training samples reaching 900 pixels. Finally, a cover classification model for coastal wetlands based on three-layer BP neural network was built, and cover classifications were completed for the research area. In addition, we used ENVI 4.8 software to make cover classifications of the research area by the Minimum Distance method and the Likelihood Classification method, on the premise that the training sample nodes were unchanged. We used an Artificial Visual Interpretation method to get standard classifications for the research area, based on field surveys. We calculated interpretation accuracies of the previous three classification results, compared with the standard classification results.The results showed that this coastal wetland classification model provides efficient land cover classification of the Yancheng Coastal Natural Wetlands. The overall accuracy of the BP classification was 85.91%, and the Kappa coefficient was 0.8328. Compared with the Minimum Distance method and Likelihood Classification method, the total classification accuracy was 7.99% and 6.08% higher, respectively. The Kappa coefficient was also increased. Therefore, the classification method of BP neural network provides a more effective wetland remote sensing image classification technology that can improve the accuracy of classification. In future studies, other geographic information should be considered in the input layer variables for the BP neural network, and other, better, artificial neural network models can be chosen.