基于ResNet34的高分遥感影像草本湿地信息提取

Extraction of Herbaceous Wetland Information from High Resolution Remote Sensing Image Based on ResNet34

以宁夏石嘴山Landsat8影像、高分二号影像、LIDAR数据插值生成的DEM数据和地理国情普查数据等为数据源,首先利用一年多期Landsat8影像确定提取草本湿地的最佳时相;然后对最佳时相的高分二号融合影像、DEM数据等进行多尺度叠置分割,选取NDVI、湿度、坡度、色调等特征,采用决策树分类获得草本湿地信息,构建草本湿地样本库;最后探讨卷积神经网络ResNet34用于高分卫星影像提取草本湿地的方法。试验结果表明:设计的ResNet34方法同决策树分类法提取结果相比,提取的效果明显提高,适用于草本湿地信息提取。

Based on Landsat8 image,GF-2 image,DEM data generated by lidar data interpolation and census data of geographical conditions in Shizuishan,Ningxia,the best time for extracting herbaceous wetland was determined by using Landsat8 image of more than one year.Secondly,multi-scale overlay segmentation was carried out on the best time phase of GF-2 fusion image and DEM data,and NDVI,humidity,slope,hue and other features were selected to obtain herbaceous wetland information by decision tree classification,and herbaceous wetland sample database was constructed.Finally,the convolution neural network ResNet34 is used to extract herbaceous wetland from high resolution satellite images.Compared with the decision tree classification method,the experimental results show that the ResNet34 method is more effective and suitable for herbaceous wetland information extraction.