基于改进元胞卷积网络的遥感图像识别

Remote Sensing Image Recognition Based on Improved Cellular Convolutional Network

为了加强卷积神经网络结构构建与训练效率,提出了改进的元胞卷积网络。首先分析了遥感地物间的差异性,实现像元邻域扩展,挖掘地物间邻域信息;进而以3层卷积神经网络作为元胞单元,构建元胞卷积网络集成框架,通过Boosting集成方式对各元胞单元识别结果进行决策级融合,获得最终识别结果。针对高分遥感图像的玉米种植区域进行识别预测,以人工标注真值为标准进行对比验证。结果表明,本方法能够精确地实现高分遥感图像玉米种植区的识别分类,总体精度达到0.94,比BP网络提高了20%以上,其他各项评价指标也得到大幅提升。

In order to strengthen the construction and training efficiency of convolutional neural network,an improved cellular convolutional network was proposed.Firstly,the difference between remote sensing features was analyzed to realize pixel neighborhood extension and excavate neighborhood information among features.In addition,the three-layer convolutional neural network was used as the cell unit to build the integration framework of the cellular convolutional network.Through boosting integration,the recognition results of each cell unit were fused at the decision level to obtain the final recognition results.The corn planting area of the remote sensing image recognition was predicted and compared the artificial mark true value for verification,the experimental results show that this method can accurately realize high score of remote sensing image recognition classification of corn belt,the overall accuracy reaches 0.94,more than 20%higher than that of BP network,another various evaluation indexes have been also improved.