融合压缩与激励的GoogLeNet模型云检测算法

GoogLeNet Network Cloud Detection Algorithm by Fused Compression and Excitation

Landsat系列陆地卫星受大量云和云阴影遮盖,干扰了地面信息的提取。因此,有效的云检测是遥感影像资源利用的前提。使用SE-GoogLeNet模型进行Landsat8遥感影像的云检测工作。SE-GoogLeNet模型有9个InceptionV3块,它起到了多尺度融合的作用,获取高级语义信息与低级空间信息相结合的特征,在每个InceptionV3后添加添加SeNet(压缩与激励)模块,通过自身的信息学习通道注意力权重,自动学习Landsat8遥感影像11个波段的相互依赖关系,每个波段的重要程度,然后按照这个重要程度提升有用特征,抑制无用的特征。通过实验可视化和评估指标的对比,SE-GoogLeNet模型云检测比GoogLeNet模型云检测准确率精度等均有提升。

Land satellites of Landsat series are covered by many clouds and cloud shadows,which interfere with the extraction of ground information.Therefore,effective cloud detection is the premise of remote sensing image resource utilization.The SE-GoogLeNet model is used for cloud detection of Landsat 8 remote sensing images.The SE-GoogLeNet model has nine Inception V3 blocks,which play the role of multi-scale fusion.It obtains the features of combining high-level semantic information and low-level spatial information.After each Inception V3,a SeNet(compression and excitation)module is added.SeNet automatically learns the channel’s significant weight.The Landsat 8 remote sensing image contains 11 bands that play different roles in cloud detection.According to this weight of importance,valuable features are promoted,and useless features suppressed.By comparing experimental visualization and evaluation indicators,the accuracy of cloud detection of the SE-GoogLeNet model has been improved compared with the GoogLeNet model.