半密集连接的U型卷积神经网络红树林提取方法

A mangrove extraction method combined with semi-dense connections by convolutional neural network

针对现有研究中少有针对深层分割卷积神经网络构型进行深入改进的问题,该文提出了一种半密集连接策略,以湛江市沿海5 km范围的红树林为研究区,以Sentinel-2中分辨率多光谱遥感影像为原始数据,并基于U型卷积神经网络分割架构构建了一种红树林提取方法。结果表明:该文所提方法比其他方法效果更好。网络在测试集中获得90.96%的红树林提取精度,高于基于面向对象的支持向量机和面向对象的随机森林方法。使用该文方法提取了2020年湛江市红树林分布,获得89.65%的精度。

Aiming at the problem that there were few in-depth improvements for deep segmentation convolutional neural network configuration in existing research,this paper proposed a semi-dense connection strategy,and the mangrove forest within 5 km of the coast of Zhanjiang was taken as the research area,the medium-resolution multispectral remote sensing images were used as raw data,and a mangrove extraction method was constructed based on the U-shaped convolutional neural network segmentation architecture.The results showed that the proposed method was better than other methods.Our network obtained 90・96%mangrove extraction accuracy in the test set,which was higher than the object-oriented support vector machine and object-oriented random forest methods.Using the method in this paper,the distribution of mangroves in Zhanjiang city in 2020 was extracted,and an accuracy of 89.65%was obtained.