基于深度学习的高分辨率卫星影像地表覆盖分类方法

Land cover classification of high-resolution remote sensing imagery in further learning

针对高分辨率卫星影像地表覆盖自动化分类精度与效率问题,在深度学习的地表覆盖分类方法的基础上设计了一种适用于高分辨率卫星影像的卷积神经网络模型(LCC-CNN)。用高分二号与北京二号影像数据构建训练样本集,在卷积神经网络中使用不同扩张率的扩张卷积,设计能够区分模糊地表覆盖分类边界的损失函数,实现了多尺度地表覆盖特征的融合与精准提取,利用编码-解码结构输出像素级地表覆盖分类成果。实验结果表明:LCC-CNN在实验区域地表覆盖分类总体精度达到87.17%,IOU及Kappa系数分别为0.7732与0.8291,精度优于传统的决策树与SVM方法8%以上,IOU及Kappa系数分别提高了10%和11%,且在地表覆盖分类过程中不需要人工提取分类特征与设定分类参数,降低了建模难度与时间成本,提高了自动化分类的精度与效率。

For improving the classification accuracy of high-resolution remote sensing imagery,an accurate and automatic land cover classification convolutional neural network(LCC-CNN)for high resolution remote sensing imagery based on deep learning is proposed.The training dataset is constructed by GF-2 and BJ-2 image.The LCC-CNN is designed with different sizes of dilated convolution in extracting multiscale features from remote sensing imagery.A novel loss function is designed for distinguishing the fuzzy boundary among different land cover classification regions.With the structure of encoder-decoder,LCC-CNN export pixel-level classification result,the experiments show that the overall classification accuracy,IOU,and Kappa coefficient of LCC-CNN are 87.17%,0.7732,and 0.8291.Compared with the existing approaches,such as SVM and decision tree,the accuracy,IOU,and Kappa coefficient are improved by 8%,10%,and 11%.The experiments also prove that LCC-CNN has strong applicability for high-resolution image classification without designing classification features and parameters.