基于反卷积高层特征的遥感地物图像分类

目前利用深度卷积神经网络提取图像底层特征后分类效果已比较优秀,但是对于数据量大、波段多、波段间相关性高的多光谱遥感图像并非完全适用。针对多光谱遥感地物分类中常常出现的Hughes现象,即当训练样本一定时,模型的预测能力随着维度的增加而减小,提出了一种结合高层特征空间和迁移学习网络的遥感地物图像分类算法,利用两层堆叠的反卷积网络来提取目标数据集的高层特征,利用VGG16模型的卷积层权重来构建迁移网络模型,将高层特征导入迁移网络中加强训练得到更加优越的训练模型,利用训练模型可对多光谱遥感数据集更加准确分类。实验结果表明,在Satellite、NWPU和UC Merced实验数据中,冰川、建筑群和海滩分类精度得到有效提高,达到92%左右,针对沙漠、岩石、水域等特殊环境遥感图像,总体分类精度提高5%左右。部分多光谱遥感数据的底层特征和中层特征在训练器中表现并不理想,波段的增多也会导致信息的冗余和数据处理复杂性的提高,反而高层特征在这部分数据中保留了地物信息的轮廓,能更好地适应分类器,得到更加优越的分类结果。

Remote Sensing Image Classification Based on High-Level Features of Deconvolution

At present, the deep-convolution neural network is used to extract the underlying features of images, and the classification effect is quite good. However, it is not fully applicable to hyperspectral remote sensing images with large data volume, multiple bands and high correlation between bands. Aiming at the Hughes phenomenon often appearing in the classification of hyperspectral remote sensing objects, that is, when the training samples are certain, the prediction ability of the model decreases with the increase of the dimension, and a remote sensing feature combining high-level feature space and migration learning network is proposed. Image classification algorithm firstly uses the two-layer stacking deconvolution network to extract the high-level features of the target data set, then uses the convolution layer weight of VGG16 model to construct the migration network model, and finally introduces the high-level features into the migration network to strengthen the training. The training model is more advantageous, and the training model can be used to classify hyperspectral remote sensing data sets more accurately. The experimental results show that in the Satellite, NWPU and UC Merced experimental data, the classification accuracy of glaciers, buildings and beaches has been effectively improved, reaching 92%. For remote sensing images of deserts, rocks, waters and other special environments, the overall classification accuracy is improved by about 5%. The underlying features and middle-level features of some hyperspectral remote sensing data are not ideal in the trainer. The increase of the band will also lead to the redundancy of information and the complexity of data processing. On the contrary, the high-level features retain the features in this part of the data. The outline of the information can be better adapted to the classifier and get better classification results.