High Resolution Remote Sensing Image Scene Classification Based On Deep Learning

For a long time,humankind has been full of interest in the exploration of the unknown.More and more sensors have been developed to observe the unknown universe,and observation of the earth is especially important.High spatial resolution remote sensing images have been a popular research topic because of high spatial resolution,large coverage and repeating at multiple angles.Due to the large format and periodic shooting of remote sensing images,large amounts of data are generated.To ful y utilize remote sensing images,it is necessary to classify remote sensing images according to the scenes.This research focuses on the scene classification of highresolution remote sensing imagery.Based on this,several methods are proposed in this thesis for high spatial resolution image scene classification,and it involves:(1)Recently,Convolutional Neural Networks(CNN)have been found to be excelent for scene classification.However,only using the deep models as feature extractor on the aerial image directly is not proper,because the extracted deep features can not capture spatial scale variability and rotation variability in HSR remote sensing images.To relieve this limitation,a bidirectional adaptive feature fusion strategy is investigated to deal with the remote sensing scene classification.The deep learning feature and the SIFT feature are fused together to get a discriminative image presentation.The fused feature can not only describe the scenes effectively by employing deep learning feature but also overcome the scale and rotation variability with the usage of the SIFT feature.By fusing both SIFT feature and global CNN feature,our method achieves state-of-the-art scene classification performances on the UCM,the Sydney and the AID datasets.(2)Using deep learning method to directly solve the scene classification problem of high spatial resolution remote sensing image,it is necessary to enhance the ability to express rotation and scale information.To design a new deep net,rotation and scale information can be included as a priori information in the deep net.In this paper,an active rotation convolutional kernel direction invariant network has been proposed to solve the scene classification problems.The convolution kernel can generate convolution kernels with multiple direction channels during convolution operation.At the same time,multiple directional feature maps are generated.Afterwards,the maximum pooling method is used to maximize the pooling of feature maps on the direction channel to ensure that the generated rotation invariant information is included in the deep network,so that the deep network has the ability to express the rotation invariant.The proposed method achieves state-of-the-art scene classification performances on the UCM,the Sydney and the AID datasets.