Research On High Resolution Satellite Image Classification Based On Deep Learning

Satellite images have rich feature information,and the classified image data is used in many industries.The performance of traditional extraction methods depends on the selection of manual features and cannot be adapted to complex high-resolution images of large samples.Compared with traditional satellite image extraction methods,deep convolutional neural networks have achieved good results in computer vision work such as target detection,semantic segmentation,and image classification due to their excellent feature representation capabilities.However,the current deep learning methods for classifying target images have the problems of poor classification of small targets and excessive segmentation.In view of the above problems,this paper proposes a new network model AA-SegNet for satellite image target classification based on a large number of domestic and foreign studies in the field of satellite image target classification.The network model is based on the improved attention and enhanced spatial pyramid mechanism of the SegNet architecture.Combined.The main structure of the AA-SegNet network model consists of three parts.The first part is the encoding network.The AA-SegNet structure is improved to add an enhanced spatial pyramid module A-ASPP at the end of the encoder.The parallel expansion convolution in the A-ASPP structure uses different expansion factors to obtain more dense sampling and collection.Higher-level local information,and aims to accurately extract small targets.In order to improve learning ability,the A-ASPP structure uses an expansion factor to expand and then reduce to maintain the advantage of multi-scale information acquisition.The expansion factor gradually expands,making the receptive field more dense and able to feel more detailed background information;then reducing the expansion factor can aggregate local information to enhance feature extraction of small objects.The second part is the decoding network.Non-linear upsampling is performed on the decoding network.The structure is changed to establish a spatial attention fusion module between the convolutional layer of the encoding network and the decoder upsampling.The spatial attention fusion module is used to guide low-level feature maps to help high-level features restore pixel positioning details and reduce over-segmentation of recognized images.The fusion structure of the attention mechanism in the decoder aims to enhance the feature propagation and can effectively transmit higher-level feature information to suppress the noise of low-level features.The third part is the random field of the fully connected condition of the post-processing module.Because of the effect of image post-processing stitching,there are some holes and fragments.After the AA-SegNet network model,fully connected CRFs are added for post-processing to punish small segmentation areas.In this paper,based on the high-resolution No.2 remote sensing image as the original data source,through radio calibration,band fusion,data cutting,down-sampling,and data enhancement techniques,data sets of different orders of magnitude and different attributes are produced,and the data sets are input to AA-SegNet For training.The experimental results show that the overall recognition accuracy rate of the AA-SegNet network is 96.65%,and it is also better than the SegNet,U-Net,DeepLab-V3 networks in terms of recognition rate,F1 score and training time,and AA-SegNet can reduce satellite images.Excessive segmentation,and can strengthen the recognition of small targets,this method has some practical value in the field of remote sensing image recognition.