Ground Objects Extraction Of Optical Remote Sensing Image Based On Deep Neural Networks

In order to gather earth data,remote sensing is a highly effective technique.Optical remote sensing images record spectral information about the ground.In land resources management,atmospheric environment detection,natural disaster prevention,and many other fields,optical remote sensing image typical ground objects extraction technology is widely used.Information contained in optical remote sensing images is becoming more abundant as remote sensing technology develops and image acquisition methods improve.Due to the complexity and diversity of information,the traditional methods of extraction of typical ground objects are no longer applicable.Recently,deep learning methods have made significant progress in computer vision.In addition,optical remote sensing images and natural images possess similar data characteristics.Optical remote sensing allows extraction of typical ground objects from images with new ideas.The thesis proposes a series of methods based on deep learning in response to the abundance of data information,large image scale,and complexity of characteristics,which is committed to improving the accuracy in identifying typical ground objects from optical remote sensing images.The main research is as follows:In this thesis,a combined image classification network and semantic segmentation network is proposed.Downsampling and upsampling are the two parts of a semantic segmentation network.Downsampling can be done with the image classification network,and it is carefully designed to extract more appropriate features.The technique uses Mix Net for the downsampling part,which has multi-scale lightweight characteristics,and U-Net for the upsampling part to obtain multi-scale features and restore the original image.The MIoU accuracy of this method is about 0.03 better than that of SegNet on both the Landsat8 and BDCI2017 datasets,while the number of parameters of the network is reduced by 46.14 M.This thesis proposes a multi-flow network based ground object extraction method with dual features.With this method,a network is constructed with multiple information flow paths between the input and the final output,so that more complex features can be captured through the network.To measure the difference between the predicted result and the actual result,this method combines the cross-entropy loss function with the dice loss function.MIo U accuracy is improved by about 0.01 by this method compared to the U-Net method on both the Landsat8 and BDCI2017 datasets.In this thesis,a multi-directional network-based method for extracting ground objects is proposed that is based on the multi-headed self-attention module.By introducing an attention mechanism into the sampling network structure,this method improves the accuracy of the extraction of typical ground objects by catching the multi-directional position information in the image and helping the network to better recognize multidirectional position information in the image.MIo U accuracy is improved by about0.02 using this method in comparison to ResUNet on both Landsat8 and BDCI2017.