High Resolution Remote Sensing Imagery Classification Based On Deep Learning

Target classification and recognition of high-resolution remote sensing images is an important part of information extraction and processing of earth observation systems and automatic identification systems.High-resolution remote sensing images,as a spatial big data,which are widely used in fields such as emergency and mitigation,modern military,and precision agriculture,put forward higher requirements for image real-time processing efficiency,classification accuracy and automation level.With the massive increase of high-resolution remote sensing data volume,the diversity of data representation forms,and the complexity of remote sensing images scenes,artificially designed features can no longer meet the precise classification and recognition tasks of high-resolution remote sensing images.Therefore,it has a great significance to study the process and method of deep learning aiming at improving the extraction accuracy and classification efficiency of remote sensing information.To deal with these issues,several researches are conducted in this thesis as follows:First of all,a method to build datasets with small intra-class distance and big interclass distance based on support vector machine is proposed.The method solves the problem of time-consuming and laborious of acquisition of large label in the deep learning supervision classification.Then a series of data augment operations are adopted to extended data sets,and it enriches the diversity of samples to make the model more generalizable.Secondly,the SVM-FCN model to solve the problem of losing spatial information caused by fully connected layer in convolutional neural network compressing the feature map into one-dimension is proposed.SVM-FCN is modified from VGG-16,which is network with strong learning ability.Intending to improve the discriminating accuracy of the object,SVM-FCN also incorporates hierarchical features.Furthermore,a magic combining a sliding step greater than one and bilinear upsampling is adopted in predict stage,which improve classification efficiency.Finally,SegNet and DeconvNet both with Encoder-decoder structure are applied to improve poor semantic segmentation of remote sensing images in FCN-32 s and FCN-8s.The convolution network in the VGG-16 model is used as the encoding structure of SegNet and DeconvNet to automatically extract better performance features.The symmetry-based decoding structure is used to recover convolutional neural networks to extract image feature information.