High Resolution Remote Sensing Imagery Land-cover Classification With SDFCN Model

In recent years,with the development of satellite and unmanned aerial vehicle(UAV)technology,remote sensing images have become the main data sources for earth observation and analysis on a global and regional scale.Interpretation of high-resolution remote sensing(HRRS)images,which contain a large amount of spectrum,texture,and context information,has become the mainstream method for intelligent analysis of spatial geographic information data as well as land cover classification.At present,in the actual production process of land cover classification mapping,the manual interpretation based on remote sensing images is still widely used,which is time-consuming,labor-intensive and inefficient.Therefore,in view of the characteristics of land cover classification,it is of great theoretical significance and practical value to study automatic classification methods of HRRS images.Semantic segmentation of remote sensing images is a basic method for automatic classification of land cover classification,and also a core issue of intelligent interpretation of remote sensing images.At present,the deep learning based full convolution networks(FCN)model has great application potential in many datasets than the traditional pixel-based and object-based classification method.However,most of the existing researches on remote sensing semantic segmentation directly utilize the models and methods of natural scene image semantic segmentation.Most of them do not improve the model structure,training and post-processing methods for the particularity of remote sensing images.In addition,the training of FCN models requires large-scale datasets.The semantic segmentation dataset for remote sensing images,especially the manual annotation data corresponding to the image is still relatively lacking,which limits the popularization and application of the deep network model in the production practice.Based on the above analysis,a model based on deep Symmetrical Dense-hybrid Fully Convolutional Network(SDFCN)is proposed.On this basis,an automatic land cover classification framework for HRRS images and a deep transfer learning method for small samples in the target domain are designed.The main contents and contributions of this thesis are as follows:1.This thesis sorts out the problems of the current mainstream FCN models when processing remote sensing images.Based on the guiding principles of expanding the receptive field and reducing the number of parameters,a model based on deep Symmetrical Dense-hybrid Fully Convolutional Network(SDFCN)is proposed.In the SDFCN model,a hybrid block that enhances the receptive field range and sparseness of the model,a weight calibration structure based on spatial and channel fusion squeeze and excitation(SCFSE),a symmetric ”encoder-decoder” structure,a shortcut connection across the ”encoder-decoder” structure,and a multi-scale output training method are adopted.Experiments on four public datasets and two automatically constructed datasets show that the SDFCN model greatly improves the generalization ability and prediction accuracy,and has the practical feasibility of being applied to land cover classification tasks.2.Taking the national geographical conditions land cover classification task as an example,this thesis puts forward a SDFCN model based land cover classification framework,including the method of dataset automatic construction,model prediction and post-processing method,which greatly improves the automation and intelligence of HRRS image land cover classification task and reduces labor costs.On the one hand,the framework can make full use of the large-scale remote sensing images and vector archive data.On the other hand,the fusion method of interpretation results based on mask-weighted voting method in the framework combines the strategies of overlapping prediction,rotation prediction and mask weighting policies,which can effectively alleviate the grid effect caused by the patch-based prediction,and smooth the contours among objects in the final classification result.3.A method based on migrating the minimization of central moment discrepancy(CMD)into the transfer learning of SDFCN model is first proposed in this thesis,which realizes the marginal distribution adaptation of the SDFCN model.On this basis,the method of fine-tuning transfer learning method based on the minimization of CMD is put forward,so that the model can fully learn the supervision knowledge in the source domain and the unsupervised prior knowledge of a large number of unlabelled remote sensing images in the target domain.Under the situation of ”relatively surplus images and relatively insufficient labelled data”,this method could improve the transfer learning ability and efficiency in the case of small samples in the target domain,and reduce the requirements of labelled data in the target domain for transfer learning based remote sensing semantic segmentation.In summary,this thesis focuses on the FCN model and deep transfer learning theory to carry out research on the semantic segmentation of remote sensing images,especially the classification model,framework and transfer learning method of land cover mapping.Accuracy evaluation and result analysis of a large number of comparative experiments have verified the effectiveness and reliability of the proposed methods.