Research On Recognition Of Urban Objects Based On Multi-source Remote Sensing Data

Recently,with the rapid development of the economy,the urban spatial structure has been continuously optimized.It is a great significance to grasp the spatial structure of the whole city and formulate scientific and rational planning by identifying urban objects and their spatial distribution characteristics.Previous urban object recognition techniques were usually based on static data or dynamic data.However,subject to subjective factors,it is difficult to obtain better recognition results.The rich spatial and spectral information captured in remote sensing images can be used to identify the urban objects,especially the functional partitions and fine objects.Image recognition based on deep learning is one of the most popular research techniques adopted in the field of computer vision.The deep learning model can be utilized to learn the discriminative feature representation of remote sensing images,and realize the intelligent recognition of urban objects.The strong discriminative data features help to better mine the potential information in the original image data and further improve the performance of the image recognition model.There-fore,it becomes an urgent matter that how to extract effective visual representation from the high-dimensional remote sensing image based on deep learning model and employ it to perform the urban functional partitions fast recognition with high accuracy.Intelligent recognition of urban objects has important implications for helping peo-ple discover valuable information hidden in the city.This dissertation is based on the deep learning models,which are utilized to extract meaningful and discriminative data features from multi-source remote sensing data(such as aerospace and aerial remote sensing da-ta),and transform low-level remote sensing features into high-level semantic features;It's used to analyze the potential information hidden in the original data,and realize the intelligent recognition of urban objects.Thus,it provides technical support for building smart cities.Specifically,the main research contents and innovations completed in this dissertation are as follows:(1)In order to solve the problem of traditional urban target recognition methods that cannot identify urban targets quickly,efficiently and intelligently,a novel hyperspectral remote sensing images classification model is proposed,namely R-3D-CNN,which com-bines the 3D-CNN model and the RNN architecture.The R-3D-CNN model makes full use of the spatial and spectral fusion features,and takes more consideration of the spa-tial context information of the pixel.With the 3D convolution operation,the accuracy of hyperspectral images classification can be improved by combining spectral features and spatial features.Finally,the proposed models are trained on the many published hyperspectral datasets.And the experimental results prove that the proposed supervised classification deep learning models outperform the traditional machine learning methods and some deep learning models(such as 2D-CNN and 3D-CNN).(2)In order to solve the problem of weak generalization ability of deep learning models for hyperspectral remote sensing classification,a novel Synergistic Convolutional Neural Network model,namely SyCNN that consists of a hybrid module of 2D/3D CNNs,is proposed for the classification of aerospace hyperspectral remote sensing images.It is utilized for hyperspectral image classification by mixing 2D-CNN and 3D-CNN models and generating the new 2D and 3D features with interacting 2D features and 3D features.Due to the data interaction module,there is more training samples for the next 3D-CNN and the spectral information is fused in the 2D features.The proposed SyCNN model not only achieves better classification results on small sample datasets,but also has better generalization capabilities.Finally,the proposed SyCNN models are experimented on three hyperspectral remote sensing datasets.The experimental results demonstrate that the proposed SyCNN models outperform the current deep learning models based on 2D-CNN or 3D-CNN.And the proposed model has better generalization capabilities,and could get better results on a variety of small sample datasets.(3)In order to solve the problems of occlusion,shadow and noise in urban road identification,a novel end-to-end recurrent convolutional neural network U-Net model,namely RCNN-UNet,is proposed for aerial remote sensing images feature extraction.This proposed RCNN-UNet model is mainly used to solve the problem of the occlusion,shadow and noises in aerial remote sensing images,and improve the accuracy of urban road recognition.It could not only make full use of the spatial context information,but also combine the advantages of the "U" architecture to make full use of the low-level features.Finally,extensive experiments were carried out based on two publicly avail-able benchmark datasets,and nine state-of-the-art baselines(deep learning models)were used in a comparative evaluation.The experimental results demonstrate that the proposed RCNN-UNet model outperforms other comparison methods both on the road detection and the centerline extraction tasks.(4)In order to solve the problem of less information and noise available in low-resolution aerial remote sensing data for urban target recognition,a novel end-to-end recursive Dense network model,namely RDN,is proposed to restore the low-resolution aerial remote sensing images to high-resolution images.The proposed RDN module has a larger effective field,and makes full use of the spatial context information.The Dense network and local residual connection operations used in the proposed model,which takes advantage of the low-level feature map generated by each layer of convolution.It can gen-erate high-resolution images directly from the low-resolution images,and then introduce the accuracy of urban object recognition.Finally,extensive experiments were carried out based on four publicly available benchmark datasets for image super-resolution,and one experiment was carried out based on one low-resolution dataset(PatternNet dataset).The experimental results show that the proposed RDN model achieves better performance on the four public datasets,and the accuracy of classification based on the RDN model is significantly better than that based on low-resolution image classification.In summary,a series of deep learning models based on remote sensing images are proposed,and successfully applied to a variety of urban target recognition.The validity of the proposed model is fully verified by a large number of experiments based on multi-source remote sensing datasets.