Deep Learning For Remote Sensing Image Processing

The field of Deep learning attracts great attention from the researchers of various fields such as Artificial Intelligent,image processing,text mining,computer vision,and language processing and also from other traditional sciences such as physics,biology,and manufacturing.Deep Neural Networks such as convolutional neural networks,sequence processing models such as recurrent neural networks,and regularization implements such as dropout,are used widely.Nevertheless,fields such as physics,Aerospace,medical science and biology,and manufacturing are ones that their current models failed to overcome their demands so they are trying to move toward the use of Convolution Neural Network and Recurrent Neural Network,new needs arise from deep learning.In this thesis three new deep learning models have been proposed to the field of Remote Sensing Image processing.In the first part of this thesis a new deep network structure for pixel-wise sea-land segmentation in complex and high-density remote sensing images have been proposed.Deep neural architectures achieved the best performance in image segmentation.So far,a few approaches used Convolutional Network for sea-land segmentation and the overall results still need to be improved.Here we presented a robust and accurate deep convolution network,called Extended U-Net(EU-Net),for sea-land pixel-wise segmentation in remote sensing images.Similar to the U-Net,EU-Net has down and up sampling paths to extract the best resolution result.However,in EU-Net,each down-and up-sampling layer,in addition to the convolution layers,has the blocks of advanced dense networks which contains multilevel convolution layers and combined connection layers.These proposed blocks have proper stepwise learning capability resulting in more accurate segmentation results.In the second part of the thesis a deep learning method for the resolution recovery have been proposed.The low-resolution objects and points in the remote sensing and surveillance records are up sampled using a deep Convolutional Neural Network and convex optimization.To avoid problems of image boundary the data padded with zeros.Dissimilar to the outdated methods which operate components individually,our model performs combined optimization for all the layers.The proposed model has a lightweight structure and minimal data pre-processing and computation cost.In the third part of the thesis we presented a novel network pipeline called CNNi N(Convolutional neural network in network)which is deeper than existing models by using appropriate layers.Then class-level sparse representation is embedded to the proposed network to map the high dimensional features space into a subspace with low dimensional discriminatory features space.The obtained spatial and spectral feature are then combined together to form a joint spatial-spectral feature map.Finally recurrent neural network with LSTM recurrent units is trained on this joint feature map which representing rich spectral and spatial property,to classify each pixel vector.The proposed model takes advantage of enhanced feature extraction from CNNs,RNN and SR.