Provide an Optimal Deep-network Method for Spectral-spatial Classifying of High Resolution Images

Ever-increasing growth and development of urbanization and rapid land-based changes have increased necessity of continuous checking of these changes for urban and environmental planning. Classification of remote sensing high resolution images can be the most effective way to achieve this goal. The classification of these images has always been challenged due to similarities between different classes and differences through one class. Dense classification, also known as semantic segmentation, is also one of the open issues in remote sensing domain. The existence of these kinds of challenges reminds the need for precise methods for classifying images. Deep learning, because of ability to extract deep and powerful features and compatible potential with images, has been known as a good choice in this domain. In this article, in order to cope with the challenges, a convolutional neural networks method based on deep learning is presented for classifying images. The reason for this choice is using deep and comprehensive features by the mentioned method. These features are captured in a supervised manner. In deep learning methods, on the other hand, there is an underlying need for training data and Because of restriction of data in remote sensing, it has been tried to ensure that the number of training samples used in the project is adequate. In this paper, the underlying goal is determination of CNN structure based on deep networks for effective classifying of aerial imagery with high spatial resolution. For this purpose, First, a deep network is designed to extract the deep and optimal features of the aerial image. Architecture and configuration of the deep network are defined in this step. Then, to evaluate the impact of different dimensions of neighborhoods on producing optimal deep features, feature extraction in image patches with different dimensions has been investigated. These patches have been used for train network. After training network with Patches in different sizes, Finally, in order to investigate the classification ability of the deep learning method, in a different approach, a support vector machine has been used for classification based on the deep features produced by the CNN. Comparison of the classification results shows almost same results in the deep learning method in comparison with the conventional support vector machine model, in the same conditions to using deep features. To evaluate the method, aerial data with a spatial resolution of one meter in Des moines area in USA and other data from Royan district in Mazandaran province have been used. Finally, the results of the evaluations show improvement in all three criteria including precision, recall and f1-score in the condition of using larger patches. Also, in general, using of deep learning methods as feature extractor and classifying these deep features by the support vector machine has a bit better evaluation results than feature extraction and classification by CNN.