Scene classification relying on images is essential in many systems and applications related to remote sensing. The scientific interest in scene classification from remotely collected images is increasing, and many datasets and algorithms are being developed. The introduction of convolutional neural networks (CNN) and other deep learning techniques contributed to vast improvements in the accuracy of image scene classification in such systems. To classify the scene from areal images, we used a two-stream deep architecture. We performed the first part of the classification, the feature extraction, using pre-trained CNN that extracts deep features of aerial images from different network layers: the average pooling layer or some of the previous convolutional layers. Next, we applied feature concatenation on extracted features from various neural networks, after dimensionality reduction was performed on enormous feature vectors. We experimented extensively with different CNN architectures, to get optimal results. Finally, we used the Support Vector Machine (SVM) for the classification of the concatenated features. The competitiveness of the examined technique was evaluated on two real-world datasets: UC Merced and WHU-RS. The obtained classification accuracies demonstrate that the considered method has competitive results compared to other cutting-edge techniques.

Remote Sensing (RS) image classification has recently attracted great attention for its application in different tasks, including environmental monitoring, battlefield surveillance, and geospatial object detection. The best practices for these tasks often involve transfer learning from pre-trained Convolutional Neural Networks (CNNs). A common approach in the literature is employing CNNs for feature extraction, and subsequently train classifiers exploiting such features. In this paper, we propose the adoption of transfer learning by fine-tuning pre-trained CNNs for end-to-end aerial image classification. Our approach performs feature extraction from the fine-tuned neural networks and remote sensing image classification with a Support Vector Machine (SVM) model with linear and Radial Basis Function (RBF) kernels. To tune the learning rate hyperparameter, we employ a linear decay learning rate scheduler as well as cyclical learning rates. Moreover, in order to mitigate the overfitting problem of pre-trained models, we apply label smoothing regularization. For the fine-tuning and feature extraction process, we adopt the Inception-v3 and Xception inception-based CNNs, as well the residual-based networks ResNet50 and DenseNet121. We present extensive experiments on two real-world remote sensing image datasets: AID and NWPU-RESISC45. The results show that the proposed method exhibits classification accuracy of up to 98%, outperforming other state-of-the-art methods.

Nowadays the rise of the artificial intelligence is with high speed. Even we are far away from the moment when machines are going to make decisions instead of human beings, the development in some fields of artificial intelligence is astonishing. Deep neural networks are such a filed. They are in a big expansion in a new millennium. Their application is wide: they are used in processing images, video, speech, audio, and text. In the last decade, researches put special attention and resources in the development of special kind of neural networks, convolutional neural networks. These networks have been widely applied to a variety of pattern recognition problems. Convolutional neural networks were trained on millions of images and it is difficult to outperform the accuracies that have been achieved. On the other hand, when we have a small dataset to train the network, there is no success to do it from a scratch. This article exploits the technique of transfer learning for classifying the images of small datasets. It consists fine-tuning of the pre-trained neural network. Here in details is presented the selection of hyper parameters in such networks, in order to maximize the classification accuracy. In the end, the directions have been proposed for the selection of the hyper parameters and of the pretrained network which can be suitable for transfer learning.