Remote sensing scene classification using an attention consistent network

Image scene classification in the remotely sensed (RS) society is an interesting subject that aims to allocate land use / cover semantic information. A big amount of Convolutional neural classification models for RS images have been proposed by the authors due to the massive behaviour of CNNs in features extracted. Despite their impressive results, there are still opportunities for advancement. To begin, local characteristics are just as important as global ones in identifying RS images. The CNNs' hierarchical organizational structure and multidimensional suitable capabilities make them good at trying to capture spatial information [1]. It's not uncommon for the feature maps to be overlooked, moreover. First and foremost, the ranges among RS image pairs must be minimised or maximised in order to achieve satisfying classifier performance. Despite this, the importance of these focuses in classification tasks is undervalued. We propose a new CNN called focus high quality infrastructure (ACNet) depending on the Siamese network in the this paper to improve the robustness discussed previous section.  [3]To begin, the input data for ACNet are the picture groups achieved through spatial rotation as a result of the network's the double framework.&nbsp