Remote sensing image scene classification based on a dual attention dense network

Abstract

Remote sensing image classification has experienced three stages: pixel-level, object-level and scene-level. With the improvement of remote sensing image resolution, pixel-level and object-level methods cannot be completely correctly classified, and thus, scene classification is the current focus of this research. We consider the complex background of remote sensing images, the existence of many small objects and the large scale of change, as well as intraclass diversity and interclass similarity. Through the salient regions and features in remote sensing images, a dual attention dense network is proposed. In addition, an adaptive spatial attention module and an adaptive channel attention module are designed. Specifically, the network combines the output of the two proposed attention modules as the feature representation. Among them, the adaptive parameter activation function is introduced into the adaptive spatial attention module, and different nonlinear transformations are performed on the input features in the spatial attention network to achieve attention on important regions. By capturing the adaptive cross channel interaction range to learn channel attention, important weights of each channel are generated and an adaptive parameter activation function is introduced to adjust the feature values of different channels, thereby acting with the global features to achieve attention on the salient features. We present extensive experiments on three scene classification datasets, including the UCM dataset, the AID dataset and the OPTIMAL dataset, and compare them with various algorithms. The experimental results demonstrate the effectiveness of our proposed dual attention model.