EGANet: Elevation-guided attention network for scene classification in panchromatic remote sensing images

Scene classification in panchromatic (PAN) remote sensing images is a challenging task due to arbitrary spatial arrangement of a variety of objects with complex background in the absence of RGB-channel information. In this paper, we propose an elevation-guided attention network (EGANet) for multimodal scene classification in panchromatic images by leveraging elevation information from digital elevation model (DEM). The proposed network helps to identify the potential regions containing prominent class-specific features in the panchromatic image scene with the attention of elevation features extracted from a convolution neural network (CNN). Then, elevation-guided features in panchromatic image scene are obtained by the correlation of these two modalities for effective scene classification. The efficacy of the proposed method is demonstrated on Cartosat-1 panchromatic remote sensing image datasets with a lot of variations in view-angle, occlusion, background, and illumination conditions. The experimental results show that our proposed EGANet achieves scene classification accuracy with an improvement of 5% in comparison with the state-of-the-art approaches.