U-ONet: Remote sensing image semantic labelling based on octave convolution and coordination attention in U-shape deep neural network

Abstract

Semantic labelling of remote sensing images is crucial for various remote sensing applications. However, the dense distribution of man-made and natural objects with similar colours and geographic proximity poses challenges for achieving consistent and accurate labelling results. To address this issue, a novel deep learning model incorporating an octave convolutional neural networks (CNNs) within an end-to-end U-shaped architecture is presented. The approach differs from conventional CNNs in that it employs octave convolutions instead of standard convolutions. This strategy serves to minimize low-frequency information redundancy while maintaining segmentation accuracy. Furthermore, coordination attention is introduced in the encoder module to enhance the network's ability to extract useful features, focusing on spatial and channel dependencies within the feature maps. This attention mechanism enables the network to better capture channel, direction, and location information. In conclusion, the U-shaped network is engineered with a completely symmetric structure that employs skip connections to merge low-resolution information, used for object class recognition, with high-resolution information to enable precise localization. This configuration ultimately improves segmentation accuracy. Experimental results on two public datasets demonstrate that our U-ONet achieves state-of-the-art performance, making it a compelling choice for remote sensing image semantic labelling applications.