Attention-augmented Domain-adaptive Semantic Segmentation Of Remote Sensing Images

With the continuous development of social economy,the human living environment is gradually being differentiated,which can be divided into urban areas and rural areas.Remote sensing image technology with high spatial resolution can help us better understand the geographical and ecological environment.High-resolution remote sensing images generated by remote sensing technology can not only be used to detect geographic location and terrain,but also be used in navigation systems,rescue and disaster relief.Therefore,the research and processing of remote sensing images have also become important.Recently,deep learning methods for semantic segmentation of remote sensing images have achieved good results,however,these methods often require a large amount of labeled data to match for reliable performance.Unlike tasks such as image classification,semantic segmentation methods using deep learning require pixel-level manual annotation.Remote sensing images themselves have the characteristics of large size and complex composition,making their annotation very laborious and expensive.The urban and rural scenes show a completely different geographical landscape.These lead to insufficient generalizability of deep learning algorithms,hindering city-level or national-level map mapping.Although existing high-spatial-resolution land cover datasets have promoted research on semantic segmentation of remote sensing images,they ignore the generalization performance of the model.This paper is based on the idea of unsupervised domain adaptation to study the generalization of semantic segmentation of remote sensing images.In order to solve the problem of cross domain remote sensing image semantic segmentation,this paper proposes an attention-augmented domain-adaptive semantic segmentation of remote sensing images algorithm based on attention-enhanced domain adaptation.The main contributions of this paper are as follows:(1)A self-training-based remote sensing image semantic segmentation algorithm is proposed.In view of the shortcomings of convolutional network model,two attention mechanism modules are introduced into the segmentation part of the model.They effectively utilize the spatial context information of the image and the inter-channel relationship,which improves the segmentation results of the model,thereby improving the effect of domain adaptation tasks.(2)In order to learn domain-invariant features,a domain classifier is integrated on the basis of the semantic segmentation model,which reduces the domain offset between the source domain and the target domain based on the idea of adversarial learning while ensuring the segmentation performance.(3)Extensive experiments are conducted to verify the effectiveness of the proposed method.The experimental results on the Love DA dataset show that the method in this paper outperforms the mainstream unsupervised domain adaptation methods,and achieves better cross-domain semantic segmentation performance both from rural images to urban images and from urban images to rural images.