Abstract: Using deep convolutional neural network (CNN) to intelligently extract buildings from remote sensing images is of great significance for digital city construction, disaster detection and land management. The color difference between multi-temporal remote sensing images will lead to the decrease of generalization ability of building semantic segmentation model. In view of this, this paper proposes the attention-guided color consistency adversarial network (ACGAN). The algorithm takes the reference color style images and the images to be corrected in the same area and different phases as the training set and adopts the consistency adversarial network with the U-shaped attention mechanism to train the color consistency model. In the prediction stage, this model converts the hue of the images to that of the reference color style image, which is based on the reasoning ability of the deep learning model, instead of the corresponding reference color style image. This model transforms the hue of the images to be corrected into that of the reference color style images. This stage is based on the reasoning ability of the deep learning model, and the corresponding reference color style image is no longer needed. In order to verify the effectiveness of the algorithm, firstly, we compare the algorithm of this paper with the traditional image processing algorithm and other consistency adversarial network. The results show that the images after ACGAN color consistency processing are more similar to that of the reference color style images. Secondly, we carried out the building semantic segmentation experiment on the images processed by the above different color consistency algorithms, which proved that the method in this paper is more conducive to the impro-vement of the generalization ability of multi-temporal remote sensing image semantic segmentation model.

Key words: multi-temporal remote sensing imagery, color consistency, generative adversarial networks, semantic segmentation, attention mechanism