Abstract: Traditional semantic segmentation of remote sensing image is to classify the pixels with similar values by using the spectral characteristics of images, but it is unable to distinguish the same kind of objects with different spectra. Aiming at this problem, a method was proposed in which the color difference information of neighborhood is integrated into the original image as input to PSPNet. Firstly, RGB was transformed into LAB. Then CIELAB formula was used to calculate the color difference value between each pixel and eight neighboring pixels, and the average value was taken as the neighborhood color difference value of the pixel. Experiment was done by using PSPNet on WHU building dataset and Massachusetts building dataset. The results show that the MIoU, ACC and F1score with neighborhood color difference are better than without. Therefore, the proposed method of merging neighborhood color difference is an effective way to improve the segmentation accuracy of PSPNet.

Key words: remote sensing image, semantic segmentation, deep learning, neighborhood color difference, convolutional neural network