Research of Land Surface Segmentation based on Convolutional Netword and Transformer Network

Throughout the decades, various applications have utilized semantic segmentation of spatial sensing imagery. As computer vision has benefited from deep learning methods, researchers have focused their efforts on transferring their superior performance to remote sensing image analysis. To achieve semantic segmentation of ultra-high resolution images, almost all neural networks have been using different methods for multi-scale fusion of image features in recent years. Although there have been numerous studies comparing the overall performance of these networks, there seems to be no study of the performance of these networks on a single class of datasets, especially for the Transformer networks that have emerged in recent years. In this paper, five networks using different multiscale fusion approaches were selected for experiments on two different datasets and similar results were obtained. We found that Transformer can achieve good results in the case where segmented objects occupy a large proportion of the image due to its focus on global image changes. However, Transformer also has some limitations, especially the image hue has a great influence on it. Our experiments found that if the color of the roof in the segmented image is similar to the color of the water surface, Transformer has a high probability to incorrectly confuse the roof as the water surface, while the traditional Convolutional Networks will work much better in this case.