|
Nowadays, semantic segmentation is commonly used to solve various remote sensing image problems. Furthermore, it has emerged as a critical technology for remote sensing image applications. In recent years, with the widespread application of deep learning in computer vision, deep learning semantic segmentation has been appreciated by more remote sensing researchers. Its performance depends on the learning of remote sensing data by the model. However, the complexity and multiplicity of remote sensing data lead to the increased difficulty of learning remote sensing data, and also reduce the ability of the model to segment high-resolution remote sensing images. To improve the model performance for deep learning semantic segmentation, many improvement methods have been proposed, such as batch regularization (BN), data augmentation, and residual networks. However, most of the methods are optimization strategies for deep learning with natural domain images. Further exploration of the effectiveness of related methods in specific application areas is necessary. In high-resolution remote sensing images, the correlation between segmentation model hyperparameter settings and residual networks has rarely been investigated, which makes the residual methods inadequately applied to the semantic segmentation of high-resolution remote sensing images. In this paper, an optimal residual semantic segmentation network based on the U-Net architecture is proposed. The ORC-UNet not only considers the loss function configuration of the model, but also the initial filter number setting of the model. With these two hyperparameter configurations, the residual method can effectively improve the segmentation performance of the model. In order to confirm the effectiveness of ORC-UNet, comprehensive experiments are conducted on the Potsdam dataset in this paper. The experimental results show that ORC-UNet achieves the best overall segmentation performance with an F1-score value of 74.44%, and it also achieves the best performance in three categories of the dataset (Background, Building, and Roads) with F1-score values of 30.78%, 89.58%, and 86.05%, respectively. Meanwhile, ORC-UNet is also close to the best results for the remaining categories. In addition, the relevant comparison results also indicate that the cross-entropy loss function and initialization value 64 of ORC-UNet are the best configurations. Meanwhile, ORC-UNet is the best segmentation model among all residual models.
|
