Improve Semantic Segmentation of High-Resolution Remote Sensing Imagery with RS-TTA

As deep learning has been widely used in many computer vision fields, semantic segmentation techniques based on convolutional neural networks are often used in remote sensing image problems. However, its segmentation performance still needs further improvement, especially for high separation rate remote sensing images. Many adjustment strategies for segmentation performance have been proposed, such as fully convolutional networks, residual connectivity, data augmentation, and so on. However, there are a few better and easy-to-operate adjustment strategies in the testing phase of segmentation networks. In this work, we propose a tuning strategy for high-resolution remote sensing images in the testing phase of segmentation networks, which we call RS-TTA. This method is used to improve the prediction performance of the segmentation network by simply rotating and flipping horizontally to enhance the prediction samples in the test phase and fusing the enhanced prediction results. This tuning method does not need to take into account the sample data during training or the training process of the network. It can improve the prediction performance of the segmentation network by performing a simple operation on the prediction samples, and it is also more compatible with the adjustment strategies and methods used in the training process. To evaluate this method, comparative experiments were conducted on the Potsdam dataset provided by the ISPRS competition. The experimental results show that the method can improve the prediction performance of the fully convolutional residual segmentation network on high-resolution remote sensing imagery. In addition, the test results of different combinations of models demonstrate that the adjustment effect of the method is well generalized for the task of semantic segmentation of remote sensing imagery.