遥感图像语义分割，就是对遥感图像进行逐像素的分类预测，分割出不同的地物区域，可以同时实现遥感地物分类和目标分割，是遥感领域十分重要且具有挑战性的任务，可广泛应用于城市规划、地理数据库更新、灾害监测、气候研究等场景中，有着非常重要的研究意义和实际应用价值。然而，随着遥感图像分辨率不断提高，遥感地物信息更丰富、目标结构更复杂、同类地物的类内特征差异大、不同地物的类间特征相似等特点，导致准确的遥感图像语义分割面临挑战，分割结果也存在许多虚假预测的区域，缺少完整性，同时如何获得明确、清晰的分割边界也是亟待解决的问题。

近年来，随着深度神经网络的发展，全卷积神经网络方法替代传统的分割方法被广泛应用于遥感图像语义分割领域，也取得了比较好的结果。但是，当前的全卷积神经网络在池化操作和上采样过程中会丢失图像中的地物边界、位置等特征信息，导致模型分割的语义区域边界模糊、有歧义；此外，由于卷积操作只针对固定感受野内的像素，没有显式利用像素间的相关关系，容易导致同一语义区域内分割结果不一致，不同语义区域的分割结果相混淆，产生细碎的错分区域，因此对现有的全卷积神经网络进行改进，提升分割效果，仍具有很重要的研究意义。

因此，本文选择带有特征融合操作的U-Net全卷积网络结构，并对其进行改进，提升分割的效果，文中首次将自注意力机制引入遥感图像语义分割领域，来融合图像的上下文语义信息，然后采用新的上采样方法以减少边界信息的损失，最后采用集成学习方式优化最终的分割结果。综上，本论文研究和讨论了基于改进U-Net全卷积神经网络的遥感图像语义分割方法。本文的具体研究工作如下：

（1）  考虑遥感图像内上下文信息的影响，改进原始的U-Net网络，引入自注意力机制，提出并实现融合自注意力机制的U-Net网络SA-UNet（Self Attention U Networks）模型，针对每个像素点生成不同的权重值，实现对特征图的加权处理，能够有效地区分目标和背景，减少同一地物区域内的错分情况，使得分割结果中的各地物区域更完整。

（2）  考虑每一个像素预测之间的关联，以新的数据相关型上采样（DUpsampling）方法替代双线性插值法，提出DU-Net（Data dependent Upsampling U Networks）模型，减少特征损失，在U-Net网络中，可以提高特征融合的效果，能够更准确地恢复像素预测，得到更准确、清晰的分割边界。

（3）  将自注意力机制和新型上采样方法同时引入U-Net网络，提出并实现SA-DUNet（Self Attention DUpsampling U Networks）模型。实验结果表明新的遥感图像语义分割方法分割精度提升明显，分割得到的地物区域边界更清晰，而且同一地物区域内的错分区域大大减少，更具完整性。

（4）  利用SA-DUNet网络和DeepLabV3+网络集成的优化策略对遥感图像语义分割结果进行优化，进一步提升分割效果。

Remote sensing image semantic segmentation is to perform pixel-by-pixel classification and prediction on remote sensing images to segment out different feature areas, which can achieve remote sensing feature classification and target segmentation at the same time. It is a very important and challenging task in the field of remote sensing, can be widely used in urban planning, geodatabase update, disaster monitoring, climate research and other scenarios, which has very important practical value and research significance. However, with the continuous improvement of the resolution of remote sensing images, the features information of remote sensing images is richer, the target structure is more complex, the intra-class features of similar remote sensing features are greatly different, and the features of different remote sensing features are similar. As a result, accurate remote sensing image semantic segmentation faces challenges, and there are many falsely predicted regions in the segmentation results, which lack completeness. At the same time, how to obtain clear segmentation boundaries is also an urgent issue.

In recent years, with the development of deep learning, the fully convolutional neural network methods have been widely used in the field of remote sensing image semantic segmentation instead of traditional segmentation methods, and have achieved relatively good results. However, the current full convolutional neural networks will lose feature information such as remote sensing boundaries and positions in the image during the pooling operation and upsampling process, resulting in fuzzy and ambiguous semantic region boundaries of the model segmentation. In addition, due to the convolution operation only for pixels in the fixed receptive field, without explicitly using the correlation between the pixels, it is easy to cause inconsistent segmentation results in the same semantic region and confused segmentation results in different semantic regions, and generate fine mismatched regions. So, it is still of great research significance to revise the existing fully convolutional neural network to improve the segmentation effect.

Therefore, we choose the U-Net fully convolutional network structure with feature fusion operation and improves it to improve the segmentation effect. In this article, the self-attention mechanism is introduced into the field of remote sensing image semantic segmentation for the first time to fuse the contextual semantic information of the image, and then a new upsampling method is used to reduce the loss of boundary information. Finally, the final segmentation result is optimized by integrated learning. In summary, we study and discuss the semantic segmentation method of remote sensing images based on improved U-Net fully convolutional neural network. Our specific research work is as follows:

(1)      Considering the influence of context information in remote sensing images, we introduce the self-attention mechanism into the original U-Net network, propose and implement the self-attention U Network (SA-UNet) model that integrates the self-attention mechanism. The model generates different weight values for each pixel and implements weighted processing of feature maps, which can effectively distinguish the target and the background, reduce the misclassification in the same remote sensing feature area, and make all remote sensing feature areas in the segmentation result more complete.

(2)      Considering the correlation between each pixel prediction, a new data-dependent upsampling (DUpsampling) method is used instead of the bilinear interpolation method. Then the DU-Net model is proposed to reduce the loss of features. In the U network, the effect of feature aggregation can be improved, the pixel prediction can be restored more accurately, and a more accurate and clear segmentation boundary can be obtained.

(3)      The self-attention mechanism and the new upsampling method are introduced into the U-Net network at the same time, and the SA-DUNet (Self Attention DUpsampling U Networks) model is proposed and implemented. The experimental results show that the new remote sensing image semantic segmentation method has obvious improvement in segmentation accuracy, the boundary of the feature areas obtained by segmentation is clearer, and the mis-segment region in the same remote sensing feature area is greatly reduced.

(4)      We use the optimization strategy of SA-DUNet network and DeepLabV3+ network integration to optimize the semantic segmentation results of remote sensing images and further improve the segmentation effect.