在遥感影像识别中地块分割是非常重要的一个环节，其识别准确率对后续的影像识别和解释会产生直接的影响。高分辨率遥感图像中的地块语义分割，不仅对于土地边界识别以及地块数据监测有一定的作用，对于促进农业现代化发展也具有重要的意义。由于遥感数据标签标注的专业性，数据中涉及的地块及背景的特征复杂多样性，加之数据量一般较为巨大，因此现行分割方法的准确度仍有局限，即使运用深度学习方法去训练模型提取地块，仍有改进的空间。

本文从语义分割任务出发，研究基于卷积神经网络的遥感地块语义分割方法及改进方式，利用卷积神经网络对于图像特征的强大学习能力对遥感图像进行地块特征提取分割，对已有模型方法进行实验优化以及改进论证，以获取分割精度更高，边界更为清晰的实验效果。

论文的主要研究内容如下：

分析并实现了三种基于卷积神经网络的典型语义分割网络结构，并且在高分辨率遥感数据集GID上构建了三种相应的模型方法，并进行了实验，分析对比了不同模型方法的模型性能、分割质量，验证了模型的有效性。实验结果表明在遥感地块语义分割中U-Net网络的分割效果最好，因此本文将其作为研究的基础网络。

由于深度卷积神经网络固有的特点，在应用于遥感地块语义分割时，存在地块语义边界模糊，整体分割效果不理想的问题。针对该问题，本文结合注意力机制对基础网络进行改进：通过嵌入不同的注意力机制，如：通道注意力机制、坐标注意力机制，优化筛选提取的特征。结合上述注意力机制，本文提出了一种考虑全局信息以及位置信息的全局坐标注意力机制，突出了目标特征，提高了分割精确度；结合U-Net网络结构考虑注意力机制的嵌入方式，比较了针对U-Net网络在编码器阶段和解码器阶段嵌入注意力机制的效果，实验结果表明，在U-Net网络的解码器阶段嵌入注意力机制，对于模型的分割效果提升更为明显；考虑网络中不同尺度数据的特征差异，引入多尺度网络模型理论，对于经过卷积后尺度改变的数据，针对性的运用不同的注意力机制，以达到降低参数，提高准确度的目标。在GID数据集上进行实验，结果表明：所提出的全局坐标注意力机制相较于同类注意力机制来说，嵌入之后的改进效果最为明显；考虑尺度问题，不同注意力机制的融合嵌入在降低参数，提高模型分割效率的同时，准确度仍可以保持较高水平；相较于同类算法，所设计的注意力机制及嵌入方式的改进，具有明显优势，在理论以及实用价值方面有一定意义。

In remote sensing image recognition, parcel segmentation is a very important link, and its recognition accuracy will have a direct impact on subsequent image recognition and interpretation.Semantic segmentation of land parcels in high-resolution remote sensing images not only plays a certain role in land boundary recognition and parcel data monitoring, but also plays an important role in promoting the development of agricultural modernization. Due to the professionalism of remote sensing data labeling, the complex and diverse characteristics of the plots and backgrounds involved in the data, and the generally huge amount of data, the accuracy of the current segmentation methods is still limited. block, there is still room for improvement.

Starting from the task of semantic segmentation, this paper studies the semantic segmentation method and improvement method of remote sensing parcels based on convolutional neural networks, and uses the powerful feature learning ability of convolutional neural networks to extract parcel features from remote sensing images. Experiment optimization and improvement demonstration to obtain experimental results with higher segmentation accuracy and clearer boundaries.

The main research contents of the paper are as follows:

Firstly, three typical semantic segmentation network structures based on convolutional neural network are analyzed and studied, and three corresponding model methods are constructed based on the public high-resolution remote sensing data set GID, and experiments are carried out to analyze and compare different models. The model performance and segmentation quality of the method verify the effectiveness of the model. The experimental results show that the U-Net network has the best segmentation effect in the semantic segmentation of remote sensing parcels, so this paper takes it as the basic network of the research.

Due to the inherent characteristics of deep convolutional neural networks, when applied to the semantic segmentation of remote sensing parcels, there is a problem that the semantic boundaries of the parcels are blurred and the overall segmentation effect is not ideal. In response to this problem, this paper improves the basic network by combining the attention mechanism, and optimizes the extracted features by embedding different attention mechanisms, such as channel attention mechanism and coordinate attention mechanism. Combined with the above attention mechanism, this paper proposes a global coordinate attention mechanism that considers global information and position information, which highlights the target features and improves the segmentation accuracy. Secondly, combined with the U-Net network structure, the attention mechanism is considered. Comparing the effect of embedding the attention mechanism in the encoder stage and the decoder stage for the U-Net network, the experimental results show that embedding the attention mechanism in the decoder stage of the U-Net network can improve the segmentation effect of the model more obviously;On this basis, considering the feature differences of data of different scales in the network, the theory of multi-scale network model is introduced, and for the data whose scale changes after convolution, different attention mechanisms are used to reduce parameters and improve accuracy. The goal. Experiments are carried out on the GID data set, and the results show that the proposed global coordinate attention mechanism has the most obvious improvement effect after embedding compared with the same attention mechanism; considering the scale problem, the fusion embedding of different attention mechanisms is reduced. parameters, while improving the efficiency of model segmentation, the accuracy can still maintain a high level; compared with similar algorithms, the designed attention mechanism and the improvement of the embedding method have obvious advantages, which have certain theoretical and practical value.