【摘要】 针对遥感图像中光伏组件的分割和提取问题，提出一种基于深度残差注意力网络的遥感图像光伏组件语义分割方法。首先基于U-Net结构搭建光伏组件遥感图像分割框架；然后，使用深度残差神经网络替换原始U-Net的特征提取部分，提升网络的图像特征提取和表达能力；最后，在网络的残差模块中引入一种高效局部注意力机制，用于进一步增强局部特征的表达能力，提高算法对光伏组件的分割和提取精度。利用该算法在遥感图像光伏组件公开数据集上进行分割提取实验，结果表明改进算法在3类不同空间分辨率的数据集上表现优于DeepLabv3+、UCTransNet、UDTransNet、HRNetV2、SegFormer等方法，相较于原始U-Net网络的mIoU、Dice、mPA和Precision等评价指标平均提升5.80%、2.91%、3.06%和3.92%。更多还原

【Abstract】 To address the challenge of segmenting photovoltaic(PV) modules in remote sensing images, this paper proposes a semantic segmentation method based on a deep residual attention network. The method builds on the U-Net architecture by integrating a deep residual network for enhanced feature extraction and representation. Additionally, a local attention mechanism is incorporated within the residual modules to further refine local feature expression, improving segmentation accuracy. Experimental results on a public remote sensing PV dataset demonstrate that the proposed method consistently outperforms DeepLabv3+, UCTransNet, and UDTransNet across multiple spatial resolutions, achieving average improvements of 5.80%, 2.91%, 3.06% and 3.92% in mIoU, Dice, mPA, and Precision, respectively, over the original U-Net.更多还原