Abstract: Manual investigation and maintenance of photovoltaic panels result in expensive human maintenance costs.Therefore,using deep learning methods to extract photovoltaic panels from remote sensing images can provide important data support for the operation and maintenance of photovoltaic power generation scenarios at a low cost. The improved DeepLabV3+ semantic segmentation model has been used to solve the problem of precise segmentation and extraction of photovoltaics using high-resolution remote sensing images. It proposes a method for extracting photovoltaic panels from ultra-high resolution remote sensing images based on DeepLabV3+ deep learning architecture. The main innovative work is reflected in:(1) a hybrid cavity convolutional spatial pyramid pooling module is proposed to address the problem of difficulty in accurately extracting photovoltaic panel information from remote sensing images; (2) to address the issue of edge details being easily lost in photovoltaic panel information extraction, attention mechanism is introduced to sensitively capture small domain features to improve the segmentation ability of the model. This research uses the multi-resolution photovoltaic data set published by the University of the Chinese Academy of Sciences in 2021 to conduct experiments. The results show that the proposed model can achieve 92.54%, 79.91% and 76.27% of the distributed photovoltaic data set IoU with 0.1?m, 0.3?m and 0.8?m spatial resolution. Compared to the original DeepLabV3+ model, the ground photovoltaic dataset with spatial resolutions of 0.3?m and 0.8?m can achieve 94.27% and 87.24% IoU improvement of 0.13~2.02?percentage points in three different resolutions and backgrounds. Compared with the classical semantic segmentation models U-Net, PSPNet, and DeepLabV3+, the proposed method improves IoU by 0.64~20.51?percentage points on rooftop distributed photovoltaic datasets with spatial resolutions of 0.1 m, 0.3 m, and 0.8 m. The above experiments demonstrate the effectiveness of this method.

Key words: high-resolution remote sensing, PV panel identification and extraction, semantic segmentation, mixed void convolution, attention mechanism

摘要： 人工调查维护光伏板导致人力维护成本昂贵。因此，使用深度学习方法通过遥感影像提取光伏板能够低成本为光伏发电场景的运行维护提供重要数据支撑。使用改进后的DeepLabV3+语义分割模型，解决了如何利用高分辨率遥感影像进行光伏的精确分割与提取的问题。提出一种基于DeepLabV3+深度学习架构的超高分辨率遥感影像光伏板提取方法。主要创新工作体现在：（1）针对遥感影像中光伏板信息难以精细提取的问题，提出混合空洞卷积空间金字塔池化模块；（2）针对光伏板信息提取中边缘细节易丢失的问题，引入注意力机制敏感捕捉小区域特征，以提高模型的分割能力。该研究采用2021年中国科学院大学发布的多分辨率光伏数据集进行实验，结果表明：提出的模型在0.1?m、0.3?m、0.8?m空间分辨率的分布式光伏数据集中IoU可达92.54%、79.91%、76.27%。在0.3?m、0.8?m空间分辨率的地面光伏数据集中可达到94.27%、87.24%，相较于原本的DeepLabV3+模型，在三种不同分辨率和不同背景的场景中的IoU提高0.13~2.02个百分点；同时在0.1?m、0.3?m、0.8?m空间分辨率的屋顶分布式光伏数据集上，提出的方法与经典语义分割模型U-Net、PSPNet、DeepLabV3+相比，IoU提高0.64~20.51个百分点。以上实验证明了该方法的有效性。

关键词: 高分辨率遥感, 光伏板识别与提取, 语义分割, 混合空洞卷积, 注意力机制