基于RDU-Net深度学习模型的电力基础设施提取方法研究

Research on Power Infrastructure Extraction Method Based on RDU-Net Deep Learning Model

摘要: 随着电力基础设施的逐步完善，如何减少电力损耗及提高发电量变成当前最迫切的需求，这就需要有效的对设施进行管理，对设施的位置、大小、面积进行识别统计是建立管理系统的前提。传统的识别统计方法如人工实地调查、遥感影像解译等都需要花费大量的人力物力和财力。本文根据地物特性，使用了RDU-Net网络模型，该模型在U-Net网络模型上进行了算法优化和完善，通过添加空洞卷积以提升模型感受野，同时引入Tversky损失函数自动平衡正负样本，最终得到了更适用于高分遥感图像分割的模型。本文实验结果表明：本文使用的RDU-Net模型能够很好的克服草木遮挡的干扰，精度较U-Net模型得到了很好得提升。可为识别典型电力基础设施相关方面研究提供思路，并能有效提高国家对基础能源设施的管理效率以及起到指导规划建设的作用。

Abstract: With the gradual improvement of power infrastructure, how to reduce power loss and improve power generation has become the most urgent demand at present, which requires effective man-agement of facilities. The identification and statistics of the location, size and area of facilities is the premise of establishing a management system. Traditional identification statistical methods, such as manual field investigation and remote sensing image interpretation, require a lot of manpower, material resources and financial resources. In this paper, according to the characteristics of ground objects, RDU-Net network model is used. This model performs algorithm optimization and improvement on the U-Net network model, improves the model’s receptive field by adding cavity convolution, and introduces Tversky loss function to automatically balance the positive and negative samples. Finally, a model more suitable for high-resolution remote sensing image seg-mentation is obtained. The experimental results show that the RDU-Net model used in this paper can overcome the interference of vegetation occlusion very well, and its accuracy is better than that of U-Net model. It can provide ideas for the identification of typical power infrastructure related research, and can effectively improve the national management efficiency of basic energy facilities and play a guiding role in planning and construction.