基于深度学习的遥感影像光伏发电板提取

Remote sensing image-based photovoltaic panel extraction via deep learning

以光伏产业为代表的太阳能资源是未来最安全和清洁的可再生能源,利用高分辨率遥感影像提取和监测光伏发电板用地分布状况对于光伏产业发展和行业监管具有重要意义。文章基于2022年鄂尔多斯高分辨率遥感影像制作光伏发电板提取样本集,构建全卷积神经网络(FCN)、金字塔场景解析网络(PSPNet)、DeepLabV3+、U-Net、U-Net++五种常用的语义分割模型,对光伏发电板进行语义分割。通过对比不同模型的提取精度和视觉效果,为光伏发电板提取工作提供自动化识别模型。结果表明,U-Net++模型精度最高,在测试集中像素准确率为96.0%,F1为95.0%,交并比达到90.4%。

Solar energy,represented by the photovoltaic industry,is considered one of the safest and cleanest renewable energy sources for the future.The utilization of high-resolution remote sensing images for extracting and monitoring the land distribution of photovoltaic panels holds immense significance for the development and supervision of the photovoltaic industry.This paper utilized high-resolution remote sensing images from Ordos in 2022 to create a sample set for photovoltaic panel extraction.Five commonly used semantic segmentation models were constructed,namely fully convolutional network(FCN),pyramid scene parsing network(PSPNet),DeepLabV3+,U-Net,and U-Net++,so as to perform semantic segmentation on photovoltaic panels.By comparing extraction accuracy and visual effects of different models,an automatic identification model was set for photovoltaic panel extraction.The results demonstrate that the U-(Net++model achieves the highest accuracy,with a pixel accuracy of 96%,an F1 of 95%,and an intersection-over-union IoU)of 90.4%in the test set.