Baseline Dataset

<b>Overview</b><br>This is a set of overhead images of wind turbines with corresponding YOLOv3 formatted labels for object detection. These labels contain the class, x and y coordinates and the height and width of the bounding boxes for each wind turbine in the corresponding image.<br><b>Why</b><br>Deep learning can help with the analysis of energy infrastructure. Extending this work to more types of energy infrastructure can create a pipeline for in-depth energy infrastructure analysis that could provide information for energy access decision makers to choose how to provide electricity to a non-electrified region (through grid extension, micro-grids or localized power generation). <br><br><b>Method</b>The majority of the images were taken from https://figshare.com/articles/Power_Plant_Satellite_Imagery_Dataset/5307364. These images were then hand labeled and converted into formatted labels, which are also contained in original_images_and_labels. This data was then preprocessed into smaller images with dimensions of 608x608 and their corresponding labels with the same YOLOv3 format of class, x, y, height, width. These values (except for class value) have relative values from 0-1 that are proportional to the size of the images. These smaller images and labels are what are contained in the dataset. These images have resolutions varying from 0.6-1m.<br>Additional images were collected through the NAIP imagery available on Earth OnDemand and then hand-labeled.<br>

<b>数据集概述</b><br>本数据集包含多幅风力发电机组的航拍图像，以及适配目标检测（Object Detection）任务的YOLOv3格式标注文件。每张图像对应的标注文件中，包含对应风力发电机组的类别、边界框（Bounding Box）的x、y坐标以及边界框的高与宽信息。<br><br><b>构建背景</b><br>深度学习可助力能源基础设施的分析工作。若将该研究拓展至更多类型的能源基础设施，即可构建一套能源基础设施深度分析流程，为电力接入决策人员提供信息支撑，辅助其选择向无电区域供电的方案（如电网延伸、微电网建设或本地化发电）。<br><br><b>数据采集与预处理流程</b><br>本数据集的绝大多数图像源自https://figshare.com/articles/Power_Plant_Satellite_Imagery_Dataset/5307364，随后对这些图像进行手工标注并转换为标准化标注格式，原始图像与标注文件均收录于`original_images_and_labels`目录中。随后将数据预处理为尺寸为608×608的小尺寸图像，并生成对应标注文件，标注格式仍采用YOLOv3标准，包含类别、x坐标、y坐标、边界框高与宽。除类别值外，其余坐标与尺寸参数均为0至1范围内的相对值，其数值与图像尺寸成正比例关系。本数据集即包含该预处理后的小尺寸图像与对应标注文件，原始图像的分辨率范围为0.6至1米。<br>额外的图像则通过Earth OnDemand平台上的NAIP影像采集，并同样进行手工标注。