Real and Synthetic Overhead Images of Wind Turbines in the US

OverviewThis dataset contains real overhead images of wind turbines in the US collected through the National Agriculture Imagery Plan (NAIP), as well as synthetic overhead images of wind turbines created to be similar to the real images. All of these images are 608x608. For more details on the methodology and data, please read the sections below, or look at our website: Locating Energy Infrastructure with Deep Learning (duke-bc-dl-for-energy-infrastructure.github.io).Real DataThe real data consists of images.zip and labels.zip. There are 1,742 images in images.zip, and for each image in this folder, there is a corresponding label with the same name, but a different extension. Some images do not have labels, meaning there are no wind turbines in those images. Many of these overhead images of wind turbines were collected from Power Plant Satellite Imagery Dataset (figshare.com) and then hand labeled. Others were collected using Google Earth Engine or EarthOnDemand and then labeled. All of the images collected are from the National Agricultural Imagery Program (NAIP), and all are 608x608 pixels. The labels are in YOLOv3 format, meaning each line in the text file corresponds with one wind turbine. Each line is formatted as: class x_center y_center width height. Since there is only one class, class is always zero, and the x, y, width, and height are relative to the size of the image and are between 0-1.The image_locations.csv file contains the latitude and longitude for each image. It also contains the image's geographic domain that we defined. Our data comes from what we defined as four regions - Northeast (NE), Eastern Midwest (EM), Northwest (NW), and Southwest (SW), and these are included in the csv file for each image. These regions are defined by groups of states, so any data in WA, ID, or MT would be in the Northwest region.Synthetic DataThe synthetic data consists of synthetic_images.zip and synthetic_labels.zip. These images and labels were automatically generated using CityEngine. Again, all images are 608x608, and the format of the labels is the same. There are 943 images total, and at least 200 images for each of the four geographic domains that we defined in the US (Northwest, Southwest, Eastern Midwest, Northeast). The generation of these images consisted of the software selecting a background image, then generating 3D models of turbines on top of that background image, and then positioning a simulated camera overhead to capture an image. The background images were collected nearby the locations of the testing images.ExperimentationOur Duke Bass Connections 2020-2021 team performed many experiments using this data to test if the synthetic imagery could help the performance of our object detection model. We designed experiments where we would have a baseline dataset of just real imagery, train and test an object detection model on it, and then add in synthetic imagery into the dataset, train the object detection model on the new dataset, and then compare it's performance with the baseline. For more information on the experiments and methodology, please visit our website here: Locating Energy Infrastructure with Deep Learning (duke-bc-dl-for-energy-infrastructure.github.io).