Rural Home Annotation Dataset Mapped by Citizen Scientists in Satellite Imagery
Description
A geolocated dataset of rural home annotations made on satellite imagery from Uganda, Kenya, and Sierra Leone. This dataset was produced through a citizen science project called Power to the People (https://www.zooniverse.org/projects/alycialeonard/power-to-the-people), which mapped rural homes for electrical infrastructure planning and computer-vision-based mapping research. 578,010 home annotations were made over 179 days by over 6,000 volunteers. Bounding-box annotations have anonymized and georeferenced. These raw annotations were found to have a precision of 49% and recall of 93% compared to a researcher-generated set of gold standard annotations. Data on roof colour and shape were also collected and are provided. Metadata about the sensors used to capture the original images and the annotation process are also attached. Links to original satellite image tiles used in annotation are also provided. While this dataset was collected for electrical infrastructure planning research, it can be useful in diverse sectors, including humanitarian assistance and public health. Usage notes: In the subjects dataset, data are provided on the count of responses to questions labelled T0 and T2. Question T0 is the first question in the annotation process, where contributors indicate whether they see any homes in the image ("Yes" or "No") while T2 was asked after contributors labelled homes to confirm whether they labelled them all ("Yes" or "There were too many to label them all"). A "None" count is also provided for those who did not answer these questions. In the annotations dataset, roof colour and roof shape are encoded as follows: roof_color 0 = white or light-coloured, 1 = brown, 2 = other; roof_shape 0 = square or rectangular, 1 = circular or rounded, 2 = other. Georeferencing information is provided in the original referencing system of the annotated satellite images, which are WGS 84 in the UTM zone appropriate to the location in question (accessible in subject metadata). Publications using this data are forthcoming; DOIs will be added here once published.
Files
Steps to reproduce
Power to the People was hosted online on the Zooniverse platform. It ran from 2nd March and ran until 28th August 2020. It was open to any global contributors with an internet connection. In the project workflow, the citizen scientist was presented with a subject (e.g. three images of the same area - one multispectral, one panchromatic, and one pansharpened). They were asked whether they could see any homes. If they did not, they were presented with a new subject. If they did, they were asked to annotate them. For each annotation, they were asked to identify the roof color from options "white or light-coloured", "brown" or "other". They were also asked to identify roof shape from the options "square or rectangular", "circular or rounded", or "other". When finished, they were asked whether they had annotated all visible homes, or whether there were too many to annotate them all. They could then confirm their annotations and move on to another subject. The imagery used in annotation was captured by the Superview-1 constellation, the Disaster Monitoring Constellation 3 (DMC-3), and the Korea Multi-Purpose Satellite (KOMPSAT) 3A. Satellite imagery was selected with resolution ≤ 1 m. Locations in Kenya, Uganda, and Sierra Leone were chosen based on research team familiarity, visual inspection, and prior knowledge of areas with underrepresented housing styles (e.g. thatched roofs, multi-structure homesteads, refugee settings).