A large and up-to-date map of Building Roof Area (BRA) is crucial for policy-making and sustainable development. In addition, as a fine-grained indicator of human activities, BRA can contribute to urban planning and energy modeling, bringing benefits to human well-being. However, existing large-scale BRA datasets, such as those from Microsoft and Google, do not include China, so China does not have a fully covered BRA map. For this purpose, we have created a multi-year Chinese Building Roof Area Dataset (CBRA) with a resolution of 2.5 meters based on Sentinel-2 images from 2016 to 2021. CBRA is China's first fully covered, multi-year BRA data. CBRA achieved good F1 scores in urban areas, with an F1 score of 62.55% based on 250000 test samples (+10.61% compared to previous BRA data in China), and a recall rate of 78.94% based on 30000 test samples in rural areas.
CBRA is organized in GeoTIFF (. tif) raster file format, with a single band and a GCSWGS_1984 coordinate system. The pixel values are 0 and 255, where 0 represents the background and 255 represents the roof area of the building. In addition, for the convenience of data usage, CBRA is divided into 215 spatial grid blocks, named "CBRA_year-E/W * * N/S * *. tif", where "year" is the sampling year and "E/W * * N/S * *" is the latitude and longitude coordinates of the upper left corner of the block data.
Sentinel-2 data, used for CBRA mapping. Sentinel-2 optical images are used for CBRA mapping. Sentinel-2 is an Earth observation mission under the European Space Agency's (ESA) Copernicus program, consisting of a constellation of two satellites, Sentinel-2A and Sentinel-2B. This product has radiation calibration through the system and terrain correction through geometry and the European Space Agency. To address cloud noise, we used GEE (Gorelick et al., 2017) to filter out more than 20% of clouds in the image, and further performed cloud and shadow removal through quality bands to obtain cloud free pixels. Finally, we performed median synthesis on the filtered images within a one-year interval.
Super resolution and semantic segmentation methods, as well as weakly supervised learning algorithms. Improve the geographic generalization of deep learning methods (i.e. extend to all regions of China), and also collect land cover data from China from 2016 to 2021 from dynamic world products. It includes 10 types of land cover and provides estimated probabilities for each type.
250000 test samples were collected in urban areas with a recovery rate of 78.94%, compared to 30000 test samples in rural areas.
This work is licensed under a
Creative
Commons Attribution 4.0 International License.
CSTR:11738.11.NCDC.ZENODO.DB6436.2024
10.5281/zenodo.7500612