Open Buildings 2.5D Temporal Dataset

Open Buildings

• Image source: High-resolution 50 cm satellite imagery.
• Format: Building polygons with additional attributes e.g., confidence values.
• Spatial coverage: Subset of Global South.
• Temporal coverage: A temporal mosaic with freshness varying per location.
• Model: See arXiv:2107.12283.

Open Buildings 2.5D Temporal

• Image source: Copernicus Sentinel-2 imagery.
• Format: Per pixel rasters: building presence, building count, building height.
• Spatial coverage: Subset of Global South (same as the V3 Open Buildings Dataset).
• Temporal coverage: Annual: 2016-2023. Each year is temporally uniform.
• Model: See arXiv:2310.11622.

A deep learning model was trained to detect per pixel building presence, fractional counts and height using Copernicus Sentinel-2 imagery. We then ran inference for this model on stacks of 32 Sentinel-2 images centered around the 30th of June of each year (2016 - 2023) for most of the Global South. Only opaque-cloud-free images were used (according to the QA60 band). Heights for non-building objects were masked out using the building presence output of the model. The accompanying technical report describes the training and evaluation of the model.

Owing to the vast scale of the data processing, covering approximately 58 million square kilometers and spanning eight years, inference was run on TPUs. The resulting float rasters were then ingested into Google Earth Engine as an ImageCollection.

The effective spatial resolution is around 4 meters (rasters files are provided at 50 cm). The mean absolute error of building height prediction is 1.5 meters and was only evaluated in North America, Europe and Japan. The building height raster is only a relative measure (height above ground) and not an absolute measure of the location’s height. Moreover, given the effective spatial resolution of around 4 meters, we expect the absolute position of the building roof will also be shifted.

Model confidence values are uncalibrated, meaning, if the model assigns a confidence value of 0.8 that there is a building at a certain pixel it does not mean the actual likelihood of building presence is 80%. As such, confidence values can only be used for relative ranking (e.g., thresholding) of pixels. Also note that the model confidence can vary across location and time based on a number of factors, such as cloud cover, imagery misalignment, etc. Score thresholds found based on our validation data can be seen in table 11 of the technical report.

The building height raster should be only used in conjunction with the building presence raster, i.e., pixels with low confidence in building presence raster should be masked in the building height raster.

The Sentinel-2 images are not aligned with the imagery in Google Maps. Moreover, the individual Sentinel-2 images are not fully aligned to each other. Due to that the model output can often be misaligned with imagery in Google Maps and there can even be a slight spatial drift between different years of the dataset. To learn a little more about satellite imagery offset see these sites (1, 2). Also see the technical report for details about data limitations and quality.

There are a few limitations that you may want to keep in mind:

• Data availability: It could be that in very cloudy places we weren't able to create a full stack of usable imagery (16 frames before and 16 after the inference date). This could have affected the quality of the results.
• Spatial stability: Detections across time are not always aligned (example of small spatial jitter, demo link). This has to do with the inherent misalignment between input imagery. More on this also in the paper.
• Temporal stability: Detections across time are not always stable, meaning the same building could be assigned different confidences or height over time. Furthermore, aggregates like building count over an area could also be unstable depending on the region (example). This is due to factors like cloud cover, misalignment between input images and fewer Sentinel-2 timeframes for years 2016-2017 (example, demo link).
• Tiling artifacts: The dataset was generated by running inference on stacks of Sentinel-2 images divided into tiles. And since the model output isn’t expected to be spatially aligned with the neighboring tiles, this might result in tiling artifacts. (example).
• Limited heights ground truth data: Ground truth data for heights was only available in the Global North, and therefore heights prediction might not be as good in the Global South (e.g., potential large errors for tall buildings). For the same reason, performance evaluation in the Global South was mostly qualitative. Heights for non-building objects were masked out using building presence output of the model, which can sometimes introduce artifacts if, e.g., a tree was occluding a building. Heights for buildings are capped at 100 meters.
• Smaller buildings: Tiny buildings / tents (e.g., sparse refugee settlements) might not be detected with high confidence. Small touching buildings might not be resolved individually and may be shown as a single blob. Output is not guaranteed to be orthorectified, especially for tall buildings.
• False positives: Some natural features, e.g., rocks, snow (example , demo link) and ground discolorations as well as human built structures like solar panels (example, demo link) and agricultural foil greenhouses (example, demo link) can be incorrectly detected as buildings. False positives are also more likely to be present in locations where only a small number of frames could be used (example, demo link).

See the paper for more details about our approach and evaluation metrics.

The data is shared under the Creative Commons Attribution (CC BY-4.0) license and the Open Data Commons Open Database License (ODbL) v1.0 license. Users can review the terms for both licenses and select the license they prefer to use the data under. However, please note the liability disclaimer in the footnotes.[ec7ba7]

We have leveraged Copernicus Sentinel data (2016-2023) processed by the European Space Agency.

We wanted to make the data compatible for ingestion by those working with ODbL-licensed datasets (namely the OpenStreetMap community) while enabling people who don't use ODbL licensing to use it under the terms of the CC BY-4.0 license. We hoped to take away the burden of figuring out whether the two licenses were compatible and simply release the dataset under both licenses.

The model is trained and run on Sentinel-2 satellite imagery, which is publicly accessible through the Earth Engine Data Catalog. However, high-resolution imagery and the associated annotations used for training the teacher model (that generates labels for the Sentinel-2 stack) is not publicly released.

To access the model, please contact open-buildings-dataset@google.com detailing your intended use case and organization. Requests will be evaluated on a case-by-case basis.

We can run inference for timestamps of interest or at higher than annual temporal frequency, so long as there are sufficient opaque-cloud-free images. Please contact open-buildings-dataset@google.com detailing your intended use case and organization. Requests will be evaluated on a case-by-case basis.

Sensitive areas, including conflict zones, were omitted from consideration to protect at-risk populations. The list of excluded areas is subject to change.

If this dataset is useful, please consider citing the technical report:

W. Sirko, E.A. Brempong, J.T.C. Marcos, A. Annkah, A. Korme, M.A. Hassen, K. Sapkota, T. Shekel, A. Diack, S. Nevo, J. Hickey, J.A. Quinn. High-Resolution Building and Road Detection from Sentinel-2. arXiv:2310.11622, 2023.

Please contact open-buildings-dataset@google.com with any feedback.