DataValidationReport.md

August 16, 2023 ยท View on GitHub

Data Validation Report

For each Satlas geospatial data product, we estimated the precision and recall of the data as of 2022-10 on a per-continent basis.

Precision is the percentage of objects in the geospatial data products that are correct (not errors), while recall is the percentage of actual objects in the world that are covered in the geospatial data (not missing).

We estimated precision by uniformly sampling 100 objects from the 2022-10 data within each continent, and manually determining whether each sampled object was correct. Sometimes, a confident determination could not be made due to the lack of sufficient data, in which case we made a best effort guess. The precision percentages reported below are equal to the number of objects we determined were correct in each continent. Missing values indicate that fewer than 100 objects were found in that continent.

We estimated recall using our hand-labeled validation sets. Each recall percentage is the percentage of ground truth objects in the validation set within each continent that was output by our model, at the same threshold we used for deployment. Missing values indicate that there were insufficient ground truth objects in that continent.

"Objects" are points for wind turbines and platforms. for solar farms, objects are predicted polygons for precision, and individual pixels for recall.

Below, Count shows the number of objects in the geospatial data products.

This report may change as the data is improved over time. The current version is last modified 16 August 2023.

Solar Farms

ContinentPrecisionRecallCount
Africa57N/A1,051
Asia9777.987,197
Europe9478.242,471
North America8689.115,493
South America7667.82,091
Oceania7457.4797
All9578.9149,100

Many incorrect solar farms were actually greenhouses, but other errors were diverse, often occurring on undeveloped land that was dark or metallic in appearance. Errors occur roughly proportionally to land area, so regions with the most actual solar farms (Asia, Europe, North America) have the highest precision.

Onshore Wind Turbines

ContinentPrecisionRecallCount
Africa100N/A2,223
Asia10096.7130,233
Europe9892.578,283
North America10087.569,161
South America99N/A8,371
Oceania94N/A3,471
All9993.2291,742

Most (6/9) incorrect onshore wind turbines were actually transmission towers. Other incorrect turbines were rare mistakes in empty sites (1/9) or undeveloped land (2/9).

Offshore Wind Turbines

ContinentPrecisionRecallCount
AfricaN/AN/A0
Asia9991.15,509
Europe10097.05,753
North AmericaN/AN/A10
South AmericaN/AN/A0
OceaniaN/AN/A0
All9994.111,272

Only one offshore incorrect wind turbine was identified, in Asia, which was actually a transmission tower in the water at (25.2260, 119.1385).

Offshore Platforms

ContinentPrecisionRecallCount
Africa9090.6842
Asia9648.36,531
Europe8866.11,087
North America9487.52,277
South America96N/A1,329
Oceania81N/A126
All9461.712,192

Incorrect offshore platforms were generally islands or locations frequently occupied by anchored fishing vessels.

Tree Cover

Data validation is currently unavailable for tree cover.