The U.S. Census Bureau has released a new experimental data product featuring monthly single-family housing starts for the West South-Central division. For select areas of interest, Survey of Construction (SOC) data collected by field representatives have been replaced with estimates obtained from satellite imagery to demonstrate the efficacy of this data collection method.
Satellite images are acquired for selected areas of interest and run through a series of convolutional neural networks (CNNs) that have been trained to identify construction stages (excavation, foundation, framing, unfinished roof, and finished roof) as well as likely building type (single-family detached, single-family attached, multi-family, and non-residential). Model output is compared for consecutive months to identify new single-family residential construction starts. These starts are combined with traditional survey data collected for the remaining areas to create a hybrid estimate for the West South-Central division.
This experimental data product provides a historical monthly data series for the division from July to November 2024, and the Census Bureau anticipates it will be updated monthly going forward.
This product demonstrates a new computer-vision based approach to monthly economic measurement that can improve geographic granularity and accuracy while also reducing respondent burden.
For more information, please review the Census Bureau’s methodology documentation. This is the first version of this experimental data. The Census Bureau plans for future refinements to its models and invites users to provide feedback on how to improve this experimental product.
Feedback can be shared by email at EID.NRC.Satellite@census.gov.