Does imagery add predictive value?

Each bar is the percent reduction in out-of-sample error versus a strong baseline (log lagged counts of both outcomes, log population and income, and census covariates). Zero means no improvement; negative means the imagery made predictions worse. Two validation schemes are shown, and the shuffled-imagery placebo shows what noise looks like.

Radar sees construction, but not crime

Sentinel-1 radar backscatter increases where new buildings are filed with the Department of Buildings. The sensor and pipeline work. They still add about one percent or less to crime prediction.

Filings are applications, not completed buildings, so counts are a noisy yardstick. This is a tract-level association, not recovery of individual construction sites.