Methods

Algorithms in this survey

Each detector below runs independently across every state COG and writes its own GeoJSON layer. They disagree by design — agreement between layers is the most reliable signal of a true pivot.

NAIP 30 cm — Hough naip_kansas_30cm

Single Hough Circle Transform algorithm run on USDA NAIP aerial imagery (30 cm native, processed at 1.2 m via COG overview). Kansas only. Recall ~75 %, precision near 100 % — high-confidence single algorithm without filtering needed.

Pipeline

For every state we (1) query the Element 84 Earth Search STAC for Sentinel-2 L2A scenes in a 45-day window, (2) mask cloud-flagged pixels using the SCL band and compute a per-pixel median across all clear timesteps, (3) reproject and mosaic into a single Cloud-Optimized GeoTIFF in the regional Albers / Lambert projection, then (4) hand the COG to every algorithm in parallel. Each algorithm's detections are stored as a GeoJSON layer with center, radius, and per-method score.