A flight-first approach for contrail detection using ground-based cameras

Abstract. Validating forecasts and models used in contrail mitigation efforts requires high fidelity observations. The ability to match individual contrails with their progenitor flights in such observations is vital. High-resolution ground cameras can offer such high-fidelity flight attributions through direct observations of contrail formation, something that low-resolution observation methods (e.g., geostationary satellites) cannot due to their optical limitations. Previous "detection-first" approaches using ground cameras have employed flight matching after detecting unattributed contrails and can thus still struggle with the attribution task due to having no inherent mechanism to distinguish between newer and older contrails, particularly suffering when in-frame contrail densities are large. Here, we present a "flight-first" approach that focally searches for contrails within dynamic bounding boxes around wind-advected flightpaths, enabling inherent flight matching for young contrails while reducing computational costs. We demonstrate our approach using ground cameras from the Global Meteor Network (GMN), the SAM2 asemantic segmentation model, and a simple random forest classifier model (RFM). In a case study of 3,506 flights across three weeks in the Southwestern USA, we detect 265 contrails with a precision of 0.99 and a recall of 0.74. The flight-first approach uniquely allows us to pinpoint where detections were missed by the SAM2-RFM pipeline, identifying cloud obscuration as a prominent driver of failure. We thus propose that our off-the-shelf approach can be deployed at-scale as a primary low-cost layer in a wider contrail detection strategy for model evaluation, serving to identify places and times for follow-up observations with more advanced observation techniques. The GMN is well suited to such a task given its global coverage and standardised components.