DeepMAP-UTA

DeepMAP-UTA: Deep Learning-based Multiple Air Pollutant mapping with a Unified architecture and Temporal Adaptation.

What you see. Hourly surface PM10, PM2.5, NO₂ and O₃ estimated at 0.1° from GEMS, CAMS, KIM meteorology and ancillary data by the final all-sky model (QC2, epoch 140). The field is the model estimate everywhere, including over sea.

Stations. Circles are observations at training stations. Triangles are separate-test stations the model never saw; the evaluation panel uses only these.

Times are observation times in UTC. Daily values are UTC days.

Satellite imagery. GEMS pseudo-colour shows the GEMS scan that went into each hourly map (normalised radiance at 490, 443 and 412 nm as red, green and blue; not true colour), over the NASA Blue Marble where GEMS did not look. GEMS clouds show the same scan only where the GEMS cloud fraction is high (0.15 transparent, 0.7 and above nearly opaque), behind the estimate. MODIS and VIIRS true colour from NASA GIBS (Worldview) are daily images for the selected date. Outside the model domain the background is dimmed. With imagery on, the night side is darkened by the solar zenith angle (84° to 96°) at the map hour, optionally with NASA Black Marble city lights.

Boundaries: OpenDataSoft world boundaries, GADM 4.1, OCHA/WFP. Lines and names are for orientation only.