AtmoLens

AtmoLens is an automated, end-to-end Python pipeline for detecting and characterising significant air-quality events in ECMWF atmospheric composition forecast fields. Moving beyond today’s fixed-threshold alerts, AtmoLens employs three parallel detection branches: an adaptive percentile-threshold with blob analysis as a transparent operational baseline; a multi-scale Hessian ridge filter that detects aerosol plumes by their physical shape and tracks them across forecast lead times using tobac; and a transfer-learned U-Net convolutional segmentation model adapted from a validated wildfire detection system built on Landsat 8 imagery, fine-tuned on weakly- supervised pseudo-labels derived from the 20-year CAMS EAC4 reanalysis. Branch outputs are fused through a configurable mask ensemble layer, and every detected event is characterised by centroid location, area, intensity, aspect ratio, orientation, and propagation velocity, exported as GeoJSON and NetCDF for direct integration into ECMWF operational workflows. A convolutional autoencoder variant provides fully unsupervised anomaly detection for rare-event variables such as volcanic SO₂. The pipeline is configurable via a single YAML file, deployable via Docker, and validated against documented wildfire smoke, Saharan dust, and volcanic eruption case studies with a rigorous benchmark comparing detection skill against computational cost across all methods.