Extending geostationary satellite AOD coverage with a lightweight spatiotemporal sequence model

Abstract

Satellite nadir retrievals of Aerosol Optical Depth (AOD) cannot be made over clouds which can limit their use for near-real-time air-quality monitoring. We evaluate a simple time-series approach that increases AOD coverage over clouds by predicting short-term transport and filling gaps using Machine Learning (ML). We use a single-channel convolutional long short-term memory (ConvLSTM) sequence model trained on NOAA GOES-18 ABI AOD dataset. Predicted plumes reproduce observed spatial organization despite the presence of clouds, and when composited into the observed AOD field by filling missing pixels, the model provides spatially coherent AOD estimates over obscured regions. With longer lead times, predictions become smoother and high-AOD extremes are unrealistically suppressed, narrowing the dynamic range relative to observations. Motion, quantified by an AOD-weighted centroid, shows frame-to-frame step distances that are smaller than observed. The distances decline with time, indicating accumulated transport error and under-advection of filaments. In addition, scene-mean AOD and its standard deviation are lower in predictions than in observations and decrease further with longer lead time (in 20-min intervals between frames). Performance degrades with increasing lead time and during prolonged, widespread obscuration, and may be less reliable for abrupt aerosol regime shifts. We compare ML model gap filling with ordinary kriging anchored to observed edges: kriging yields the smoother fields and tends to elevate AOD over broader regions, whereas ML-based filling preserves plume organization, avoids some edge amplification, and produces fewer extremes. Overall, a ConvLSTM provides timely AOD nowcasts and substantially extends coverage for short lead times, although performance degrades with lead time.