A low-latency deep learning framework for volcanic ash cloud nowcasting using geostationary satellite imagery.

Rapid assessment of hazardous aerosol dispersion is critical for emergency response, yet operational dispersion workflows can exhibit end-to-end latency that is incompatible with the first minutes of decision-making. This study develops and validates a deep learning approach for near-real-time nowcasting of volcanic ash dispersion from geostationary observations. The model was trained on an archive of volcanic ash satellite imagery from EUMETSAT's SEVIRI instrument (Ash RGB composite) and achieved a structural similarity index of 0.88 for 15-minute next-frame forecasts. The complete edge workflow, including data download and inference, runs in under five seconds on an NVIDIA Jetson AGX Orin. To illustrate how the same nowcasting pipeline can be used for hypothetical scenario exploration across particulate sources, a pixel-based event-injection algorithm is introduced to overlay synthetic plumes of varying sizes into real-time satellite frames before inference. Scenario demonstrations parameterized by nuclear-yield-inspired sizes (10 kt to 100 Mt) are presented at urban (Paris, London, Berlin), national (Iberian Peninsula), and continental (Europe-wide) scales. These scenario outputs are intended as illustrative, low-latency visualizations of kinematic transport patterns in the SEVIRI observation space, as validated predictions of nuclear plume morphology. The primary contribution is a fast, low-cost volcanic ash nowcasting system, complemented by a generalizable injection framework for rapid scenario visualization on edge computing.