A Low-Latency Deep Learning Framework for Volcanic Ash Cloud Nowcasting Using Geostationary Satellite Imagery

This codebase implements the training, evaluation, and inference components of a low latency deep learning system for volcanic ash cloud nowcasting from geostationary satellite imagery. It includes a memory efficient ConvLSTM training pipeline for next frame prediction on large satellite archives, with time based data splitting and GPU optimized data loading. The framework also provides multi step recursive forecasting and quantitative evaluation across lead times using image based error and perceptual metrics. Together, these components operationalize the methodology for a Low-Latency Deep Learning Framework for Volcanic Ash Cloud Nowcasting Using Geostationary Satellite Imagery and support reproducible development and testing of rapid ash transport nowcasts.