Object-Based Deep Learning for Probabilistic Convective Core Nowcasting from Satellite Data

Flash flooding from intense rainfall causes major damage and loss of life across Africa, particularly in the Sahel, where rainfall is dominated by mesoscale convective systems. Convective cores, which represent regions of intense convective activity, evolve rapidly, limiting short-term predictability, especially in data-sparse regions where numerical weather prediction models face substantial uncertainty in convective initiation, evolution, and organisation. This study presents an object-based deep learning framework for probabilistic convective core nowcasting using geostationary satellite data. Convective cores are identified from Meteosat Second Generation infrared imagery using a two-dimensional wavelet transform applied at mesoscale spatial scales. The framework is first demonstrated at a single location, where nearby core attributes are used to predict local core occurrence up to six hours ahead. Explainable AI analysis indicates that predictions are driven by physically meaningful factors related to core proximity, size, and intensity. The approach is then extended to the western Sahel through NCAST, a spatio-temporal Transformer architecture that models the evolution and interactions of convective core populations and produces gridded probabilistic nowcasts. Ablation experiments reveal that self-attention contributes most strongly to forecast skill, while temporal information becomes increasingly important at longer lead times. Evaluation against persistence and an operational conditional-climatology benchmark demonstrates skilful forecasts at 1-, 3-, and 6-hour lead times using multiple probabilistic and spatial verification metrics. These results highlight the potential of object-based learning to provide reliable short-range nowcasts in data-sparse environments.