SII-NowNet: A Machine Learning Tool for Nowcasting Convection Initiation and Intensification in the Tropics

Abstract Nowcasting developing convection is a crucial component of early warning systems in the tropics. While machine learning has proven effective for radar-based nowcasting, the lack of radar coverage across much of the tropics creates a significant capability gap. This study presents Simple Initiation and Intensification Nowcasting Neural Network (SII-NowNet), a machine learning tool that uses satellite brightness temperatures to produce probabilistic nowcasts of intensifying and initiating convection in the tropics. SII-NowNet is first demonstrated over Sumatra, Indonesia—a densely populated tropical island with frequent convective activity. For nowcasts of intensifying convection, SII-NowNet outperforms an optical flow model for lead times of 1–6 h but begins to overpredict events beyond 3 h, indicating its limit of capability. For nowcasts of initiating convection, SII-NowNet’s limit of capability is reached at 2 h, beyond which it overpredicts events and is outperformed by climatology. SII-NowNet is trained on 8661 samples (12 months of data), but sensitivity testing shows that the number of samples can be reduced to 3 weeks for intensification and 3 months for initiation, before it is outperformed by climatology. This has practical implications for the implementation and further development of SII-NowNet in resource-constrained settings. To exemplify generalizability in other tropical regions, SII-NowNet is tested over New Guinea, Zambia, Congo, and West Africa. Without retraining or region-specific tuning, SII-NowNet achieves skill scores comparable to those over Sumatra. Overall, SII-NowNet’s promising results, combined with ease of applicability across the tropics, make it a valuable tool for future operational nowcasting. Significance Statement In the tropics, short-term forecasts of destructive, rapidly developing storms are critical components of early warning systems—essential for protecting lives and minimizing impacts. However, many regions face challenges in accurate forecasting due to limited access to ground-based meteorological observations. This study presents a machine learning tool that uses freely and continuously available satellite data—covering the entire tropics—to forecast storm development several hours in advance. The tool performs well in five distinct tropical regions and requires minimal computational resources, showing its suitability for implementation in resource-constrained settings. Furthermore, the results highlight the challenges in capturing newly initiating storms—a focus for future studies. Continued work on the tool will focus on testing its performance in additional tropical regions.