Automated Nowcasting of High-Impact Convection over Morocco: Comparing Classical Computer Vision and Deep Learning Pipelines in Satellite Remote Sensing

Rapid detection and monitoring of deep atmospheric convection is a cornerstone of early warning systems and severe weather preparedness across the world's most vulnerable regions. Within modern data-intensive earth observation workflows, traditional manual analysis of multi-spectral satellite imagery remains inherently subjective, introducing significant inter-operator variability while lacking scalability. This challenge is further compounded in regions characterized by complex topography and pronounced microclimatic diversity, such as Morocco. This study presents a comparative analysis of two automated cloud segmentation paradigms utilizing Severe Storms RGB data for the rapid identification of high-impact convective zones. We contrast a classical computer vision baseline—HSV color-space thresholding via OpenCV—against a deep learning approach based on a U-Net convolutional neural network. The data-driven architecture was trained on a meteorological dataset of 2,552 manually annotated satellite images spanning 2023–2024 and covering a study domain of 40°N–20°N and 20°E–5°W, using a joint Binary Cross-Entropy and Dice loss formulation. Quantitative validation against expert-generated ground truth demonstrates that the U-Net pipeline significantly outperforms the classical baseline, achieving a mean Intersection over Union (IoU) of 0.733 and a Dice coefficient of 0.754, versus 0.562 and 0.601 for the OpenCV approach. Notably, the U-Net architecture yields near-perfect spatial segmentation in over 50% of test cases with lower overall variance, while Bland—Altman analysis quantifies a systematic 15.3 percentage-point performance advantage for the neural network. These findings underscore the readiness of deep learning for automated remote sensing nowcasting pipelines, establishing a scalable, high-accuracy framework to support natural hazard mitigation and real-time atmospheric situational awareness.