Large-Scale Deep Learning for Global Precipitation Type Classification

Accurate classification of precipitation type (convective vs. stratiform) from passive microwave satellite observations is fundamental for understanding global precipitation patterns and improving weather prediction models. While previous studies have demonstrated the effectiveness of deep learning approaches with accuracies above 90% on smaller temporal windows, questions remain about model generalization across extended time periods and different surface types. This study presents a comprehensive analysis using a nine-year Global Precipitation Measurement (GPM) mission dataset, followed by a detailed investigation of surface-type specialization. 

Our primary analysis leverages over 400 million samples from 2014-2022, using 32x32 pixel patches and a ResNet architecture. This large-scale model achieved an accuracy of 85% on a holdout test-year, with balanced performance across both precipitation types (F1-score of 0.85 for both convective and stratiform classes). While matching the general performance range of previous approaches, these results demonstrate robust generalization capabilities across a much longer temporal span and diverse global conditions, using a significantly larger training dataset. 

To further investigate model generalization, we conducted a specialized analysis to examine performance across different surface types creating distinct datasets for land and ocean. ResNet-50 architectures were trained for three comparative models: a baseline model using combined data, a land-only, and an ocean-only model. Analysis revealed that the baseline model achieved robust performance across both surface types (82 % accuracy over land, 86 % over ocean). Surprisingly, surface-specific models showed minimal improvement, with the land-specific model achieving 81% and the ocean-specific model reaching 85% accuracy on their respective domains. This suggests that larger, diverse datasets enable models to learn more robust features that generalize well across data subsets. 

These findings demonstrate that deep learning models can effectively maintain consistent performance when scaling to multi-year global datasets, and suggest that investing in larger, more diverse training datasets provides robust generalization across both temporal and spatial dimensions without requiring specific subsetting. The approach could lead to more efficient operational systems while maintaining reliable classification accuracy across different surface types and extended time periods.