Decoding Nonlinear Signals in Multidimensional Precipitation Observations

High-quality hydrometeor microphysical observations are essential inputs to global precipitation retrievals and numerical weather prediction models. Detailed observations of particle size distributions, shape, fall speed, and effective density provide crucial insights into particle habit and derived surface precipitation rates. However, interpreting large, multidimensional datasets is challenging due to significant regional and seasonal variability in precipitating processes. In this study, we employ both linear and nonlinear dimensionality reduction techniques to reveal global modes of precipitation variability. Using over one million minute-scale observations from the Precipitation Imaging Package (PIP) across 10 sites over 10 years, we developed a robust, unified dataset of particle microphysical characteristics named pipdb. Principal Component Analysis (PCA) of pipdb snowfall data identified six distinct process groups based on latent features representing precipitation intensity, particle density, and snowfall regime. These groups were validated through physically consistent categorizations when compared with independent, collocated ancillary observations from surface radars and meteorologic instruments. However, PCA's assumptions of linearity left nearly half of the dataset ambiguous. By contrast, nonlinear techniques like Uniform Manifold Approximation and Projection (UMAP) provided superior separation of microphysical characteristic groups, including clear categorizations of particle phase, precipitation intensity, and shape. Cluster quality analyses demonstrated an increase of +0.14 in the silhouette score using UMAP over PCA along with 35% fewer ambiguous cases, indicating a more accurate representation of the underlying global data structure. UMAP accurately classified cases into snow, rain, mixed-phase, and transition groups, offering a robust statistical method for categorizing traditionally challenging precipitation characteristics like hydrometeor phase. These findings highlight the capabilities of unsupervised machine learning in uncovering physically meaningful signals within large, multidimensional datasets, thereby enhancing our understanding of complex patterns in global precipitation processes.