Spherical Machine Learning: A Revolutionary Paradigm Shift in Earth Science Research

This article focuses on how Spherical Machine Learning (Spherical ML) addresses the projection distortion issue of traditional planar convolutional neural networks (CNNs) on global-scale Earth science data, introduces the advantages of the HEALPix grid system, and demonstrates its breakthrough applications through cases of marine heatwave detection and biodiversity correlation. Key contributions include clarifying the failure modes of planar methods, providing technical solutions combining HEALPix and spherical harmonics, empirically verifying the advantages of spherical methods, and demonstrating cross-domain versatility.

Traditional planar CNNs assume data lies on an Euclidean plane, but the Earth is spherical. Latitude-longitude projections suffer from severe distortion in high-latitude regions, causing the same physical feature to appear in different pixel forms at different latitudes. Experiments show: a planar model trained at the equator sees its accuracy in detecting marine heatwaves at the poles plummet from 100% to 50% (random level), as the model learns projection distortions rather than real physical patterns.

Spherical convolution defines operations directly on the sphere, achieving rotational equivariance (feature recognition is unaffected by position) via spherical harmonics. The HEALPix grid is an ideal base: 1. Equal-area property avoids over-sampling in high latitudes; 2. Equal-latitude ring structure facilitates spherical harmonic transformation; 3. Nested hierarchy supports multi-resolution analysis; 4. Native integration with spherical harmonic transformation.

Synthetic data experiments: planar matched filters see their accuracy drop to 50% at 70-80°N, while HEALPix-based spherical harmonic filters maintain 100% accuracy. Real case (2011 Ningaloo Niño event): NOAA OISST data aggregated onto the HEALPix grid revealed that 94% of marine biological observation records are distributed in heatwave grid cells, and the correlation analysis is statistically significant only under the spherical framework.

Spherical methods maintain 100% accuracy in cross-domain transfer, while planar methods drop to 84.5%. The project adopts open science practices: Jupyter Notebooks for experiment reproduction, GitHub Actions for continuous integration, Docker containers to eliminate environment issues, MIT/CC-BY licenses to promote sharing, and nano-publications to support knowledge graph construction.

Spherical machine learning is a fundamental paradigm for processing global data, with core contributions including problem awareness, technical pathways, empirical verification, and cross-domain value. Implications for biodiversity research: more accurate global species distribution prediction, reliable climate-ecological correlation analysis, and multi-scale integration capability. Only by respecting the geometric essence can we build a reliable global environmental intelligence system.