PyEOGPR: A Python package for vegetation trait mapping with Gaussian Process Regression on Earth observation cloud platforms

Abstract

Developed to efficiently quantify vegetation traits from satellite Earth Observation (EO) data, the here presented PyEOGPR Python package makes trained probabilistic Gaussian Process Regression (GPR) models readily accessible within cloud-computing platforms like Google Earth Engine (GEE) and openEO. PyEOGPR provides a diversity of validated hybrid GPR models targeting common vegetation traits, as well as newer, more challenging ones such as canopy nitrogen content (CNC), applicable to Sentinel-2 (S2) and Sentinel-3 (S3) data. The package also enables users to incorporate newly trained GPR models for quantifying user-defined surface properties. A key advantage of GPR models is their provision of associated uncertainty estimates, significantly enhancing retrieval reliability. PyEOGPR streamlines large-scale vegetation analysis, facilitating quantitative map generation from local to global scales with customizable time windows, eliminating the need for local image downloads or processing. This paper outlines the complete processing pipeline and demonstrates the generation of landscape-scale maps of key vegetation traits using S2 (20 m resolution) data, and global trait maps using S3 data. PyEOGPR currently supports 27 generically applicable GPR models, aiding environmental monitoring and sustainable agroecological management, with minimal coding expertise required. This integration democratizes access to advanced GPR models within cloud environments, making spatial vegetation dynamics analyses accessible to a broader user base and improving the efficiency of EO data processing.