Hexagonal discrete global grid system enhances model latitudinal spatial generalization and deep learning capabilities

Discrete grid systems are continually being developed to address uncertainties in Earth science data. The existing global parameters modeled on graticules are prone to distortions in high-latitude regions. With the explosion of Earth observation big data and Artificial Intelligence, the development of deep-learning techniques that rely heavily on massive training samples, reduce inherent grid-related biases, and improve generalization capabilities is an emerging challenge. The hexagonal discrete global grid system offers substantial potential as a robust framework for global-scale modeling, owing to its isotropy and uniform sampling. This study compares the ISEA projection-based hexagonal DGGS (HDGGS) and equal longitude–latitude grids from both machine learning and deep learning perspectives, evaluates their representational performance for various land surface parameters, and quantifies their cross-latitude generalization ability. An innovative deep learning architecture was developed, achieving a performance comparable to that of standard convolutional networks, while using fewer convolutional units. Under the HDGGS framework, the RMSE and the range of F1 scores for land-surface parameter inversion were lower than those obtained with longitude–latitude grids, indicating superior generalization performance. This study quantitatively characterizes the latitudinal consistency of the HDGGS, which can provide a spatial reference for enhancing the accuracy and reliability of big Earth data inversion.