A method based on graph convolutional neural networks for predicting the computational intensity of geocomputation

High-performance cluster infrastructure supplies essential computational capabilities for large-scale geocomputation. However, many parallel geocomputation tasks are impacted by the heterogeneity of geospatial data, particularly when complex spatial structures or uneven feature distributions cause spatially uneven computational workloads across data partitions. This uneven cost is commonly described as computational intensity (CI), which reflects the amount of computation required per data unit and plays a critical role in data partitioning for parallel geocomputation. Existing CI estimation methods usually rely on expert-driven heuristics or simplified statistical indicators, which struggle to generalize across different parallel geocomputation tasks. As a result, inaccurate CI prediction leads to load imbalance and degraded parallel performance. To address those issues, the paper proposes a Learning-based Computing Framework for Geospatial data (LCF-G) to predict CI and guide parallel geocomputation. In LCF-G, geospatial data are represented as graphs to capture spatial relationships and organizational patterns influencing computational cost. A Graph Convolutional Network regression model for Geospatial data (GCN4G), is proposed to learn the mapping between graph-structured data characteristics and CI. Experiments on three representative geocomputation tasks show that GCN4G achieves an R-squared (R2) value above 0.8 in CI prediction, while LCF-G improves parallel efficiency to over 0.8, outperforming existing approaches.