From Geostatistics to Graph Attention Networks: A Holistic GeoAI Approach for Mapping Forest Soil Particle Size Distributions with Limited Samples

Mapping fine-grained soil particle size distributions (PSDs) in complex forest ecosystems remains a significant challenge in pedometrics. Traditional pixel-based machine learning approaches often struggle to capture the spatial heterogeneity and dependencies inherent in forest soils, particularly when ground-truth sampling is limited by cost and accessibility. This study presents a novel, holistic GeoAI framework that integrates geostatistical augmentation with Graph Neural Networks (GNNs) to map fine-grained soils (<60 µm) using LiDAR and Sentinel-2 data.

Our methodology addresses the "small data" problem through a two-stage process. First, we employed CoKriging (geostatistics) to locally upscale point-based soil samples within measured forest stands. This geostatistical interpolation generated a dense set of annotated training data, effectively augmenting the dataset to train GNN models. Second, we shifted from varying pixel resolutions to object-based analysis by segmenting forests into homogeneous polygonal zones based on tree species and canopy structure, which served as nodes in a graph structure.

We evaluated five GNN architectures (GAT, RGCN, GCN, EGNN, and MPNN) to predict PSDs using ~60 covariates derived from high-resolution LiDAR (geomorphometry, hydrology) and Sentinel-2 time-series (vegetation/soil indices). The graph attention network (GAT) emerged as the superior model, demonstrating remarkable stability and predictive accuracy. By utilizing multi-head attention mechanisms, the GAT model successfully learned the importance of neighboring nodes and complex spatial dependencies that standard convolutional models often miss. The GAT model achieved R² values exceeding 0.98 across all soil particle groups. Feature importance analysis revealed that LiDAR-derived geomorphometry (specifically elevation and downslope) and Sentinel-2 derived organisms (e.g., WDVI) were the dominant covariates driving predictions. This approach demonstrates that combining geostatistical data augmentation with the relational learning capabilities of GATs offers a scalable, accurate solution for digital soil mapping in data-sparse environments, with significant implications for forest management and hydrological modelling.

Reference: Abdi, O., Laamanen, V., & Uusitalo, J. (2025). Mapping forest fine-grained soil particle size distributions: a holistic GeoAI approach via graph neural networks, LiDAR, and Sentinel-2. International Journal of Applied Earth Observation and Geoinformation, 143, 104807. https://doi.org/10.1016/j.jag.2025.104807.