Mapping forest fine-grained soil particle size distributions: a holistic GeoAI approach via graph neural networks, LiDAR, and Sentinel-2

Abstract

Fine-grained soils are crucial for assessing forest diversity and soil disturbances. Existing models for predicting particle size distributions (PSDs) often rely heavily on soil samples or lack necessary spatial dependencies, scalability and flexibility. This study introduces a holistic GeoAI model using graph neural networks (GNNs), LiDAR, and Sentinel-2 data to address these limitations. We collected 330 soil samples from 47 forest stands with a random-stratified method in southwestern Finland. The samples were pre-processed and analyzed for PSDs using a laser diffraction method, and classified into four groups: <2 µm, 2–6 µm, 6–20 µm, and 20–60 µm. To increase the number of annotations, we predicted soil PSDs at unmeasured locations using CoKriging within stands. The forests were segmented into small homogeneous polygons to construct the graph layer. We mapped 61 covariates using LiDAR and Sentinel-2 based on scorpan model, which were then summarized into the graph layer. Subsequently, we established the pipelines of five GNN models regarding the top covariates. The results indicate that geomorphometry and organisms covariates accounted for the majority of importance. The graph attention network (GAT) recorded high stability during training and remarkable prediction accuracy after testing with R² values above 0.98 in predicting fine-grained soil PSDs across all four soil groups. Conversely, the relational graph convolutional networks (RGCN) also achieved R² values above 0.97, but with lower stability and longer training times. However, the high accuracy of the predictive models is partly due to the large number of annotations derived from CoKriging, which may introduce uncertainties. Our GAT model demonstrated strong transferability when applied to an independent test stand using CoKriging-derived data (R²: 0.98–0.99) and showed robust performance when evaluated against real ground-truth samples (R²: 0.88–0.95). The observed prediction errors (RMSE: 0.68–2.82) reflect a combination of uncertainties originating from the CoKriging training data (RMSE: 0.34–2.46) and model-induced errors during training (RMSE: 0.37–1.46). Nevertheless, the consistently high R² values indicate a strong agreement between predicted and measured soil PSDs. Future studies should focus on training the model with a larger number of ground-truth soil samples and evaluating its transferability across diverse boreal forest landscapes.