Transforming noisy archives into high-fidelity soil physical and chemical property maps: a structure-aware weakly supervised framework

Accurate mapping of soil physical and chemical properties (SPCPs) constitutes a fundamental prerequisite for parameterizing the terrestrial water cycle in Earth system models, but continental-scale prediction is hindered by sparse observations and noisy legacy soil maps; it remains a formidable scientific bottleneck. Here, we address this challenge through a structure-aware Weakly-Supervised Learning (WSL) framework that treats legacy soil maps not as ground truth, but as imperfect, noisy priors. We introduce SoilViT, a specialized Vision Transformer architecture that integrates a dual-head disentanglement mechanism with hybrid geographic encoding to disentangle latent pedogenic signals from spatially structured errors. By enforcing edge-preserving total variation regularization, the model autonomously identifies and discards interpolation striping while preserving genuine high-frequency soil texture across diverse biogeographical gradients. Quantitative assessment reveals that our approach achieves an average R² of 0.8630 and RMSE of 0.0248. Notably, when validated against external profile data, the framework yields an R² improvement of up to 17.35%, with markedly reduced spatial autocorrelation of residuals. We demonstrate the framework’s robust performance by generating a 500 m resolution SPCPs dataset across China, which inherently corrects the systematic striping errors found in legacy inventories. These results demonstrate a scalable, resource-efficient pathway for global environmental monitoring in data-scarce regions.