Correlation of excavated soil multi-source heterogeneous data using multimodal diffusion model

Abstract The sustainable utilization of excavated soil as a geomaterial requires a comprehensive understanding of its multi-dimensional properties, but correlating heterogeneous data (e.g., visual, mechanical, and electrical characteristics) remains a challenge. To address this, an excavated soil information collecting system was developed to acquire multi-source data including RGB images, cone index (CI) curves, and TDR waveforms—from China’s largest soil transfer platform, establishing a database of 23,122 sets. A generative-model-aided correlation analysis framework was proposed, leveraging a denoising diffusion probabilistic model to explore inherent relationships between soil properties. Performance metrics, such as SSIM, LPIPS, and RMSE, were employed to analyze the model's training results. Key findings reveal that: (1) soil images encode water content information, which correlates with CI curves and TDR waveforms; (2) CI and TDR data cannot capture color-based mineral composition details from images; and (3) TDR waveforms uniquely detect pollution indicators (e.g., electrical conductivity), undetectable via other methods. This AI-driven approach provides a novel methodology for analyzing multi-dimensional property correlations in geotechnics, enhancing sustainable soil reuse.